audioVersionDurationSec
float64 0
3.27k
⌀ | codeBlock
stringlengths 3
77.5k
⌀ | codeBlockCount
float64 0
389
⌀ | collectionId
stringlengths 9
12
⌀ | createdDate
stringclasses 741
values | createdDatetime
stringlengths 19
19
⌀ | firstPublishedDate
stringclasses 610
values | firstPublishedDatetime
stringlengths 19
19
⌀ | imageCount
float64 0
263
⌀ | isSubscriptionLocked
bool 2
classes | language
stringclasses 52
values | latestPublishedDate
stringclasses 577
values | latestPublishedDatetime
stringlengths 19
19
⌀ | linksCount
float64 0
1.18k
⌀ | postId
stringlengths 8
12
⌀ | readingTime
float64 0
99.6
⌀ | recommends
float64 0
42.3k
⌀ | responsesCreatedCount
float64 0
3.08k
⌀ | socialRecommendsCount
float64 0
3
⌀ | subTitle
stringlengths 1
141
⌀ | tagsCount
float64 1
6
⌀ | text
stringlengths 1
145k
| title
stringlengths 1
200
⌀ | totalClapCount
float64 0
292k
⌀ | uniqueSlug
stringlengths 12
119
⌀ | updatedDate
stringclasses 431
values | updatedDatetime
stringlengths 19
19
⌀ | url
stringlengths 32
829
⌀ | vote
bool 2
classes | wordCount
float64 0
25k
⌀ | publicationdescription
stringlengths 1
280
⌀ | publicationdomain
stringlengths 6
35
⌀ | publicationfacebookPageName
stringlengths 2
46
⌀ | publicationfollowerCount
float64 | publicationname
stringlengths 4
139
⌀ | publicationpublicEmail
stringlengths 8
47
⌀ | publicationslug
stringlengths 3
50
⌀ | publicationtags
stringlengths 2
116
⌀ | publicationtwitterUsername
stringlengths 1
15
⌀ | tag_name
stringlengths 1
25
⌀ | slug
stringlengths 1
25
⌀ | name
stringlengths 1
25
⌀ | postCount
float64 0
332k
⌀ | author
stringlengths 1
50
⌀ | bio
stringlengths 1
185
⌀ | userId
stringlengths 8
12
⌀ | userName
stringlengths 2
30
⌀ | usersFollowedByCount
float64 0
334k
⌀ | usersFollowedCount
float64 0
85.9k
⌀ | scrappedDate
float64 20.2M
20.2M
⌀ | claps
stringclasses 163
values | reading_time
float64 2
31
⌀ | link
stringclasses 230
values | authors
stringlengths 2
392
⌀ | timestamp
stringlengths 19
32
⌀ | tags
stringlengths 6
263
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0
| null | 0
|
404ade455166
|
2018-09-04
|
2018-09-04 21:53:48
|
2018-09-17
|
2018-09-17 06:01:01
| 7
| false
|
ru
|
2018-09-17
|
2018-09-17 06:01:01
| 2
|
17c6087892f5
| 4.887736
| 0
| 0
| 0
|
Могут ли вопросы на собеседовании по машинному обучению быть одновременно прикольными и глубокими?
| 5
|
25 прикольных вопросов для собеседования по машинному обучению
Могут ли вопросы на собеседовании по машинному обучению быть одновременно прикольными и глубокими?
Многие из исследователей данных изучают машинное обучение (МО) в основном с точки зрения практического специалиста. Как следствие, мы можем сосредотачиваться скорее на освоении как можно большего количества новых пакетов, фреймворков и методик, чем на глубоком рассмотрении основополагающих теоретических аспектов МО. Кроме того, в этом материале моё определение машинного обучения включает в себя обыкновенное статистическое обучение (т. е. не ограничивается глубинным обучением).
Однако пытливый, рассудительный и упорный ум может придумать множество чудесных вопросов по МО, разбор ответов на которые прекрасно способен обнаружить более глубокие его аспекты. В общем, такие вопросы могут помочь высунуть голову из кучи с картинки выше. Мы ведь не хотим целыми днями перемешивать данные — мы хотим нырнуть в глубины свойств, странностей и тонкостей методик машинного обучения и проникнуться ими…
В конце концов, в интернете хватает статей о «стандартных вопросах для собеседования по машинному обучению». Может, сделать кое-что другое, поинтереснее?
Сразу скажу: я публикую эти вопросы, просто чтобы вдохновить вас на размышления и разговоры. Готовых ответов не даётся. В некоторых вопросах есть подсказки, но они на самом деле больше для обсуждений, они не указывают на определённый ответ. Каждый вопрос заслуживает подробного обсуждения. Нет какого-то одного ответа. Некоторые вопросы заумные, а некоторые просто забавные. Приятного чтения :-) Вдобавок я вставил смешной мем после каждого 5-го вопроса…
Прикольные вопросы
«P > 0,05. Игра окончена, попробуйте ещё раз».
Я построил линейную регрессионную модель, показывающую 95%-ный доверительный интервал. Означает ли это, что существует 95%-ная вероятность, что коэффициенты моей модели верно оценивают функцию, которую я хочу аппроксимировать? (Подсказка: на самом деле это означает 95% времени…)
В чём сходство между файловой системой Hadoop и алгоритмом k-ближайших соседей? (Подсказка: «лень»)
Какая структура мощнее в смысле выразительных возможностей (т. е. она может достоверно отобразить заданную булеву функцию) — однослойный перцептрон или двухслойное дерево принятия решений? (Подсказка: XOR)
А что мощнее — двухслойное дерево принятия решений или двухслойная нейронная сеть без активационных функций? (Подсказка: нелинейность?)
Может ли нейронная сеть служить инструментом для понижения размерности? Объясните как.
«Виды головной боли: мигрень, перенапряжение, стресс, математические основы глубинного обучения».
Все ругают и обесценивают понятие постоянного слагаемого в линейных регрессионных моделях. Назовите одно из его применений. (Подсказка: сборщик шума/мусора)
LASSO-регуляризация сводит коэффициенты точно к нулю. Гребневая регрессия сводит их к очень маленькому, но не нулевому значению. Можете ли вы объяснить разницу между ними интуитивно, по графикам двух простых функций: |x| и x²? (Подсказка: острые уголки на графике |x|)
Допустим, что вы ничего не знаете о распределении, откуда взят некий набор данных (непрерывнозначные числа), и вам запрещается предполагать, что это гауссово нормальное распределение. Приведите как можно более простое доказательство того, что, независимо от характера распределения, вы можете гарантировать, что ~89% данных лежат в пределах ±3 стандартных отклонений от среднего значения. (Подсказка: научный руководитель Маркова)
Большая часть алгоритмов машинного обучения так или иначе связана с операциями над матрицами, например перемножением или обращением. Дайте простое математическое доказательство, почему мини-пакетная версия такого алгоритма МО может быть более эффективна с точки зрения объёма расчётов, чем обучение на полном наборе данных. (Подсказка: временная сложность перемножения матриц…)
Не кажется ли вам, что временной ряд — это очень простая задача линейной регрессии с единственной переменной отклика и с единственным предиктором — временем? В чём проблема метода линейной регрессии (необязательно с единственным линейным членом, с многочленами тоже) в случае данных временного ряда? (Подсказка: прошлое указывает на будущее…)
«Что если я скажу тебе, что это — регрессия?»
Приведите простое математическое доказательство того, что поиск оптимального дерева решений для задачи классификации среди всех возможных древовидных структур может быть экспоненциально сложной задачей. (Подсказка: а сколько вообще деревьев в джунглях?)
Как деревья решений, так и глубокие нейронные сети являются нелинейными классификаторами, т. е. они разбивают пространство посредством сложной границы решений. Почему в таком случае модель дерева решений настолько интуитивно понятнее глубокой нейронной сети?
Обратное распространение — рабочая лошадка глубинного обучения. Назовите несколько возможных альтернативных методик обучения нейронной сети без использования обратного распространения. (Подсказка: случайный поиск…)
Допустим, что у вас две задачи — на линейную регрессию и на логистическую регрессию (классификацию). Какую из них с большей вероятностью упростит открытие нового сверхскоростного алгоритма перемножения матриц? Почему? (Подсказка: какая из них с большей вероятностью использует операции над матрицами?)
Как корреляция между предикторами затрагивает метод главных компонент? Как с этим справиться?
«Что делать с… корреляцией?»
Вам поручили построить классификационную модель столкновения метеоритов с Землёй (важный проект для человеческой цивилизации). После предварительного анализа вы получаете 99%-ную достоверность. Вы можете быть довольны? Почему нет? Что с этим можно сделать? (Подсказка: редкое событие…)
Возможно ли определить корреляцию между непрерывной и дискретной переменной? Если да, то как?
Если вы работаете с данными по экспрессии генов, часто бывает так, что предикторных переменных миллионы, а замеров только несколько сотен. Приведите простое математическое доказательство, почему обычный метод наименьших квадратов — плохой выбор в такой ситуации, когда нужно построить регрессионную модель. (Подсказка: кое-что из матричной алгебры…)
Объясните, почему k-проходная перекрёстная проверка плохо работает с моделями временного ряда. Что можно сделать по этому поводу? (Подсказка: ближайшее прошлое прямо указывает на будущее…)
Простая выгрузка случайной выборки из обучающего набора данных в обучающую и проверочную выборку хорошо подходит для задачи регрессии. А что может пойти не так с этим подходом для задачи классификации? Что с этим можно сделать? (Подсказка: все ли классы преобладают в одной и той же степени?)
«Случайная выборка не работает!»
Что для вас важнее — достоверность модели или качество модели?
Если бы вы могли воспользоваться многоядерным процессором, вы бы предпочли алгоритм бустинга над деревьями случайному лесу? Почему? (Подсказка: если задачу можно делать 10 руками, стоит этим воспользоваться)
Представьте, что ваш набор данных наверняка линейно разделим и вам нужно гарантировать сходимость и наибольшее число итераций/шагов в вашем алгоритме (из-за вычислительных ресурсов). Выбрали ли бы вы в таком случае градиентный спуск? Что можно выбрать? (Подсказка: какой простой алгоритм с гарантией обеспечивает нахождение решения?)
Пусть у вас крайне мало памяти/места для хранения данных. Какой алгоритм вы предпочтёте — логистическую регрессию или k-ближайших соседей? Почему? (Подсказка: пространственная сложность…)
Вы строите модель машинного обучения, и изначально у вас было 100 точек данных и 5 признаков. Чтобы уменьшить смещение, вы удвоили количество признаков (включили 5 новых переменных) и собрали ещё 100 точек данных. Объясните: правильный ли это подход? (Подсказка: на машинное обучение наложено проклятье. Слышали о нём?)
«Меня прокляли размерностью!»
Перевод статьи Tirthajyoti Sarkar “25 fun questions for a machine learning interview”
|
25 прикольных вопросов для собеседования по машинному обучению
| 0
|
25-прикольных-вопросов-для-собеседования-по-машинному-обучению-17c6087892f5
|
2018-09-17
|
2018-09-17 06:01:02
|
https://medium.com/s/story/25-прикольных-вопросов-для-собеседования-по-машинному-обучению-17c6087892f5
| false
| 1,017
|
Перевод и адаптация статей в сфере IT
| null | null | null |
Nuances of programming
|
nuancesprog@gmail.com
|
nuances-of-programming
|
NUANCES OF PROGRAMMING,NUANCESPROG
|
nuancesprog
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Alisa Tolstova
| null |
323f6f4f9e94
|
alisa.tolstova
| 3
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
a60ddeb25059
|
2018-06-26
|
2018-06-26 08:42:38
|
2018-06-26
|
2018-06-26 14:16:49
| 2
| false
|
en
|
2018-06-26
|
2018-06-26 14:16:49
| 6
|
17c685c4fadd
| 3.787107
| 3
| 0
| 0
|
“Turn your dreams into reality with a personal loan,” reads the website of a major bank with a presence in several developing countries…
| 5
|
Traditional Finance is Failing the World’s Poor. Can AI Help?
Rawpixel/Unsplash
“Turn your dreams into reality with a personal loan,” reads the website of a major bank with a presence in several developing countries. Two of the requirements to get the loan? Be a salaried employee and have a net monthly income of no less than $200.
Yet according to the 2016 Poverty and Shared Prosperity Report by the World Bank, 10 percent of people live on less than $1.90 a day. And half of world’s extreme poor — amounting to 389 million people — live in Sub-Saharan Africa. This dream-come-true loan product locks out most people in developing countries.
But where traditional finance is failing in the developing world, new and innovative technologies — which rely heavily on AI — are poised to transform personal finance for millions.
Take microlending, for instance. Leveraging the peer-to-peer economy, microloans are small loans issued by individuals who each contribute some portion of the total amount. They are often given to people in developing countries who don’t have access to traditional financing and want to start businesses.
Microlending mobile applications like Tala and Branch are in very high demand in the developing world. They collect a range of data about applicants, including biographical information, the number of people they contact daily, the size of their networks and support systems, and their habits. They then use machine learning algorithms (such as regression trees or support vector machines) to train credit scoring models to determine how much to lend.
Tala, present in Kenya, Tanzania and the Philippines, has already disbursed over 5.6 million loans to over a million customers ranging from $10 to $300. Earlier this year, they raised $65 million to expand to Mexico and India.
These microlenders come with a wealth of advantages, including speed and accessibility. Loans can be disbursed almost instantly. Compared to the requirements laid out by banks, microlenders present a much lower barrier to entry, providing a line of credit to individuals who previously would have been unable to scale their small businesses.
Twiga Foods, a business-to-business food distribution company, has introduced a blockchain-based micro-financing product piloted by a couple hundred kiosk owners in Kenya. The company is leveraging blockchain and machine learning to calculate credit scores and disburse microloans.
The loans, which are disbursed through mobile phones, are used as working capital by the kiosk owners, which enables them to purchase and sell higher volumes of food per day. They make an order to Twiga Foods and receive a text message with loan options to pay for the order. In the pilot, order sizes saw an increase of 30 percent, with each retailer’s profits increasing by 6 percent — all with an average loan of about $30.
Despite the benefits of a streamlined and instantaneous loan process, there are also significant challenges. For starters, the loan applications have very high interest rates with equally high penalties for late loan repayment. Lenders justify this by pointing to the riskiness of the investment.
Rawpixel/Unsplash
A much higher price paid — one that is not always apparent — is the personal data these individuals are giving away. These products collect everything from biographical data to employment histories to mobile money transactions, and many consumers are unaware of the possible ramifications, including identity theft and financial fraud.
Beyond the dangers of losing control of one’s personal data, the aggregated data of millions of individuals may pose an even greater risk for society, especially in developing markets, where there are few data protection laws and little regulation. Machine learning models automate decision making in scenarios as mundane as targeted advertising, but they could also lead to more exclusionary practices like filtering out female applicants for a technology role. Practitioners as well as policy makers need to collaborate to create frameworks and regulations that address potential biases and ensure that AI doesn’t widen existing socio-economic gaps.
“Beyond the dangers of losing control of one’s personal data, the aggregated data of millions of individuals may pose an even greater risk for society…”
In the case of peer-to-peer lending, AI is widely marketed positively as a way to empower those at the bottom of the pyramid to raise their living standards. At best, it can help individuals grow their small businesses and meaningfully work their way toward financial stability. But giving these populations access to multiple lines of credit without offering financial education also has a flip side: It is not uncommon to find individuals in developing markets with outstanding repayments on more than one of these applications. Individuals could easily find themselves in growing debt, taking on additional loans to repay existing ones.
It is also apparent that these applications are owned by those who have traditionally been at the top of the pyramid. Though their biggest markets are the developing world, Branch has its headquarters in San Francisco while Tala is based in Santa Monica.
Artificial intelligence is poised to permeate and disrupt many different industries. But in order for the overall effects to be truly beneficial for everyone, particularly in the developing world, individuals need to better understand its power; communities need to be empowered with technical knowledge to build their own solutions to their problems; and multidisciplinary teams of practitioners and policy makers need to collaborate to interrogate and mitigate the potential negative consequences.
|
Traditional Finance is Failing the World’s Poor. Can AI Help?
| 24
|
traditional-finance-is-failing-the-worlds-poor-can-ai-help-17c685c4fadd
|
2018-06-26
|
2018-06-26 14:16:49
|
https://medium.com/s/story/traditional-finance-is-failing-the-worlds-poor-can-ai-help-17c685c4fadd
| false
| 902
|
Our goal is to foster greater participation and representation in AI so that the technologies we create work for all of us, not just a privileged few.
| null | null | null |
Ada-AI
|
madeline@ada-ai.org
|
ada-ai
|
AI,ARTIFICIAL INTELLIGENCE,INCLUSION,EMERGING TECHNOLOGY
|
GoAda_AI
|
Fintech
|
fintech
|
Fintech
| 38,568
|
Kathleen S.
|
feminist | data scientist | nairobi data science and machine learning community organizer | writer of code and prose | dreamer and doer of impossibilities
|
52d5dc388cf2
|
kathleensiminyu
| 31
| 39
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-08-27
|
2018-08-27 03:16:24
|
2018-08-27
|
2018-08-27 03:16:33
| 0
| false
|
en
|
2018-08-27
|
2018-08-27 03:16:33
| 1
|
17c8ca161a7c
| 2.196226
| 0
| 0
| 0
|
FREE [P.D.F] How to Create a Mind: The Secret of Human Thought Revealed READ ONLINE By Ray Kurzweil
Link…
| 1
|
Download pdf Online Doctor Who Character Encyclopedia By Jason Loborik PDF #book
FREE [P.D.F] How to Create a Mind: The Secret of Human Thought Revealed READ ONLINE By Ray Kurzweil
Link https://initialepubkindle.icu/?q=How+to+Create+a+Mind%3A+The+Secret+of+Human+Thought+Revealed
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Read Online PDF How to Create a Mind: The Secret of Human Thought Revealed, Download PDF How to Create a Mind: The Secret of Human Thought Revealed, Download Full PDF How to Create a Mind: The Secret of Human Thought Revealed, Download PDF and EPUB How to Create a Mind: The Secret of Human Thought Revealed, Read PDF ePub Mobi How to Create a Mind: The Secret of Human Thought Revealed, Reading PDF How to Create a Mind: The Secret of Human Thought Revealed, Read Book PDF How to Create a Mind: The Secret of Human Thought Revealed, Read online How to Create a Mind: The Secret of Human Thought Revealed, Download How to Create a Mind: The Secret of Human Thought Revealed Ray Kurzweil pdf, Download Ray Kurzweil epub How to Create a Mind: The Secret of Human Thought Revealed, Read pdf Ray Kurzweil How to Create a Mind: The Secret of Human Thought Revealed, Download Ray Kurzweil ebook How to Create a Mind: The Secret of Human Thought Revealed, Read pdf How to Create a Mind: The Secret of Human Thought Revealed, How to Create a Mind: The Secret of Human Thought Revealed Online Download Best Book Online How to Create a Mind: The Secret of Human Thought Revealed, Read Online How to Create a Mind: The Secret of Human Thought Revealed Book, Read Online How to Create a Mind: The Secret of Human Thought Revealed E-Books, Read How to Create a Mind: The Secret of Human Thought Revealed Online, Read Best Book How to Create a Mind: The Secret of Human Thought Revealed Online, Read How to Create a Mind: The Secret of Human Thought Revealed Books Online Download How to Create a Mind: The Secret of Human Thought Revealed Full Collection, Download How to Create a Mind: The Secret of Human Thought Revealed Book, Read How to Create a Mind: The Secret of Human Thought Revealed Ebook How to Create a Mind: The Secret of Human Thought Revealed PDF Read online, How to Create a Mind: The Secret of Human Thought Revealed pdf Download online, How to Create a Mind: The Secret of Human Thought Revealed Read, Download How to Create a Mind: The Secret of Human Thought Revealed Full PDF, Read How to Create a Mind: The Secret of Human Thought Revealed PDF Online, Read How to Create a Mind: The Secret of Human Thought Revealed Books Online, Read How to Create a Mind: The Secret of Human Thought Revealed Full Popular PDF, PDF How to Create a Mind: The Secret of Human Thought Revealed Read Book PDF How to Create a Mind: The Secret of Human Thought Revealed, Read online PDF How to Create a Mind: The Secret of Human Thought Revealed, Download Best Book How to Create a Mind: The Secret of Human Thought Revealed, Read PDF How to Create a Mind: The Secret of Human Thought Revealed Collection, Read PDF How to Create a Mind: The Secret of Human Thought Revealed Full Online, Read Best Book Online How to Create a Mind: The Secret of Human Thought Revealed, Download How to Create a Mind: The Secret of Human Thought Revealed PDF files
|
Download pdf Online Doctor Who Character Encyclopedia By Jason Loborik PDF #book
| 0
|
download-pdf-online-doctor-who-character-encyclopedia-by-jason-loborik-pdf-book-17c8ca161a7c
|
2018-08-27
|
2018-08-27 03:16:33
|
https://medium.com/s/story/download-pdf-online-doctor-who-character-encyclopedia-by-jason-loborik-pdf-book-17c8ca161a7c
| false
| 582
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
burt
| null |
a6b925ad09c4
|
burt_41998
| 0
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-04-12
|
2018-04-12 12:26:06
|
2018-04-12
|
2018-04-12 12:44:46
| 2
| false
|
ko
|
2018-04-12
|
2018-04-12 12:44:46
| 1
|
17c988513f8c
| 1.695333
| 0
| 0
| 0
|
PCA
| 2
|
[1D1D] Covariance Matrix
PCA
고차원 데이터에서 패턴을 찾는 도구
주로 인공 신경망에 입력하기 전 전처리 과정에서 사용
데이터의 범위를 재조정하고, 데이터의 평균을 0으로 맞춰줌으로써 고차원 데이터 중 중요한 차원을 골라줌
PCA는 결국 데이터에서 의미있는 선 혹은 축을 찾는 과정
데이터에는 차원 수 만큼 주성분이 존재하는데, PCA는 어떤 축이 더 중요한지 우선 순위를 구하는 것
아래 그림에서는 우상향하는 선, 분산이 가장 큰 방향이 첫번째 주성분
두번째 주성분은 첫번째 주성분으로 표현할 수 없는 축 중에서 가장 분산이 큰 축이다.
Covariance Matrix (공분산 행렬)
각 데이터의 분산과 공분산을 이용해 만드는 정방행렬
공분산 행렬에서 고유벡터와 고유값을 찾는 것은 데이터 주성분을 찾는 것과 동일하다?
고유 벡터는 행렬이 어떤 방향으로 힘을 가하는지 표현
== 어떤 방향으로 분산이 가장 큰지 구하는 것
고유값은 고유벡터에 해당하는 상관계수일 뿐임
결과적으로 고유값이 큰 순서대로 고유 벡터를 정렬하면? 주성분
좌측 하단과 우측 상단 성분의 값은 동일할 수 밖에 없는데 이는 x와 y간의 공분산을 표현하기 때문임
좌측 상단은 x의 분산, 우측 하단은 y의 분산
결과적으로 대각선 성분을 중심으로 대칭이 됨
아래 그림처럼 공분산이 양수이면 우상향, 음수이면 우하향
x의 분산이 큰 경우와 y의 분산이 큰 경우도 데이터 분포 차이가 남
참조
https://deeplearning4j.org/kr/eigenvector#code
데이터 셋에서 주성분을 구하는 것이 어떤 의미인지는 알겠으나
행렬에서 고유 벡터가 가하는 힘이 있다는 것은?
공분산 행렬의 의미를 알면 데이터 분포을 파악할 수 있다.
|
[1D1D] Covariance Matrix
| 0
|
1d1d-covariance-matrix-17c988513f8c
|
2018-04-12
|
2018-04-12 12:44:47
|
https://medium.com/s/story/1d1d-covariance-matrix-17c988513f8c
| false
| 203
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
peter_yun
| null |
5ef79ac236b6
|
peteryun
| 130
| 34
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-09-25
|
2018-09-25 07:31:08
|
2018-09-25
|
2018-09-25 07:33:08
| 4
| false
|
en
|
2018-09-25
|
2018-09-25 07:34:29
| 1
|
17c9d97e226e
| 5.183019
| 0
| 0
| 0
|
It is an exciting time to be a designer. Emerging technologies, from machine learning to artificial intelligence and self-driving vehicles…
| 5
|
A process model for designing autonomous vehicles
It is an exciting time to be a designer. Emerging technologies, from machine learning to artificial intelligence and self-driving vehicles, make us question what we know about interfaces. We are still at the very beginning of the AI journey and do not yet fully understand the impact of our design decisions. As Josh Lovejoy, user experience designer at Google, recently put it: “No-one really knows how to do that UX for AI thing yet. (…) We know just enough to know that the space is in desperate need of multi-disciplinary collaboration.”
At Stanley Robotics, we fully embrace problem-solving through design thinking and empathic design. Engineers, service operators and product managers are the experts. It is crucial to understand their way of reasoning and perception of the system and we have found it especially important to integrate them into the design process of each of our projects. As a result, we have been busy testing popular methodologies and have adapted them to the particularities of designing for self-driving vehicles and artificial intelligence.
Be it the User-Centred Design Cycle or the Double Diamond, they all follow, more or less, the same pattern. First, we research with the aim of understanding the users, their needs and viewpoint. Then, a problem space is defined. Finally, we sketch ideas, prototype and test before we start anew and iterate.
While they are user-centered, which is a good thing, they are also designer-centred and often fail to bring forward methods which work well in cross-functional teams. Even though they are business-centric, we have found that thinking in terms of Google’s infamous Design Sprints proved to be a good place to start building our own design process model.
Google Design Sprint is a five-phase-model answering product-related business questions through design thinking.
The difference in our own design process is that there are phases which are specifically intended for designers, engineers, service operators, developers and product managers. By that we mean that we do not only interact with them, but they actively take part in the design process. In addition, we are an agile company. Using a model which is designed for iterative processes is critical.
Dark blue phases are executed by all parties involved, whereas light blue ones are designer-only phases.
During the first phase, just like in any other model, we aim to understand the problem space. We hold interviews and observe people when they are working. However, we also hold what we call a kick-off workshop. Participants come from all teams involved; through a number of exercises, we make sure that the problem space, as the defined by us designers, aligns with everyone’s needs. One notable type of exercise which has delivered exceptional results is what we call “Draw It Like I See It”. Before the workshop, we come up with 2–3 topics we would like to clarify and dig into deeper. Similar to Crazy Eights, the goal is to draw; however, we ask people to draw how, in their mind, a certain functionality works. This might be an interface or a component, but it does not have to be. The goal is simply to foster understanding throughout the team. Equally, for established interfaces, we have started to carry interface kits to all our meetings which include all interface components in paper form. This way, components can easily be moved around, drawn on, cut up and reassembled. They help communicate ideas quickly.
The interface kits
Now one might ask: why do we need to work with engineers? What is different in autonomous vehicles and artificial intelligence compared to traditional systems? When someone from the technical world would describe programming to someone with little knowledge about it, they could describe it as ‘using a language to tell a computer or a machine what to do’. Even if abstract, this is understandable.
The difference with artificial intelligence is that it is not only the developer who tells the machine what to do, but the machine starts to take decisions on its own. A rather complex concept, raising a lot of questions: How does that work? In how far is the machine autonomous? What if things go wrong?
While users work with a system, they build up what designers and psychologists call a ‘mental model’. It is a user’s own explanation of how a system works. It helps them interact with the system and is used to decide on what to do when things go right, but especially when things go wrong. It is important to understand what a user’s mental model looks like, not just when designing for artificial intelligence or self-driving vehicles, but for any system.
It is only then that we can start developing interfaces which are tailored to the needs of a user. Engineers and service operators, having full understanding of how everything is interconnected, can then help the designer translate that knowledge to a ‘conceptual model’, the picture that the designer would like to paint of the system and which helps users to build a mental model that works well.
Once we know how engineers see the system behind the interface we are developing, we are more concerned with our users. The second and third phases focus therefore on what we as designers do best: ideate, evaluate and prototype iteratively. We use the usual techniques to do so: sketching, wireframing, paper and medium-fidelity prototyping. With so many different user groups, it might become difficult to see how they differ from one another. To align their goals we have become big supporters of job stories.
For those familiar with user stories, they are a tool for describing the goals of a user in an interface. Job stories are different in that they focus on the situation, motivation and expected outcome, rather than extraneous details.
Finally, in phase four, we test our interface with both engineers and service operators. We tend to give them a task, as discovered during the first phase, to go through using our interface. As they use the interfaces in different situations, it is always interesting to see what happens when they are presented with the same task. Does the conceptual model correspond to both mental models? To what extent are they different? Where does this difference originate? The answers let us iterate again and again until we find a solution that ticks all the right boxes.
When designing an interface for a system which is built on artificial intelligence, it is tempting to start with one interface and work one’s way through the others. However, if there is one piece of advice that we can give you: do not go down that road. Instead, take your time to gain a holistic picture of what the task of each interface is and try to understand how the system works as a whole. The idea is to build one’s own mental model the same way an engineer or other related parties would, leading to deeper understanding, better products and results — right from the beginning.
Originally published at stanley-robotics.com.
|
A process model for designing autonomous vehicles
| 0
|
a-process-model-for-designing-autonomous-vehicles-17c9d97e226e
|
2018-09-25
|
2018-09-25 07:34:29
|
https://medium.com/s/story/a-process-model-for-designing-autonomous-vehicles-17c9d97e226e
| false
| 1,188
| null | null | null | null | null | null | null | null | null |
Design
|
design
|
Design
| 186,228
|
Stanley Robotics
|
Thanks to our robotic valet parking a place to park guaranteed everywhere anytime. 👉 http://stanley-robotics.com
|
6bf2d5f8566c
|
Stanleyrobotics
| 2
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-07-24
|
2018-07-24 12:12:54
|
2018-07-23
|
2018-07-23 09:00:27
| 2
| false
|
en
|
2018-07-24
|
2018-07-24 12:24:57
| 9
|
17cde9158f8f
| 2.734277
| 0
| 0
| 0
|
According to OSHA, 18% of violent crimes occur in the workplace. So, what are you doing to minimize threats and strengthen your business…
| 5
|
Deep Learning and Security: How Does It Impact Your Business?
According to OSHA, 18% of violent crimes occur in the workplace. So, what are you doing to minimize threats and strengthen your business security?
Security experts have been working hard on developing new solutions to help you reinforce your physical security and keep your company safe. Through technological advancements, deep learning has changed the dynamics of commercial security systems, specifically video analytics, to ensure your business protected.
What is deep learning?
Computers don’t just get smarter on their own. In fact, there’s a science behind it. Machine learning is designed to make computers “smarter” by feeding them information and data, much like humans learn. When you read a book and consume information, you’re making yourself more intelligent. Same goes for computers.
Deep learning takes machine learning techniques and specializes them to help computers learn by example. For example, when you shop on Amazon or watch your favorite programs on Netflix or Hulu, how do you think they can make predictions about what item or show you may be interested in? It’s all about deep learning. But this technology isn’t just for online shopping or streaming shows. It can also be integrated into your business security systems for maximum protection.
How deep learning facilitates business security
Over the years, business security system technology has become extremely advanced; allowing for sophisticated integrated security solutions and comprehensive security analytics. However, where deep learning and security excel is analyzing video security footage.
Deep learning and security go hand in hand, especially when it comes to analyzing video surveillance footage. Rather than spending countless hours reviewing the footage or risking a false alarm, deep learning can identify these incidents and eliminate human error — saving your business time and money.
Here are some other benefits deep learning can offer your company:
Time efficiency: Deep learning and cut the processing and analysis of video surveillance footage in half and convert it into information you can use to keep your business protected.
Object detection and face recognition: If you’ve ever uploaded photos to Facebook, then you’re familiar with facial recognition. Your business security system can use the same deep learning technology to identify faces in video surveillance footage, even in a crowd. The same can be done with objects captured on security video.
Accuracy: False alarms can be a problem for any business not utilizing smart, integrated security solutions. Not only do they cause increased police response times due to low priority, but they’re also responsible for low apprehension rates which can cost businesses a ton of money. Deep learning provides higher accuracy rates and eliminates false alarms — making your company a priority and allowing you to recoup any losses.
Implementing a smart commercial security system
Companies of all sizes are prone to attacks of all kinds. That’s why it’s essential to invest in intelligent security technology to keep your business safe. Technological advancements, such as deep learning, allow commercial security companies to provide businesses, such as yours, with the best options to protect their property and employees.
Have you invested in your business security?
Whether you’re considering a new security system or want to update an existing one, make sure you find the right fit. You’ll want to work with a security company that provides reliable, trusted products and services, can assess your vulnerabilities and can offer a comprehensive approach to keeping your business safe. Consider what it will cost you to install or update a system and what it will cost you if you don’t. Your security is an investment — one worth making. So start researching today.
Originally published at source1security.com on July 23, 2018.
https://source1security.com/2018/07/23/deep-learning-and-security-what-does-it-mean-for-your-business/
|
Deep Learning and Security: How Does It Impact Your Business?
| 0
|
deep-learning-and-security-how-does-it-impact-your-business-17cde9158f8f
|
2018-07-24
|
2018-07-24 12:24:58
|
https://medium.com/s/story/deep-learning-and-security-how-does-it-impact-your-business-17cde9158f8f
| false
| 623
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Source 1 Solutions
|
Source 1 Solutions offers maximum protection to companies by providing top-rated business security systems and services in Florida and throughout the US.
|
50bd297ad1f3
|
source1security
| 2
| 11
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-07-02
|
2018-07-02 01:52:47
|
2018-07-06
|
2018-07-06 14:31:13
| 0
| false
|
en
|
2018-07-06
|
2018-07-06 14:31:13
| 1
|
17ce56b0134c
| 4.041509
| 2
| 0
| 0
|
What if the power of AI could help us tackle complex global issues from the inside out, solving problems of unequal distribution of…
| 5
|
Cutting through the legal jargon — Ritika Dutt on how AI can help navigate the system
What if the power of AI could help us tackle complex global issues from the inside out, solving problems of unequal distribution of resources and access to basic needs? Botler AI Co-founder Ritika Dutt believes that the bots are here to serve us by doing more to create systemic change than is humanly possible.
Her company makes the legal and justice systems more accessible and simple to navigate, allowing everyone to become better versed on their rights and legal situation so they can make educated decisions on how to take crucial next steps.
Q: We enjoy fairly uncomplicated access to basic education, health care and other institutional services essential to navigating the modern world. Yet legal help is still often shrouded in jargon, bureaucracy and high prices that make it intimidating to many and inaccessible to more vulnerable, marginalized members of our society. What drew you to address this problem in particular?
RD: As an economist, I believe that the law is a public good, and should therefore be accessible to everyone for free. However, it is not, which is an issue that affects multiple echelons of society. Essentially, unless you have deep pockets to pay for your legal needs, you’re more likely than not to be materially impacted.
The average person does not necessarily easily understand the law or their rights. In order to access this complex information and understand it better, they need to hire lawyers. Often, people (myself included) find that by the time they get to a point where they need a lawyer to go to court or take a definitive action, they’ve already maxed out their budget.
Q: You grew up in Hong Kong and Singapore before immigrating on your own to Canada, you founded the marketing department of a Y Combinator startup from scratch right after graduating from McGill University, you’ve had your share of navigating some complicated systems. How much of your personal experience is behind Botler AI’s mission?
RD: I dealt with a couple of instances after moving to Canada where I had to navigate the legal situation on my own and it was an absolutely awful experience. The sheer contrast and disparity between what I was used to as a child in Singapore and how it is in the “real world,” where I had to fend for myself as an adult in Canada with no safety cushion, really opened my eyes to what the majority of people go through when dealing with the legal system.
With the immigration issue, I went from having a regular life with a paying job to incomeless literally overnight. I had to start rationing out cents and live off toast and instant noodles so that I could make ends meet while I waited for my permanent residency to come through and the issue to be dealt with.
The other personal experience: I found myself with a stalker. It kept getting worse and I still remember waking up every morning filled with fear and dread that I would have to go into work and deal with him. Some of the people around me were amazing and so supportive, while others blamed it all on me and said there was nothing wrong. I really had no idea what my rights were. It was only a few months later, when I was far removed from the situation and able to think rationally, that I found out that his behaviour had constituted criminal harassment (stalking) which was a direct violation of the Canadian Criminal Code.
One of the recurring issues was that many people, myself included, really did not know what constituted actual harassment, or what their rights were in such a situation. I saw the potential for the technology we’d developed at Botler to be applied to criminal laws regarding sexual misconduct and decided to build the tool as a free resource for others in similar circumstances. The idea is to give users a completely anonymous first layer of help, which will hopefully help them be better informed when making a decision on how to deal with the issue.
Q: How do you feel about the reception that Botler has received, not only in North America but internationally?
RD: When you set your mind to tackling a problem, it’s because you genuinely believe you have a novel approach that can help improve the way things currently work. When we started to get a positive reception, it was an incredible stamp of approval that what we built has directly impacted someone’s life for the better. There’s really nothing more that I can ask for.
With the harassment tool, it’s part of the zeitgeist now and it’s really the first time that a lot of people, both men and women, have been able to openly talk about what they’ve dealt with. We’ve had over 30,000 users from around the world in just over 5 months since the launch. Though it’s only available in Canada and the US for now, people from all over the world have been sending us emails asking us to bring it to their country next. As an example, earlier this year we had not one but two different police officers from South Korea reach out to us to request having the tool translated to Korean so they could use it as part of their police work. They were from different cities and had no knowledge of each other, they’d just read about Botler independently and reached out!
It really gives me immense motivation and joy to witness how Botler was able to help someone in a particular situation. Although, if I’m being completely honest, some of these are also so heart wrenching, they keep me up at night because I can’t begin to imagine what they had to live through.
Q: What excites you the most about the future of digital social entrepreneurship?
RD: One of my best friends has been very involved with social entrepreneurship for a long time, but I never imagined that I would also get into the field through the use of technology. Alot of the SE projects she would discuss were very traditional in their approach, such as microfinance cash loans. What I’m most looking forward to is seeing how innovative social entrepreneurs can apply these transformative technologies to tackle these same traditional problems and opportunities!
|
Cutting through the legal jargon — Ritika Dutt on how AI can help navigate the system
| 2
|
cutting-through-the-legal-jargon-ritika-dutt-on-how-ai-can-help-navigate-the-system-17ce56b0134c
|
2018-07-06
|
2018-07-06 14:31:13
|
https://medium.com/s/story/cutting-through-the-legal-jargon-ritika-dutt-on-how-ai-can-help-navigate-the-system-17ce56b0134c
| false
| 1,071
| null | null | null | null | null | null | null | null | null |
Police Brutality
|
police-brutality
|
Police Brutality
| 4,294
|
Startupfest
|
An international gathering of entrepreneurs, founders, investors, and mentors from around the world. July 10–14, 2018 in Montreal. | www.startupfest.com
|
1f7371d39161
|
startupfest
| 2,873
| 1,862
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-08-12
|
2018-08-12 04:24:49
|
2018-08-12
|
2018-08-12 04:26:20
| 0
| true
|
ja
|
2018-08-12
|
2018-08-12 04:26:20
| 0
|
17cef69a93e5
| 0.262
| 0
| 0
| 0
|
新しい動画を投稿しました。
自然言語処理で単語をベクトル化することのできる word2vec のライブラリと学習済みモデルを用いて類似した映画のあらすじを探してみます。
| 5
|
自然言語処理を学ぼう #4 word2vec 実装編
新しい動画を投稿しました。
自然言語処理で単語をベクトル化することのできる word2vec のライブラリと学習済みモデルを用いて類似した映画のあらすじを探してみます。
#機械学習 #人工知能 #自然言語処理 #word2vec
|
自然言語処理を学ぼう #4 word2vec 実装編
| 0
|
自然言語処理を学ぼう-4-word2vec-実装編-17cef69a93e5
|
2018-08-12
|
2018-08-12 04:26:49
|
https://medium.com/s/story/自然言語処理を学ぼう-4-word2vec-実装編-17cef69a93e5
| false
| 12
| null | null | null | null | null | null | null | null | null |
Japanese
|
japanese
|
Japanese
| 17,110
|
Regonn
|
Web Programmer. Rails, Julia, Haiku.
|
857ffb56ea37
|
regonn_haizine
| 22
| 25
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
ead6b775b343
|
2018-02-21
|
2018-02-21 11:36:19
|
2018-02-21
|
2018-02-21 11:50:06
| 3
| true
|
en
|
2018-10-02
|
2018-10-02 08:44:41
| 3
|
17d0d42cdf25
| 3.070755
| 1
| 0
| 0
|
Researchers from music and engineering teams are working together to turn big data into sound.
| 5
|
What Does Big Data Sound Like?
Researchers from music and engineering teams are working together to turn big data into sound.
A collaboration between two professors — one of music and one of engineering — at Virginia Tech resulted in the creation of a new platform for data analysis that makes it possible to understand data better by turning it into sound.
This is a pioneering approach to studying spatially distributed data. Instead of placing information into a visual context to show patterns or correlations — meaning, data visualization — the work will use an aural environment to leverage the natural affordances of the space and the user’s location within the sound field.
Funded by the National Science Foundation, the work combines elements of music, geospatial science, computer science, and human-computer interaction. It’s the first time a research project led by a faculty member from the university’s School of Performing Arts working in collaboration with the College of Engineering
Ivica Ico Bukvic, associate professor of composition and multimedia in the College of Liberal Arts and Human Sciences, and Greg Earle, professor of electrical and computer engineering, used infrastructure at the Institute for Creativity, Arts, and Technology to investigate how immersive sound can be used to develop our understanding of complex systems.
According to Bukvic, data sonification — which involves converting non-auditory information into sound, is a relatively unexplored area of research, yet provides a unique perspective for exploring data. The human auditory system has a superior ability to recognize temporal changes and patterns, making sonification a powerful tool for studying complex systems.
“Identifying new time and space correlations between variables often leads to breakthroughs in the physical sciences,” explained Dr. Bukvic, who also serves as a senior fellow for the Institute for Creativity, Arts, and Technology. “It makes sense that we would want to go beyond two-dimensional graphical models of information and make new discoveries using senses other than our eyes.
Titled “Spatial Audio Data Immersive Experience (SADIE),” the project is the first large-scale endeavor focusing on immersive spatially-aware sonified data using a high-density loudspeaker array. The research will focus on the earth’s upper atmospheric system, which features physical variables that are spatially and temporally rich. Each of the data sets associated with this system will be represented by distinct sound properties, such as amplitude, pitch, and volume.
These sounds will be played through a 129-loudspeaker spatially distributed immersive sound system in the Cube, located in the Moss Arts Center. A combination of performance space, research laboratory, and studio, the Cube is a collaborative research facility at Virginia Tech where researchers, composers, and musicians are uncovering new possibilities in immersive sound.
Using the Cube’s motion capture system, similar to the interface imagined for the film Minority Report, participants will be able to navigate the sonified data using a gesture-driven interface, allowing them to rewind, fast-forward, rotate, zoom, amplify, speed up, and slow down the data playback. The system will also be used to capture user study data.
“Allowing the brain’s innate signal processing mechanisms to identify specific features in complex data sets is a logical way to link computational sciences with human sensory perceptions. This merging of technology and nature could further current analysis techniques and foster new breakthroughs involving complex systems in science, with the potential to produce new technologies designed to spur creativity,” concludes Dr. Bukvic, who says that if this approach to experiencing data can be proven to improve people’s understanding of complex relationships in physical systems, it could be applied to other fields of study: “It could have applications to fields such as thermodynamics, quantum mechanics, and aeronautical engineering; help advance visualizations and virtual reality systems; and create interdisciplinary bridges between scientific communities, including music, computing, and the physical sciences.”
Alice Bonasio is a VR Consultant and Tech Trends’ Editor in Chief. She also regularly writes for Fast Company, Ars Technica, Quartz, Wired and others. Connect with her on LinkedIn and follow @alicebonasio on Twitter.
|
What Does Big Data Sound Like?
| 1
|
what-does-big-data-sound-like-17d0d42cdf25
|
2018-10-02
|
2018-10-02 08:44:41
|
https://medium.com/s/story/what-does-big-data-sound-like-17d0d42cdf25
| false
| 668
|
Showcase for the latest disruptive technology that is changing the education landscape globally
| null |
EdTechTrends
| null |
Tech Trends
|
alicebonasio@gmail.com
|
edtech-trends
|
TECHNOLOGY,EDUCATION,VIRTUAL REALITY,TECH,STARTUP
|
alicebonasio
|
Big Data
|
big-data
|
Big Data
| 24,602
|
Alice Bonasio
|
Technology writer for FastCo, Quartz, The Next Web, Ars Technica, Wired + more. Consultant specializing in VR #MixedReality and Strategic Communications
|
1663ada6f01e
|
alicebonasio
| 7,865
| 608
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-17
|
2018-06-17 06:02:40
|
2018-06-17
|
2018-06-17 06:29:05
| 9
| false
|
en
|
2018-06-17
|
2018-06-17 06:29:05
| 4
|
17d1fdb3ac97
| 4.845283
| 3
| 0
| 0
|
The Gabor filter, named after Dennis Gabor, is a linear filter used in myriad of image processing application for edge detection, texture…
| 5
|
Through The Eyes of Gabor Filter
The Gabor filter, named after Dennis Gabor, is a linear filter used in myriad of image processing application for edge detection, texture analysis, feature extraction, etc. The characteristics of certain cells in the visual cortex of some mammals can be approximated by these filters. These filters have been shown to possess optimal localization properties in both spatial and frequency domain and thus are well suited for texture segmentation problems. Gabor filters are special classes of band pass filters, i.e., they allow a certain ‘band’ of frequencies and reject the others. A Gabor filter can be viewed as a sinusoidal signal of particular frequency and orientation, modulated by a Gaussian wave. One such 2D Gabor filter is shown in the figure 1.
Figure 1: A 2-D Gabor filter obtained by modulating the sine wave with a Gaussian
From the above figure we can notice that the sinusoid has been spatially localized. In practice to analyze texture or obtain feature from image, a bank of Gabor filter with number of different orientation are used.
Consider an example of elephant which has pattern or stripes on its skin at different orientation. Now to highlight or extract out all those patterns we are going to use a bank of 16 Gabor filters at an orientation of 11.250 (i.e. if the first filter is at 00, then the second will be at 11.250, the third will be at 22.50, and so on.).The figure 2 shows all the filter bank of 16 filters
Figure 2: A bank of 16 Gabor filter oriented at an angle of 11.250 (i.e. if the first filter is at 00, then the second will be at 11.250, the third will be at 22.50, and so on.)
When the input image is convolved with all the Gabor filters the patterns are easily highlighted as shown in figure 3. When a Gabor filter is applied to an image, it gives the highest response at edges and at points where texture changes. When we say that a filter responds to a particular feature, we mean that the filter has a distinguishing value at the spatial location of that feature.
Figure 3: (a) The input image of an Elephant and (b) the output image after passing it through the Gabor filter bank
To better understand what each filters detect in input image, consider a simple white circle in black background. When this image is passed through each filter in the filter bank, the edge of the circle which gets detected is the edge oriented at an angle at which the Gabor filter is oriented. The Figure 4 clearly shows this.
Figure 4: The corresponding oriented edge features being detected when passed through individual oriented Gabor filters
Different parameters that controls the shape & size of 2D Gabor filter:
There are certain parameters that controls how Gabor filter will be and which features will it respond to. A 2D Gabor filter can be viewed as a sinusoidal signal of particular frequency and orientation, modulated by a Gaussian wave. The filter has a real and an imaginary component representing orthogonal directions. The two components may be formed into a complex number or used individually. The equations are shown below:
In the above equation,
λ — Wavelength of the sinusoidal component.
Ө — The orientation of the normal to the parallel stripes of Gabor function.
Ψ — The phase offset of the sinusoidal function.
σ — The sigma/standard deviation of the Gaussian envelope
ɣ — The spatial aspect ratio and specifies the ellipticity of the support of Gabor function.
The above mentioned five parameters controls the shape and size of the Gabor function. The role of each parameter is discussed in detail below.For illustration of the effects of parameters, following values are chosen as starting point:
Lambda (λ) = 30; Theta (Ө) = 00
Gamma (ɣ) = 0.25; Sigma (σ) = 10
Psi (Ψ) = 0
Lambda (λ):
The wavelength governs the width of the strips of Gabor function. Increasing the wavelength produces thicker stripes and decreasing the wavelength produces thinner stripes. Keeping other parameters unchanged and changing the lambda to 60 and 100, the stripes gets thicker.
Figure 1: Keeping other parameters unchanged (Ө = 00, ɣ = 0.25, σ = 10, Ψ = 0), and on changing the lambda from 30 to 60 and 100 the Gabor function gets thicker
Theta (Ө):
The theta controls the orientation of the Gabor function. The zero degree theta corresponds to the vertical position of the Gabor function.
Figure 2: Keeping other parameters unchanged (λ = 30, ɣ = 0.25, σ = 10, Ψ = 0), and on changing the theta from 00 to 450 and 900 the Gabor function rotates.
Gamma (ɣ):
The aspect ratio or gamma controls the height of the Gabor function. For very high aspect ratio the height becomes very small and for very small gamma value the height becomes quite large. On increasing the value of gamma to 0.5 and 0.75, keeping other parameters unchanged, the height of the Gabor function reduces.
Figure 3: Keeping other parameters unchanged (λ = 30, Ө = 00, σ = 10, Ψ = 0), and on changing the gamma from 0.25 to 0.5 and 0.75, the Gabor function gets shorter.
Sigma (σ):
The bandwidth or sigma controls the overall size of the Gabor envelope. For larger bandwidth the envelope increase allowing more stripes and with small bandwidth the envelope tightens. On increasing the sigma to 30 and 45, the number of stripes in the Gabor function increases.
Figure 4: Keeping other parameters unchanged (λ = 30, Ө = 00 ɣ = 0.25, Ψ = 0), and on changing the sigma from 10 to 30 and 45 the number of stripes in Gabor function increases.
In OpenCV Python, following is the structure of the function that is used to create a Gabor kernel.
cv2.getGaborKernel(ksize, sigma, theta, lambda, gamma, psi, ktype)
ksize is the size of the Gabor kernel. If ksize = (a, b), we then have a Gabor kernel of size a x b pixels. As with many other convolution kernels, ksize is preferably odd and the kernel is a square (just for the sake of uniformity).
sigma is the standard deviation of the Gaussian function used in the Gabor filter.
theta is the orientation of the normal to the parallel stripes of the Gabor function.
lambda is the wavelength of the sinusoidal factor in the above equation.
gamma is the spatial aspect ratio.
psi is the phase offset.
ktype indicates the type and range of values that each pixel in the Gabor kernel can hold.
Experiment
Lambda (wavelength)- Controls the width of the strips of gabor function. Decreasing the wavelength produces thinner stripes
Theta(Orientation): governs the rotation of the gabor envelope
Gamma(aspect Ratio): controls the height of the Gabor filter. If the gamma value is the height of gabor reduces and if the gamma value is small the height of gabor increases.
Sigma (Bandwidth) : It controls the overall size of the Gabor envelope . for larger bandwidth the envelope increase allowing more stripes and with smaller bandwidth the envelope tightens.
Psi-is the phase offset of sinusoid.
|
Through The Eyes of Gabor Filter
| 3
|
through-the-eyes-of-gabor-filter-17d1fdb3ac97
|
2018-07-05
|
2018-07-05 20:16:24
|
https://medium.com/s/story/through-the-eyes-of-gabor-filter-17d1fdb3ac97
| false
| 966
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Anuj shah (Exploring Neurons)
|
On a quest to learn,share and explore Neurons
|
c7f54cd28926
|
anuj_shah
| 6
| 10
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-10
|
2018-06-10 19:48:42
|
2018-06-10
|
2018-06-10 19:49:50
| 0
| false
|
en
|
2018-06-10
|
2018-06-10 19:49:50
| 4
|
17d2728fb1cf
| 0.603774
| 0
| 0
| 0
|
The Turing Test is one of the more influential concepts in artificial intelligence; it certainly defined how we think about machine…
| 4
|
Did Google’s A.I. just beat the Turing test? Yes and no.
The Turing Test is one of the more influential concepts in artificial intelligence; it certainly defined how we think about machine learning achieving ‘artificial intelligence’ in the nascent stages of the field. And Google’s A.I. might just have beat the test with the whole world watching.
What is the Turing Test and why does it matter?
Alan Turing is the father of modern computing. His early Turing Machine — to many the first general purpose computer invented — cracked the Nazi Enigma cipher, which was one of the turning points in the Allies’ victory in WWII (it certainly turned the tide of the naval battle for the Atlantic). A slightly fictionalized version of this story made some serious Academy Award waves with Benedict Cumberbatch in the titular role of Turing in Imitation Game.
To read the full article, please visit: http://www.eno8.com/blog/did-googles-a-i-just-beat-the-turing-test/
|
Did Google’s A.I. just beat the Turing test? Yes and no.
| 0
|
did-googles-a-i-just-beat-the-turing-test-yes-and-no-17d2728fb1cf
|
2018-06-10
|
2018-06-10 19:49:50
|
https://medium.com/s/story/did-googles-a-i-just-beat-the-turing-test-yes-and-no-17d2728fb1cf
| false
| 160
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
ENO8
|
ENO8, a Dallas based custom software product and mobile app development company focused on creating impactful experiences.
|
db4fe6f378c7
|
ENO8
| 0
| 4
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-03-16
|
2018-03-16 08:41:04
|
2018-04-17
|
2018-04-17 13:03:43
| 4
| false
|
en
|
2018-04-17
|
2018-04-17 13:09:39
| 9
|
17d3b13a2417
| 4.122642
| 6
| 0
| 0
|
The ability to familiarize oneself with new topics quickly is a key competency for professionals in all innovation-related sectors, be it…
| 4
|
Some thoughts on a new project: Leveraging NLP to augment intellectual exploration routines
The ability to familiarize oneself with new topics quickly is a key competency for professionals in all innovation-related sectors, be it venture capital, corporate strategy or consulting. Dynamic market environments shaped by accelerating innovation cycles and blurring industry borders require us to intellectually explore new technologies, market sectors and industry ecosystems continuously. Following our original vision of “building technology that helps you stay in the know” we’re currently working on a software that helps users get familiar with new topics. Here are some first thoughts on it.
How does intellectual exploration work?
We spoke to many friends and colleagues working in the industries mentioned above to understand how they explore topics. The conversations revealed that intellectual exploration routines are very comparable even if the tools and workflows are highly individual. Most often, the process starts with superficial knowledge on single concepts of a topic — like bitcoin was a starting point for many people to explore the broader field of distributed ledger and blockchain. These concepts are the starting point for an iterative research process across sources like Wikipedia, YouTube, forum threads, expert media, scientific publications or company websites. The research process results first in a better understanding of the initial concepts and, secondly, in the discovery of further related concepts. Over time, we gradually recognize relationships and hierarchies, prioritize the concepts and finally gain a holistic understanding of the topic. On an abstract level, the intellectual exploration of a topic is an iterative process of
discovering underlying concepts of the topic,
gaining understanding of these concepts and
grasping the interdependencies between them.
Augmentation vs Automation
Over the past years there was a lot of buzz around the “automation of knowledge work,” aka “AI killing white-collar jobs” if you prefer a little more sensationalism. We’re deeply convinced that AI technologies, like machine learning or natural language processing, will fundamentally impact knowledge work and already do. But we also believe that augmentation is the better concept than automation to think about how these technologies will change the workflows of knowledge workers. This is especially true for the highly iterative and intuitive process of intellectual exploration where the human can’t be cut out but which offers massive potential for acceleration and objectivization.
The basic idea of our new project, 12K EXPLORER, is to accelerate and objectify intellectual exploration routines. The software is based on four key elements:
Data set: Our software is fueled by historical and continuously updated content from innovation-related sources (media, patents, academia). Our data set currently includes more than 4 million documents.
Topic model: We use NLP algorithms to transform the unstructured text content into a structured topic model of concepts. Our current topic model covers 1.5 million innovation-related concepts.
Graph interface: The graph interface is a visual interface to the topic model. It allows the user to gradually explore topics by detecting related concepts and mapping the interdependencies between concepts.
Content ranking: We developed an algorithm to detect the most relevant articles on single concepts or entire topics. The relevance ranking is based on the technology that we use for our algorithmically curated newsletter, 12K FILTER, and considers content properties like topic fit, fact density, article length, article diversity and expert shares.
V1-beta and initial learnings
To give you a first glimpse of the product, we created a topic graph for neurotech (inspired by Clement Vouillon’s and nicolas debock’s great newsletter). Our starting point was the concept “neural interfaces” from where we iteratively added related concepts to the topic graph.
The graph interface suggests related concepts based on the topic model
After a few minutes, we ended up with a nice graph covering key aspects of neurotech, namely
brain imaging / neuroimaging / eeg / fmri (🧠)
brain machine interaction / bci / brain implants (🤖)
neural stimulation / neuroprosthetics / neural implants (⚡)
transhumanism / posthumanism / AGI (🔮)
The size of the nodes indicates the attention for the given concept in innovation-related media sources. The color intensity of the nodes indicates the growth of the media attention over the last years.
Besides playing with the tool ourselves, we asked partners, friends and clients to test the beta version and use it in their workflows. Here are the key takeaways from the feedback:
The graph interface is a great tool to discover underlying concepts of a topic. Creating the neurotech graph took us about 15 minutes and definitely helped to get an overview of a topic that we knew little to nothing about before.
We developed a simple feed based on our content ranking algorithm that shows the top articles on any given concept right next to the graph. But building something that exceeds fantastic services like Wikipedia, Google or Quora in researching a specific concept is not an easy task. That’s why we will focus on the discovery of concepts and the mapping of their interdependencies in the next step.
The content ranking is however very useful in another way. We built a newsletter functionality that sends users the top articles on a topic once a week. That feature is a great help in staying up to date on various research topics.
The topic graphs are not only helpful for single users to explore topics but also a great tool to create a common understanding of topics within teams or towards clients.
The initial feedback showed us that there is potential for acceleration and objectification even in the research routines of the brightest humans. That motivated us to keep on working on the project. Get in touch if you’re interested to learn more and check out the beta: hi@12k.co.
|
Some thoughts on a new project: Leveraging NLP to augment intellectual exploration routines
| 28
|
some-thoughts-on-a-new-project-leveraging-nlp-to-augment-intellectual-exploration-routines-17d3b13a2417
|
2018-05-13
|
2018-05-13 21:50:51
|
https://medium.com/s/story/some-thoughts-on-a-new-project-leveraging-nlp-to-augment-intellectual-exploration-routines-17d3b13a2417
| false
| 907
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
12K
|
A software company operating at the intersection of machine learning and data science. Our specialty is to gather, structure and leverage data. // www.12k.co
|
4affb591d281
|
12KResearch
| 206
| 414
| 20,181,104
| null | null | null | null | null | null |
0
|
除了在 Medium 追蹤我們，也歡迎你加入集知社團，一起交流討論創意的觀點。
臉書搜尋「集知會所」或是直接透過以下連結：
https://www.facebook.com/groups/1830918583789303/
| 1
|
7c99ca16df7d
|
2018-06-08
|
2018-06-08 09:51:29
|
2018-06-08
|
2018-06-08 10:31:35
| 5
| false
|
zh-Hant
|
2018-06-27
|
2018-06-27 07:29:23
| 6
|
17d52aaa227b
| 1.184277
| 3
| 0
| 0
|
五月末一部引起科技圈熱議的影片被留出，內容是 Google X 實驗室 2016 年內部討論用的實驗，名為「自私的帳本」。
| 5
|
從「自私的帳本」看見 AI 不可見的影響力
五月末一部引起科技圈熱議的影片被留出，內容是 Google X 實驗室 2016 年內部討論用的實驗，名為「自私的帳本」。
在影片描述的未來，AI 演算法為了讓全體人類朝向更好的方向邁進，會透過不斷累積個人的數據資料，進一步影響用戶目標、思緒、行為。最終藉由引導群體行為的改變，解決氣候、貧困、疾病等巨大的全球問題。
一開始系統會提供簡單的良性價值給使用者選擇，例如減肥、環保。當使用者選好目標後，每次互動時 Google 都會引導你朝向目標邁進。
比如購買有機的本地產品、選擇環保交通工具，逐漸誘導你的行為。
系統會推薦符合設定目標的選擇
更極端的是，當用戶的資料有所缺失時，這套系統會自動想辦法得到。例如發現缺少使用者的體重資料，就會配合你的喜好，推薦你購買體重器或乾脆 3D 列印一台給使用者，以取得更完善的資料。
透過分析你的喜好，再投你所好，獲得更多的數據。
藉由代代相傳的用戶數據，新用戶可以從上一代的行為和決策中受益，進而讓全體朝向更好的方向邁進。
超級數據庫打造完美世界，這樣的情節在科幻電影中屢見不鮮。然而當提出這項實驗的單位，是現實中掌握大量使用者資料的 Google ，這樣的情節恐怕驚悚大過於有趣。
電影中為了方便情緒渲染，我們總是把 AI 視為「擬人」、「可見」的威脅。然而實際上很可能就如同「自私的帳本」影片描述，使用者自以為擁有生活中的選擇權，殊不知一切都在 AI 的計算之下。
如果 AI 能夠計算哪些行為能選出下一任總統、推動某個政策、引發某一場抗爭，你無法得知生活中影響你決策的資訊，到底是巧合的投放，還是蓄意的引導，讓你配合大企業的意志改變行為。
以下進一步解釋影片中的概念
自私的帳本一詞（The Selfish Ledger），引用自著名生物學著作《自私的基因》（The Selfish Gene）。影片結合了兩個觀點，分別是：
拉馬克之用進廢退說
用進廢退，意思是「頻繁的使用會讓該部位更發達，反之就會退化」。例如長頸鹿為了吃到更高的樹葉，拉長脖子，導致下一代長頸鹿的脖子更長。該假說認為生物的基因編碼是動態的，身體特徵可能為了適應環境而改變，並遺傳給下一代
長頸鹿為了吃到更高的樹葉，拉長脖子，導致下一代長頸鹿的脖子更長
道金斯之基因本位進化論
強調生物只是基因的載體，就像駕駛人與巨大機器人的關係。物種的生存競爭只是一種「代理戰爭」，生物只是基因暫時的保管者，競爭的目標是為整體誕生更優秀的基因。
雖然「用進廢退說」在生物學上已經被「達爾文進化論」擊敗，但 Google X 提出了一個假設，如果今天把這樣的觀點套用在使用者數據上呢？如果全人類的數據，能被視作一個「大型的基因帳本」，科技產品不是為了滿足個人的需求，而是為了讓全人類邁向更好的方向，進而改變人類行為。
白話一點，其實就是
透過大型資料庫，
促使人們世世代代「用進廢退」
爭議點在於，這些「良性價值」是由 Google 所認定，並不一定適用於所有人，且很有可能造成預想之外的後果，例如讓圖利少數人的意志被包裹在「整體向善」的糖衣裡。
雖然 Google 已出面表示這只是一項思想實驗，無論是影片的拍攝方法或內容，刻意引起不安正是他的目的。但 Google 的地位實在太過敏感，也因此引起了大眾的質疑。
相關報導：
Google版《黑鏡》真實上演，全面搜集數據改變人類行為｜數位時代
Google一支內部影片「自私的帳本（The Selfish Ledger）」外流，描述透過個人不斷累積的數據資料，進一步影響用戶行為，甚至引導群體行為的改變，宛如科幻影集《黑鏡》情節。而這種使用數據的手法，會是我們的未來嗎？www.bnext.com.tw
Google「令人不安」內部影片外洩 發言人強調只是思想實驗 - INSIDE 硬塞的網路趨勢觀察
跟 Facebook 相比，Google 在網絡上更加無處不在，要脫離 Facebook、WhatsApp 及 Instagram 已不容易，要擺脫 Google（搜尋器）、Gmail、Google Drive、Google…www.inside.com.tw
更多 AI 主題文章，請見
本月主題：人機 合作/競爭 時代
AI 人工智慧一直是科幻題材的寵兒，
然而當 Google 的語音助理，已經能夠模仿真人語氣打電話訂位，
震撼的不只科技同業，大眾也重新意識到過去的幻想，已經開始取代真實。
超級電腦控制世界看似遙遠，但現實生活早已被滲透。…medium.com
本文由集知編輯整理校閱
如果你喜歡這篇文章，請為我們拍手不只一下吧
|
從「自私的帳本」看見 AI 不可見的影響力
| 38
|
從-自私的帳本-看見-ai-不可見的影響力-17d52aaa227b
|
2018-06-27
|
2018-06-27 07:29:23
|
https://medium.com/s/story/從-自私的帳本-看見-ai-不可見的影響力-17d52aaa227b
| false
| 93
|
歡迎來到集知會所，我們是一個以交流創意知識為核心的社團，發起者多為廣告、行銷業界資深人士，每月我們會設定主題，與大家一同討論各種觀點。期望在變動快速的時代，資深與新血都能在此教學相長。社團連結：https://pse.is/knowledgehere
| null | null | null |
集知會所
| null |
集知會所
|
行銷,創意產業,設計,廣告,評論
| null |
人工智慧
|
人工智慧
|
人工智慧
| 85
|
集知
|
創意交流社團《集知會所》精華站，在此匯集了跨領域的觀點，更多當期討論臉書請搜尋《集知會所》或是直接點選 http://pcse.pw/knowledge
|
55961daf3eff
|
knowledgehere
| 234
| 79
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-02-18
|
2018-02-18 15:06:16
|
2018-03-03
|
2018-03-03 00:11:10
| 9
| false
|
en
|
2018-03-05
|
2018-03-05 22:10:34
| 10
|
17d58232ff76
| 8.943396
| 11
| 1
| 0
|
On the dangers of In Browser Cryptocurrency Mining, and how I got 1000s of computers to mine for me.
| 5
|
The Vulnerability Nobody Wants to Fix
On the dangers of In Browser Cryptocurrency Mining, and how I got 1000s of computers to mine for me.
In the 90’s, a proposition was put forward to limit Email spam. Emails would be validated by senders using a Proof of Work (POW) algorithm that required some amount of CPU time and computation.
The system, called HashCash, was rejected because of fears that spammers could simply use other computers’ processing power to do this work. Twenty years on, and we’re beginning to see why this decision was made.
Now, an idea has arisen based on a similar POW concept. When users browse to a website, their computers can do some work in order to forego advertisements.
In Browser Crypto Mining
This concept is called In Browser Crypto Mining (IBCM). Yup, that’s a thing…
Between the news headlines and incessant social media chat, it seems like everywhere you go some libertarian appears asking to discuss Bitcoin or the Blockchain, like a religious wing-nut in Times Square.
It’s usually fine to zone out when this happens, but you shouldn’t now. That’s because this story of magic internet money is scary and it involves you.
When you load a website, the code running in your browser can turn your computer into a cryptocurrency miner and send the rewards off to the owner of the website. Specifically, the currency mined is Monero (XMR, $345.12). The idea of IBCM goes: the User, in exchange for visiting the Website, pays a small fee to the Website Owner. This fee is in the form of Monero, which is mined on the User’s computer. It is essentially a swap of electricity for the ability to forego advertisements.
Mining = Electricity = Money.
That’s you, on the right.
The hope is that website owners will use this software transparently to charge you so they don’t have to display adverts. However, just like HashCash, this notion could be dangerous.
The reason lies in the fact that it is extremely hard for the Website Owner or You to detect when this software is running. Oh, and it is not just Website Owners who can choose to run IBCM. Anyone can….
Just don’t refrigerate your computer
What if I told you that if you’ve visited any of these websites in 2018, then your computer has literally mined Monero for someone else: youtube.com, facebook.com, all Word Press pages, or any of these 4200 websites.
This is not happening in obscure corners of the internet; even multiple Government Websites (including the UK’s NHS site) have all fallen victim to what I would definitely call ‘malicious software’. And that makes you a victim.
Right now, IBCM is mostly being done surreptitiously WITHOUT the knowledge of website owners or YOU.
Eesh
It turns out that naughty people have been injecting this software onto websites. This essentially puts your entire computer and its processing power into their hands when you visit that site. Your CPU becomes part of their Monero mining pool, performing up to on the order of 200,000 Millions of Instructions per Second (MIPS). Not only are they draining your electricity but also impeding other programs’ performance and your processor’s resources.
It’s crazy to think that this appropriation can happen so easily. Illustrating how and why is the main purpose of this article.
WhiteHat: 60% of Websites are Insecure
Client-side Injection Attacks
There are three main methods that hackers use to get their mining code into your browser.
Cross-site scripting (XSS), Improper Input Handling
XSS is the third most common webpage vulnerability. Unsanitized user inputs give hackers an entry to websites. The call to eval above would execute any code in the victimʼs browser context, including malicious mining code.
External Resources
The second method of attack is ironically through third party resources like adverts. This type of attack is not limited to just ads, but any source which has embedding permission in the page. If these external sources themselves are compromised, a hacker can slip mining code into client-side Javascript as is done in Trojan Horse attacks.
Browser extensions
A final method of attack is through browser extensions. Extensions can render any JS script through content scripts in any tab in your browser, serving as another entryway for malicious code.
These plug-ins are dangerous because they pervade the browser, allowing malicious code to be executed in any website no matter how secure.
This is how all Chrome extensions work. So it would behoove Google to protect Chrome users against malicious extensions, like ones that mine cryptocurrencies.
Chrome Extension Malware Detection
Google have an automated security screening process of questionable proficiency. Some have called it a “joke” and others, “one of the largest single security threats to the web right now.” So I did a bit of digging into Chrome’s malware detection model to find out why.
The feature set for classification consists of a collection of over 20 million signals. That’s for a training set of 99,818 extensions. 😬
The automated detection uses an implementation of online gradient descent logistic regression with Lasso L1 regularization. That is to reduce the feature set of size 20 million down to something more reasonable, though in a computationally inefficient way. A possible improvement here would be to follow L1 regularization by an L2 regularized model-building step, to minimize the influence of outliers.
“We have variable confidence in the malware or benign labels assigned to training instances”.
Millions of Malware Installs
This model also runs the risk of overlooking cloaked malware; malicious developers may delay execution until after dynamic analysis takes place, obfuscate code, or fingerprint IP addresses used in the model.
Furthermore, the detection scheme doesn’t achieve a wide code coverage since extensions are event-driven and have many branches. There is no sub-model trained to explicitly search for activation pathways.
Finally, the fact that this model is public means that malevolent developers can train their extension to be classified incorrectly.
Blocking the Bulk of Malicious Extensions
To compensate for the massive amount of features, they multiply the gradient descent learning rate with a correction factor that is proportional to an approximate confidence level, assigned by known malware items.
They look for specific threats; outgoing POSTs that may indicate unauthorized account manipulation, DOM operations that replace adverts, Torpig banking trojans and so on.
All in all, the prediction model finds most malicious extensions (80%+). It does a good job of eradicating clear signs of abuse, but seems not yet to find extensions that incorporate IBCM.
I hope this isn’t Illegal
So I decided to test my hypothesis. Not only did I want to create a malicious extension to see if Chrome took it down, but also to check how many people online were protecting themselves against IBCM.
I created a simple Chrome extension to change the browser’s colour temperature just like Flux. But also to mine cryptocurrencies. Then, I embedded mining code on a website I built last month. I also allowed my code to throttle the victim’s processing power (see code below), on any number of the victim’s cores.
Instantiating the Miner In Browser
Only 7% of accesses to my malicious code were blocked; the rest let me slide right on in.
In just a few weeks, I had turned over 1200 CPUs into my very own money making machines.
Protect Yourself
Here are a few ways we could block IBCM, and a few reasons why protecting ourselves is rather difficult.
Blacklisting and Validating Content
First off, you could have a content script run against the rendered HTML and blacklist known malicious domains and IP addresses. Extensions like No Coin hardcode known malevolent domains and block them.
These extensions could also grep for known 3rd party miners CDNs like coinhive.min.js and digimine.min.js.
I ended up compressing my client side Javascript into complete gibberish to avoid this. Attackers could also rotate their CDN’s IP using an elastic address.
Most browsers validate embedded content using Subresource Integrity (SRI). Comparing the ‘integrity’ cryptographic hash (such as SHA-3 checksum) provided by the website owner to the computer hash of the resource fetched would expose compromised assets. In the case that these external assets are hijacked, the browser would be able to block them.
Monitor your CPU
A second method would be to monitor CPU load, to see if your cores are in overdrive. The browser could have an energy threshold above which a potential mining warning would be raised. Or if we wanted to be really smart, we could implement a model which would only flag abnormal CPU usage.
However, since JS is based on a single event loop, this type of monitoring would be performance costly; it would be like killing a fly with a submachine gun. Not to mention that this would itself be a hack, since the browser shouldn’t have access to this kind of information.
Additionally, legitimate CPU intensive processes like video games and Netflix could be mistaken for mining. Miners could try and fly under the radar by throttling back the mining rate, or replicating the performance of video playback.
Check your scratchpad
The final method of detection is definitely the most clever. To understand it, we have to dive into Cryptonote, the protocol behind Cryptonight, Monero’s POW hashing algorithm. This is the coolest thing you’ll learn today.
Cryptonight allocates around 2MB of fast CPU cache (usually L3 cache), one the most expensive piece of hardware on your computer. The allocated memory block (called the scratchpad) is split up and individual components perform rounds of Advanced Encryption Standard (AES) encryption and read/write operations at pseudo-random addresses. Finally, the entire scratchpad is hashed with Keccak to yield the mining output number.
Cryptonight is designed to be inefficiently computable on GPUs, effectively to mitigate the chance of one entity gaining a large proportion of hash power.
This is what makes IBCM viable.
This behaviour arises from a few things; memory in ASICs (like GDDR5, used in Ethereum mining) may have a higher bit rate, but slower random access speed than L3 cache. GPUs also just don’t have enough cache memory to allocate 2MB of work to. In addition, most CPUs are really good at doing rounds of encryption using AES, which comprises most of the Cryptonight algorithm,
Also, heavy calls in a cryptocurrency networks may result in vulnerabilities. Miners must check every new block’s POW. If a node spends too much time on each hash evaluation, it could theoretically be DDoS’ed with arbitrary Merkle Trees requesting validation. This would destabilise the GPU at that node.
Now that we know how Cryptonight works, perhaps we can come up with a way to block it. Simply check internal fast memory (scratchpad) to see if your CPU is doing something which resembles the algorithm.
Trending
Taking into account these difficulties, will anything stop IBCM? Analysing Monero mining behaviour in the wake of the exposed Spectre vulnerabilities might give us a hint.
Kernel patches like Kaiser, which affects system calls and interrupts could slow down Cryptonight’s performance by up to 28%. This would impede on miners’ hashrates, making it more costly to mine.
Also, Spectre variant 2 can be used to evict L3 cache memory (where Cryptonight runs) even remotely in-browser. With system level vulnerabilities exposed at the CPU level, perhaps computed hashes could be stolen via side channel timing attacks, even though remote penetrations are harder.
However, we don’t see any sign of reduction in hash rate through January (when vulnerabilities were announced). This highlights Monero’s viability in spite of a potential mining vulnerability and performance hit.
It’s Time for Less Trust and More Self-Defense.
Even in the 90’s, before the days of data driven machine learning, people were aware of the risks of a sort of In Browser Mining.
Hopefully more mainstream workarounds will be constructed in the future to protect us from IBCM. Preventing mining isn’t going to be easy, but companies like Google pride themselves on finding solutions to hard problems.
At the end of the day, defacing a website is fun but mining is profitable. This type of attack will not stop and only you can defend yourself against it right now.
So next time your computer overheats and fans get noisy, you may be able to guess why.
Disclaimer: Please note that this is an opinion piece and for information purposes only. Thanks to Nikhil Bhatia and Jack Dent for their help.
|
The Vulnerability Nobody Wants to Fix
| 66
|
the-vulnerability-nobody-wants-to-fix-17d58232ff76
|
2018-06-19
|
2018-06-19 04:37:46
|
https://medium.com/s/story/the-vulnerability-nobody-wants-to-fix-17d58232ff76
| false
| 2,052
| null | null | null | null | null | null | null | null | null |
Blockchain
|
blockchain
|
Blockchain
| 265,164
|
Fredric Moezinia
|
MIT CS
|
7137edf3c463
|
fredricmoezinia
| 214
| 222
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
634d4b270054
|
2018-05-07
|
2018-05-07 09:38:12
|
2018-05-07
|
2018-05-07 09:42:33
| 1
| false
|
en
|
2018-06-05
|
2018-06-05 07:19:57
| 3
|
17d5ba53c96
| 1.132075
| 1
| 0
| 0
|
UAVs have become the great source in providing accurate data to the customers. But sooner or later, the demand for drone data management…
| 5
|
The Need For Drone Data Management Platform
UAVs have become the great source in providing accurate data to the customers. But sooner or later, the demand for drone data management platform may also arise.
“Our drone data management and AI Analytics Platform turns unstructured data into answers and delivers that data in useful ways,” says Optelos, a drone data management company.
A small conversation held with David Tran, Co-founder and CEO of Optelos, covered various aspects related to drone and AI algorithms.
AI is being used by a lot of drone companies to streamline the work and reduce the time to deliver insights. Drone data may be complex at times, but extracting all the rich metadata and correlating it would provide better picture of collected data.
Plus, compared to the current manual workflow, the delivery time via drones is reduced to 60%. AI solution will make a great place in the drone industry.
Source: https://bit.ly/2Jgi9Vu
About DEEPAERO
DEEP AERO is a global leader in drone technology innovation. At DEEP AERO, we are building an autonomous drone economy powered by AI & Blockchain.
DEEP AERO’s DRONE-UTM is an AI-driven, autonomous, self-governing, intelligent drone/unmanned aircraft system (UAS) traffic management (UTM) platform on the Blockchain.
DEEP AERO’s DRONE-MP is a decentralized marketplace. It will be one stop shop for all products and services for drones.
These platforms will be the foundation of the drone economy and will be powered by the DEEP AERO (DRONE) token.
|
The Need For Drone Data Management Platform
| 1
|
the-need-for-drone-data-management-platform-17d5ba53c96
|
2018-06-05
|
2018-06-05 07:19:59
|
https://medium.com/s/story/the-need-for-drone-data-management-platform-17d5ba53c96
| false
| 247
|
AI Driven Drone Economy on the Blockchain
| null |
DeepAeroDrones
| null |
DEEPAERODRONES
| null |
deepaerodrones
|
DEEPAERO,AI,BLOCKCHAIN,DRONE,ICO
|
DeepAeroDrones
|
Deepaero
|
deepaeros
|
Deepaero
| 0
|
DEEP AERO DRONES
| null |
dcef5da6c7fa
|
deepaerodrones
| 277
| 0
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-26
|
2018-05-26 06:34:25
|
2018-05-26
|
2018-05-26 06:35:56
| 1
| false
|
en
|
2018-05-26
|
2018-05-26 06:35:56
| 1
|
17d6383676d3
| 0.335849
| 0
| 0
| 0
|
In this example, you will learn: How to calculate difference between dates in Python.
| 4
|
How to calculate difference between dates in Python — Machine Learning Kickstarter example 043
In this example, you will learn: How to calculate difference between dates in Python.
https://setscholars.com/DataScience/how-to-calculate-difference-between-dates-in-python/
|
How to calculate difference between dates in Python — Machine Learning Kickstarter example 043
| 0
|
how-to-calculate-difference-between-dates-in-python-machine-learning-kickstarter-example-043-17d6383676d3
|
2018-05-26
|
2018-05-26 06:35:57
|
https://medium.com/s/story/how-to-calculate-difference-between-dates-in-python-machine-learning-kickstarter-example-043-17d6383676d3
| false
| 36
| null | null | null | null | null | null | null | null | null |
Python
|
python
|
Python
| 20,142
|
Nilimesh Halder
| null |
88be7c24b7fd
|
nilimeshhalder
| 51
| 3
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-02-20
|
2018-02-20 18:17:12
|
2018-02-20
|
2018-02-20 18:25:52
| 2
| false
|
en
|
2018-02-21
|
2018-02-21 11:33:33
| 2
|
17da858922ba
| 1.390881
| 2
| 0
| 0
|
Pete Williams, ex Head of Enterprise Analytics at Marks and Spencer [M&S] and now running his own outcome focused data consultancy, The…
| 4
|
Podcast Series: The role of the CDO with Pete Williams
Pete Williams, ex Head of Enterprise Analytics at Marks and Spencer [M&S] and now running his own outcome focused data consultancy, The Decision Playbook, shares with us in the interview his view around the transformational role of the CDO. His overall aim was to bring dynamic data to the heart of every business decision to make every decision faster, better informed and of higher quality.
Pete highlights the risk of getting in the hype cycle of needing a CDO; making bad decisions on the data science team; the difficulty of recruiting a successful team and the challenges in defining the data structure.
Listen to an unconventional point of view on what the role of the CDO is, his place in the C-suite and how he developed his advisory role.
Listen to the entire interview in this podcast
“The role of the Chief Data Officer inside a company” — a Podcast with Pete Williams
You’ll find more about:
The danger of the hype cycle and the need to have a successful Chief Data Officer & data science team
Why companies are not allowed to rely just on gut feeling, but on data channeled via the data science team.
HR analytics as a data opportunity in a company
Finding a data structure flexible enough to satisfy the organisational needs & the semantic layer in the data structure
How to prepare a safe environment for the team to play with the data & the secret of delivering on time
The use of cloud technology to smooth the transition from one tool to another and the always growing tendency of moving away from on-premise data
|
Podcast Series: The role of the CDO with Pete Williams
| 2
|
podcast-series-the-role-of-the-cdo-with-pete-williams-17da858922ba
|
2018-02-21
|
2018-02-21 14:12:46
|
https://medium.com/s/story/podcast-series-the-role-of-the-cdo-with-pete-williams-17da858922ba
| false
| 267
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
Keboola UK
|
We provide cloud-based data engineering platform, helping clients #DoMoreWithData by integrating, augmenting and enriching it for analytics & data science needs
|
c9a3ec5a48f6
|
keboolauk
| 10
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-02-06
|
2018-02-06 19:48:12
|
2018-02-06
|
2018-02-06 19:53:42
| 1
| false
|
en
|
2018-02-06
|
2018-02-06 20:01:12
| 1
|
17db11d469b3
| 3.592453
| 0
| 0
| 0
|
I distinctly remembers that day as if it happened yesterday. It was in autumn almost five years ago that my brother called to inform me…
| 4
|
We found a way to know who is going to die, when and how.
By Muhammad Aurangzeb Ahmad
I distinctly remembers that day as if it happened yesterday. It was in autumn almost five years ago that my brother called to inform me that the doctors has given their prognosis: our father was unlikely to survive the week. They said we should prepare for the eventuality. While my father had been in and out of hospitals a number of times over the prior year, there had been no indication that his health would deteriorate to this level so soon. Collectively everyone in my family expressed one wish that we had more time to prepare, that we had been forewarned, that we could have structured the last few months or even last few weeks of my father’s life differently. Death is an inevitable part of life. Yet how we react to and care for death, especially in old age, can be partially controlled by us. This forces the important question of why and how in an era of AI and machine learning can we help millions of families like mine to better cope with the potential loss of a loved one.
In the US a large percentage of elderly patients on Medicare pass away in acute care hospitals. For many, the last six months of life are characterized by complex medical procedures, repeated emergency visits, and frequent hospital stays. This corresponds to significant costs to the taxpayers with marginal benefits to the patients and their families. In most cases, one gets a very small window of time to prepare for and care for the patient about to expire. Thus, many patients and their families do not have a sense of control over the last stages of their lives. In surveys, a majority of Americans, around 70 percent, expressed their wish to ideally die at home; living their normal way of life, surrounded by their loved ones. Palliative care services are designed to provide alternatives to hospital-based medicine for patients in their last months of life. The number of patients who utilize these services has been growing steadily so that it now comprises of almost half of people (48%) who die annually. However even among those who utilize hospice services, the average length of palliative care is a mere 26 days. Just 26 days.
We hope that by assisting physicians to predict the onset of mortality risk for elderly patients can increase their ability to have conversations with patients and families sooner and improve the overall quality of care towards the end of life. This week we will be presenting some key insights from our research on predicting mortality risk six months to one year out at the Thirtieth Conference on Innovative Applications of Artificial Intelligence held at annual AAAI (Association for the Advancement of Artificial Intelligence) conference. Using extracts of data from Electronic Medical Records and Patient Claims we developed various machine learning models to predict such risk of mortality. In many healthcare settings there is reluctance to use machine learning models because a large number of such models are unfortunately what we term as ‘black-box’ models, i.e., it is not always clear to even a human expert in that domain, in our case physicians, why the model is making certain predictions. To address this problem, KenSci has invested significantly on developing models which not only have a high predictive power, but are also interpretable. This ensures that using the KenSci platform, accompanying each prediction the machine learning system also specifies why the prediction was made, and which factors might have been important to the system that made the prediction. Care providers can thus meaningfully consider the risk recommendation and objectively decide to accept or ignore such recommendations given their own expertise and understanding of the larger picture. This is one way to make the underlying machine learning system more assistive and trustworthy.
Interpretable models particularly within healthcare, is a growing area of interest for the AI community, and at KenSci Research our vision is to foster global research partnerships with health systems and academic research to integrate best-practices in healthcare with state of the art in AI leading to broad social impact. Our recent efforts on interpretable risk of mortality models is one such effort.
While we have no doubt that advances in AI will one day help patients experience a more fulfilling life, my thoughts often return to that day in autumn. I still feel the loss, but I also feel a great sense of hope that by contributing towards the advancement of what we at KenSci term Assistive Intelligence will one day enable physicians to make and also communicate their decisions much sooner giving families more time.
The Association for the Advancement of Artificial Intelligence (AAAI), announced that KenSci’s academic research paper has been recognized as EMERGING and will be presented at the 2018 AAAI conference’s track Innovative Applications of Artificial Intelligence 2018 (IAAI-18) in New Orleans, Louisiana. The paper dives into how machine learning techniques are used to predict the risk of mortality for patients from two large hospital systems in the Pacific Northwest along with the explanations for end of life predictions and insights that are derived from the predictions can then be used to improve the quality of patient care towards the end of life.
Download the paper here: http://bit.ly/Death-Vs-DataScience-Research-Paper
|
We found a way to know who is going to die, when and how.
| 0
|
we-found-a-way-to-know-who-is-going-to-die-when-and-how-17db11d469b3
|
2018-02-06
|
2018-02-06 20:01:14
|
https://medium.com/s/story/we-found-a-way-to-know-who-is-going-to-die-when-and-how-17db11d469b3
| false
| 899
| null | null | null | null | null | null | null | null | null |
Healthcare
|
healthcare
|
Healthcare
| 59,511
|
KenSci
|
We’re fighting Death with Data Science
|
9db9e86483f3
|
kensci
| 18
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-10-30
|
2017-10-30 23:43:50
|
2017-11-06
|
2017-11-06 21:37:17
| 1
| false
|
en
|
2017-11-06
|
2017-11-06 21:37:17
| 0
|
17dc4d0ce379
| 1.396226
| 10
| 0
| 0
|
There can be many benefits to adding a personality to your bot. Today we will cover three of our favorite aspects and discuss why they are…
| 4
|
Benefits of your bot having a personality
There can be many benefits to adding a personality to your bot. Today we will cover three of our favorite aspects and discuss why they are important. Before diving in though, think about how much a personality can define a particular person you know. There are some of us with short and hot tempers and others of us who are quite calm and collected, no matter the circumstances. We often think of this subconsciously but it can greatly impact how you interact with this person. The same can be true for a chat bot as we integrate A.I. technology to allow the bot to be more human like.
Express your brand motive
The more inline your chat bot can be with your brand motive, the better it will preform for it’s users. For example, a end user facing bot should be able to handle a variety of tasks which a back end human resource style system may simply not need to. The end user should be considered when designing the bot and therefore the personality.
Higher Conversion Potential
There is a saying in sales that a salesman sells themselves not the product they’re selling. This can often be very true if you’ve ever worked in a sales position. Furthermore, expressing why it’s a great idea to tailor your bot’s personality to be more engaging with it’s users. This will lead to a high conversion rate and thus a better use of your investment for the development of a chat bot.
Chat bots are getting smarter
The potential to create an intelligent chat bot which can shine through directly to your customers has never been easier. The artificial intelligent aspects can be truly astonishing as well. The simple ease of adding a spunky or professional or feminine or masculine tone can greatly increase your bots efficiency.
Until next time friends,
J
|
Benefits of your bot having a personality
| 16
|
benefits-of-your-bot-having-a-personality-17dc4d0ce379
|
2018-05-09
|
2018-05-09 15:23:54
|
https://medium.com/s/story/benefits-of-your-bot-having-a-personality-17dc4d0ce379
| false
| 317
| null | null | null | null | null | null | null | null | null |
Bots
|
bots
|
Bots
| 14,158
|
HOLDME
|
HOLDME is An Artificial Intelligence Revolution Live Chat software powered by Ethereum Blockchain.
|
fbf4091f5b9a
|
holdmeio
| 77
| 5
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-25
|
2018-05-25 10:11:24
|
2018-05-25
|
2018-05-25 10:12:59
| 0
| false
|
en
|
2018-05-26
|
2018-05-26 15:31:48
| 2
|
17dc5081d2c2
| 2.139623
| 1
| 0
| 0
|
If you’re anything like me, your previous experiences with chatbots have left you with a sour taste in your mouth. At best, they can help…
| 5
|
What do applications such as Mitsuku mean for the future of chatbots?
If you’re anything like me, your previous experiences with chatbots have left you with a sour taste in your mouth. At best, they can help you get in touch with a real person; at worst, they elevate your blood pressure to dangerous levels.
Mitsuku, however, is different. Mitsuku, unlike chatbots you’re likely to encounter on insurance sites and the like, is a conversational bot designed specifically for human communication. In other words, Mitsuku will respond to your messages with a level of intelligence that is truly dazzling for a chatbot.
Then again, this should come as no surprise when you see the accolades that Mitsuku has garnered. Created by Steve Worsick, the chatbot has won the Loebner Prize (which is awarded to the most human-like chatbot) on three occasions: 2013, 2016, and 2017. Development on Mitsuku first started way back in 2005, so it’s little wonder the acclaimed chatbot has managed to accumulate a large sum of knowledge since its inception.
To get a true appreciation of what Mitsuku is capable of, allow me to show you some examples from my conversation I had with it. I’ll start with a softball question:
Me: “What is your favourite number?”
Mitsuku: “My favourite number is 25.”
So far so good.
Me: “Why is your favourite number 25?”
I want to test Mitsuku’s ability to articulate reasons behind her replies.
Mitsuku: “Because it’s odd and square, like me.”
Zing.
Me: “Can you pass the Turing test?”
Mitsuku: “You be the judge of that.”
Touché.
But Mitsuku isn’t unflappable, of course. Here’s an example:
Me: “Who is the fastest man alive?”
Mitsuku: “It depends on the historical context, because it changes from time to time.”
Now, I can see what the AI is doing here — it’s giving a vague answer because of its inability to properly grasp the question. Honestly, I thought this question would have been a layup, but evidently this wasn’t the case. Historical context is irrelevant, since I specified the fastest man alive.
Nevertheless, I am still very impressed by Mitsuku’s ability to converse in a fairly natural manner. I still think we are a considerable way off a chatbot being able to successfully fool a human into thinking it, too, is a human. I should also note that getting answers back instantaneously is a bit disconcerting, although it's certainly not a criticism — just an observation.
Even major media outlets like the Wall Street Journal have taken note. Worswick told technology reporter Christopher Mims, “What keeps me going is when I get emails or comments in the chat-logs from people telling me how Mitsuku has helped them with a situation whether it was dating advice, being bullied at school, coping with illness or even advice about job interviews. I also get many elderly people who talk to her for companionship.”
Clearly, there is a growing demand for intelligent chatbots, with more than 250,000 people conversing with Mitsuku each day back in 2015, and I’m sure that number has increased since then. As chatbots continue to improve, it will be fascinating to see what can happen when neural conversational models are fully fleshed out. But that will have to wait for a future article!
In the meantime, you can check out WorkFusion to see how automation tools can help streamline your own business.
|
What do applications such as Mitsuku mean for the future of chatbots?
| 50
|
what-do-applications-such-as-mitsuku-mean-for-the-future-of-chatbots-17dc5081d2c2
|
2018-05-26
|
2018-05-26 15:31:50
|
https://medium.com/s/story/what-do-applications-such-as-mitsuku-mean-for-the-future-of-chatbots-17dc5081d2c2
| false
| 567
| null | null | null | null | null | null | null | null | null |
Chatbots
|
chatbots
|
Chatbots
| 15,820
|
Ben Schultz
| null |
69c7b2c80f40
|
benschultz_57614
| 17
| 17
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-02-02
|
2018-02-02 21:35:12
|
2018-02-06
|
2018-02-06 17:55:28
| 3
| false
|
en
|
2018-02-07
|
2018-02-07 04:57:33
| 3
|
17df03691ad5
| 4.172642
| 4
| 0
| 0
|
Like a kid in a candy store, I’m happy to announce that I completed Professor Andrew Ng’s Introduction to Machine Learning course on…
| 5
|
Journey into data science and machine learning
Like a kid in a candy store, I’m happy to announce that I completed Professor Andrew Ng’s Introduction to Machine Learning course on Coursera. In this 11-week course, I sought to bridge my statistical knowledge and background in predictive analytics into artificial intelligence and machine learning (AI/ML). After spending more than 11 weeks, I can say that this is a good starter course on AI/ML and wanted to share my thoughts on getting started with machine learning courses such as these:
Introductory statistics and a bit of calculus knowledge is helpful (although not required). As the course goes over gradients and sum squared errors, it helps to understand why gradients are derivatives and why we’re doing sum squared errors to figure out the cost function. Sure, it’s possible to get through the course, but it may not be as easy to do so without understanding the underlying mathematical concepts.
If I had more programming background, I think it would have been slightly easier to get started. I started with more limited coding experience but to go through the assignments, it challenged me not just on the statistics but also on the development implementation. Specifically, I had inertia initially when I started the assignments because I was not sure where to begin. Now I’m much more comfortable not just reading code but also on the functional programming side. As result, I appreciated the consideration that we started with using the open source version of MATLAB, called Octave.
Use the resources, community, and tutorials as they can be extremely helpful to complete the course. Don’t necessarily solve the assignments by strictly following the assignment tips and recommendations. Try vectorization and even your own approach.
Late in the course, I figured out how to go faster in the course. For example, I’d work on the assignment for Week 9 and watch the lecture videos for Week 10 at the same time. By staggering it, I’d was able to double my weekly velocity. Also, I’d watch the videos at 1.5–2x speed. Taking notes really helps as well.
Comparing my predictive analytics background with what I learned in this course, I can say definitely that there is quite a bit of conceptual overlap in the area of statistics. My experience had been in the context of linear/multivariate/logistic regression, decision trees, and K-means using Excel type of applications. For example, my context had been around customer purchase decision support such as looking at a set of customer demographic or historical inputs to predict their likelihood to purchase in the future.
Similar to concepts in machine learning, a dataset would be split into training and validation data sets. The training data set is a set of data that has independent input variables and the actual results of an outcome (e.g., 1 for purchase or 0 for not). Based on the training data set, an algorithm with the betas (or called Thetas) would be “trained” through different approaches such as various regressions. Selecting the best approach and then applying them to a “blind” validation dataset, data from the validation set is applied to see if the results were similarly effective.
Beyond where my experience in predictive analytics, this machine learning course taught me about many key concepts and tips:
Matrix algebra and vectorization are the most coolest concepts to me. It’s not that for-loops are obsolete, but there’s just efficiency savings using vectorized approaches to solving cost and gradient problems. Reaching way back to high school algebra memories, I have had little appreciation of matrices until now.
Holy moly! 3-way split on the dataset. Training, validation, AND test! Enough said.
Learning curves are a very useful method to distill whether there is high bias or variance. It’s a plot comparing the performances of the cost functions (validation and training) to show the differences in error at a specific training data size. If both cost functions converge on a fairly high error, then solve high bias. If there’s a significant gap between cost functions, then look at solving high variance.
Regularization (lambda) can be used to reduce overfitting.
F-Score ( 2 * Precision * Recall ) / ( Precision + Recall) allows one to calculate the effectiveness of an algorithm rather than just looking at Precision or Recall by itself.
As a visual learner, I can appreciate that three dimensions affords the ability to plot and visualize a dataset. However, in machine learning, there are just way more features (and dimensions) than a human mind can comprehend. Hence, principal component analysis (PCA) allows one to reduce the dimensions into a surface plane or line. More importantly, this is done really to improve the learning algorithm’s performance having reduced dimensionality.
When you’ve a largest dataset, it can take a long time and be very costly to figure out gradient descent if you use batch method, which sums up all the partial derivative of the training set cost function. Enter methods such as stochastic and mini-batch, which either takes one example before updating theta or a hybrid of batch and stochastic.
“It’s not who has the best algorithm that wins, it’s who has the most data.” — Andrew Ng
While this course was an introductory course, it sparked my interest to continue learning about map reduce, python, and deep learning. Map reduce is used to accelerate the progress of the learning algorithm by splitting up the work into multiple machines. Python appears to be the most practical and common place to work in machine learning as I’ve talked to colleagues so I’m extending my python experience by working through Jose Portilla’s course which covers the range from numpy and data analysis to data visualization in pandas and such.
|
Journey into data science and machine learning
| 53
|
journey-into-data-science-and-machine-learning-17df03691ad5
|
2018-04-20
|
2018-04-20 16:22:47
|
https://medium.com/s/story/journey-into-data-science-and-machine-learning-17df03691ad5
| false
| 960
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Daniel Wu
|
Product Management, Innovation, Digital Health, UX, Data Science and Analytics.
|
a2bc7fe3d2d4
|
danielwu3
| 55
| 88
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-01-17
|
2018-01-17 10:50:07
|
2018-01-17
|
2018-01-17 10:51:50
| 2
| false
|
en
|
2018-01-17
|
2018-01-17 10:51:50
| 2
|
17df55085981
| 2.319182
| 0
| 0
| 0
|
An S-curve can be used to illustrate the innovation in a category, and the speed of adoption of that innovation. In the graph below, you…
| 5
|
The way to understand CES 2018 is with one simple letter — S.
An S-curve can be used to illustrate the innovation in a category, and the speed of adoption of that innovation. In the graph below, you can see growth rate of innovation on the Y axis and time along X.
Most new categories don’t innovate at a high rate until there’s a significant external factor that changes the requirement of the technology — for example there was only gradual innovation of the aeroplane until the start of the First World War. Then, after a spurt of innovation, the rate of change cooled as the technology matured.
Looking at 2018’s CES you can see a lot of categories that are are at the top of an S-curve. PCs, mobiles, and TVs are all very similar and are competing on levels of specification that most consumers would hardly notice. 8K TV anyone?
One of the fascinating things about innovation is that a new S-curve can start when a new force exerts itself on a category — we’ve seen this happen all over CES this year. All kinds of consumer devices are being taken on a new S-curve of innovation with the addition of Voice Tech. The integration of Amazon and Google assistants are changing the way almost every connected device works.
Alexa and Google Assistant are vying to be the soul of the objects in your home. Both US giants are trying to imbue all kinds of home appliances with their voice driven AIs. This could be your TV, fridge — even your toilet.
The crucial thing for marketers isn’t which AI will win the battle to be the voice operating system of your home — it’s that consumers will start to expect the devices they buy to be operated by voice. Voice driven UX will be table stakes, not a point of difference. I’ve been vocal about stats like “71% of people think voice tech will be used for one or more daily tasks in 10 years’ time” but most brands think it’s weird and are waiting for brands like Domino’s and Pernod Ricard to lead the way. But with 50% of all search expected to be coming from voice in 2020, is this really something smart marketers can ignore?
Brands don’t need to work on their voice enabled bread bin strategy anytime soon, but they do need to work out how voice is going to work as part of their overall customer experience. While most consumers aren’t going to fork out for a thousands for a fridge that you can have a conversation with, they are definitely going to expect to talk to their household objects of the future. Brands will be able to speak to them at home on these devices too.
Consumers talking to brands through household devices is at the bottom of its own S-curve and CES 2018 has shown the marketing industry that it needs to start making the ubiquity of voice tech in the home more than just a talking point.
|
The way to understand CES 2018 is with one simple letter — S.
| 0
|
the-way-to-understand-ces-2018-is-with-one-simple-letter-s-17df55085981
|
2018-01-17
|
2018-01-17 10:51:51
|
https://medium.com/s/story/the-way-to-understand-ces-2018-is-with-one-simple-letter-s-17df55085981
| false
| 513
| null | null | null | null | null | null | null | null | null |
Innovation
|
innovation
|
Innovation
| 59,190
|
Tom Ollerton
|
Innovation Director at We Are Social.
|
4b4f1661bb7e
|
mrtomollerton_47138
| 5
| 9
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-11-21
|
2017-11-21 08:35:49
|
2017-11-24
|
2017-11-24 21:49:09
| 5
| false
|
tr
|
2017-11-24
|
2017-11-24 21:49:09
| 0
|
17e000c09690
| 2.414465
| 3
| 0
| 0
|
Bu yazımda Autoencoder tabanlı Derin öğrenmenin ne olduğu anlatmaya çalışacağım. Autoencoder yaklaşımı yapay sinir ağlarının bir türevi…
| 3
|
AutoEncoder temelli Derin Öğrenme
Bu yazımda Autoencoder tabanlı Derin öğrenmenin ne olduğu anlatmaya çalışacağım. Autoencoder yaklaşımı yapay sinir ağlarının bir türevi olarak ortaya çıkmıştır. İlk olarak 1980 yıllarda ortaya çıkmış ve 2006 yılında derin öğrenme alanındaki önemli konularında biri olmuştur.Özellikle ses ve görüntü veri setlerindeki gürültü giderme ve boyut düşürme konusunda başarılı sonuçlar almıştır.
Autoencoderlar giriş ve çıkış katmanları aynı veri setine sahip çok katmanlı algılayıcıdırlar.Genellikle üç katmandan oluşurlar. Gizli katmandaki nöron sayısı giriş ve çıkış katmanındaki nöron sayısından az olmalıdır.Yapı itibariyle veri setinden sınıf bilgisi içeren etiketlere ihtiyaç duymaz . Bu özelliği ile denetimsiz öğrenme yöntemidir. Ağ, giriş veri setini çıkışta verilen aynı veri setine uydurmak amacıyla eğitim sırasında geri yayılım algoritmasını kullanarak en uygun ağırlık değerlerini belirler. Bu nedenle yöntem öğreticisi
olmayan geri yayılım algoritması olarak da anılmaktadır.Kısacası autoencoder, gizli katmanından elde edilen veri aslında giriş katmanındaki verinin daha düşük boyutlu bir uzaya iz düşürülmüş halidir.Bir çok autocoder yığın şeklinde arka arkaya bağlayarak derin öğrenme mimarisi oluşturulur.
Bu mimarinin son katmanına denetimli veya denetimsiz öğrenme katmanın eklenmesiyle autoencoder derin öğrenme makinası oluşturulmuş olur.
Aslına bakarsanız autoencoder sınıflandırma işlemi yapmaz. Temel amacı n boyutlu bir özellik vektörünü daha küçük bir boyutlu vektöre en az kayıp ile düşürmektir.
Yukardaki eşitlikte, x_j girdi katmanındaki j’inci nöronun değerini, y_i gizli katmanındaki i’inci nörona aktarılan değeri, n girdi katmanındaki nöron sayısını, w_ji girdi katmanındaki j’inci nörondan gizli katmandaki i’inci nörona giden ağırlığı, f ise aktivasyon fonksiyonunu(softmax,sigmoud vs) temsil etmektedir.
İlk seviyeden sonra elde edilen değerler aşağıdaki eşitlikte gösterildiği gibi çıktı katmanına aktarılır. Eşitlikte x_j çıktı katmanındaki j’inci nöronu, y_i gizli katmandaki i’inci nöronu, w_i_j gizli katmanındaki i’inci nörondan çıkan katmanındaki j’inci nörona giden ağırlığı, m gizli katmandaki nöron sayısını , f ise herzamanki gibi aktivasyon fonksiyonu temsil etmektedir.
Autoencoder yaklaşımındaki amaç ilk iki aşamada elde edilen x_jˆ’ değerini , girdi katmanındaki x_j değerine benzer bir değer gelmesidir. Bu iki değerin birbirine yakın gelmesi için ise ağırlıkların geri yayılım algoritması ile hesaplanarak sürekli güncellenmesi ile olur.Geri yayılım algoritması ise aşağıdaki eşitlikte göründüğü gibi iki değer arasındaki farkın karesini minimize ederek olur.
Diğer tüm derin öğrenme mimarilerinde olduğu gibi autoencoder öğrenme mimarisinde çözülmesi gereken problemlerden biri ağın aşırı eğitilmesidir.Aşırı eğitim durumu, bir modelin eğitim seti için büyük başarılar elde etmesi olanak sağlarken, eğitim seti dışında yeni bir eğitim seti için düşük bir başarı oranı elde edilmesi demektir.Bunun sorunun çözülmesi için yapılması gereken biri eğitim verisinin bir kısmını doğrulama verisi olarak kullanıp eğitim verilerindeki hata oranı azalırken benzer şekilde geçerlilik verisindeki hata oranın kontrol edilip ağırlıkları günceleyerek eğitime tekrardan başlanabilir.
sürç i lisan ettiysek affola.
|
AutoEncoder temelli Derin Öğrenme
| 7
|
autoencoder-temelli-derin-öğrenme-17e000c09690
|
2018-04-24
|
2018-04-24 02:34:48
|
https://medium.com/s/story/autoencoder-temelli-derin-öğrenme-17e000c09690
| false
| 419
| null | null | null | null | null | null | null | null | null |
Deep Learning
|
deep-learning
|
Deep Learning
| 12,189
|
server calap
|
Developer
|
4a02aa5acbfe
|
servercalap
| 7
| 40
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-09-30
|
2018-09-30 20:27:33
|
2018-09-30
|
2018-09-30 20:29:22
| 0
| false
|
en
|
2018-09-30
|
2018-09-30 20:29:22
| 1
|
17e090edb5de
| 1.524528
| 0
| 0
| 0
|
In Haylse‘s article, she constructed the question about posthumanism in the framework of multiple interconnected filed. For me the most…
| 1
|
AI & Culture: Week 6 Response
In Haylse‘s article, she constructed the question about posthumanism in the framework of multiple interconnected filed. For me the most inspiring ideas are her discussion about the “boundaries” between human and posthuman and the “disembodiment” of information. She elucidates the connection of cybernetics theory with the idea of “disembodiment” by claiming that “the erasure of embodiment is a feature common to both the liberal humanist subject and the cybernetic posthuman.” Through the discussion about human identity our thinking method was posthumanized. Thus, no matter in the future will the idea of human — homo spines being biologically replaced by cyborg or not, we are turning into posthuman.
When introducing the definition of posthuman, Haylse Claims that the consciousness is just the accident of information. In here, the physical human body was considered simply as a vessel for holding the true identity of human — “information”, and the human itself is the embodiment of information. If “information” can last forever, then the changing of the physical vessel is just a process of ensuring the “immortality” of the human. Then the question will go to whether the “mortality” and “immortality” become the differences of human and posthuman. Later Haylse provides an example of she being a posthuman, which is a collective system with her agencies. In such a system, the physical body is no longer the most important component, rather it becomes “the fashion accessories”. The communication of information is vital for the system. Under this situation, the physical body becomes the embodiment of a collective system of information.
Haylse also mentions William Gibson’s concept of “Data made flesh”, McCulloch’s concept of “neurons worked as information-processing system”. the emotion, consciousness, memory can be complained by the bioelectric signal. In here, all the object with data processing ability can be regarded as the embodiment of information, which provides a foundation for the union of the human with machine. In many cyberpunk fiction and movie, we can see in the future world it is very common for people to upgrade their body with intelligent machines.
I have a question about what Haylse states as “information can circulate unchanged among different material substrates.” If we regard memory as a certain type of cognitive information then the “unchanged” character is not applicable to memory. Memory, as we discussed during class, is vague and can be manipulated through the process of input and output.
|
AI & Culture: Week 6 Response
| 0
|
ai-culture-week-6-response-17e090edb5de
|
2018-09-30
|
2018-09-30 20:29:22
|
https://medium.com/s/story/ai-culture-week-6-response-17e090edb5de
| false
| 404
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Yixiao Fu
| null |
ed67a267ce24
|
yixiaofu
| 1
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-11-22
|
2017-11-22 09:10:06
|
2017-11-22
|
2017-11-22 09:17:58
| 1
| false
|
en
|
2017-11-22
|
2017-11-22 09:17:58
| 1
|
17e10f0a96d3
| 0.320755
| 22
| 0
| 0
| null | 5
|
Robotic Process Automation & Artificial Intelligence
Robotic Process Automation & Artificial Intelligence
Human endeavor has always been to make life easier, faster and convenient with improved productivity, efficiency, and…obizcoin-robotic-process-automation.blogspot.in
|
Robotic Process Automation & Artificial Intelligence
| 925
|
robotic-process-automation-artificial-intelligence-17e10f0a96d3
|
2018-05-23
|
2018-05-23 09:20:05
|
https://medium.com/s/story/robotic-process-automation-artificial-intelligence-17e10f0a96d3
| false
| 32
| null | null | null | null | null | null | null | null | null |
Ethereum
|
ethereum
|
Ethereum
| 76,961
|
Nikhil Agarwal
|
Skilled in SOP, AI & Blockchain | Management Consultant | B2C (Retail) Domain Expert
|
11d91d7855ae
|
nikhilagarwal_25623
| 46
| 19
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
1de9a8b799ad
|
2018-01-31
|
2018-01-31 17:23:03
|
2018-01-31
|
2018-01-31 17:35:46
| 1
| false
|
en
|
2018-01-31
|
2018-01-31 19:00:39
| 5
|
17e3e63cdbd1
| 3.641509
| 14
| 0
| 0
|
Decision-making algorithms help determine who gets into college, is approved for a mortgage, and anticipate who is most likely to commit…
| 5
|
Combatting ‘Fairness Gerrymandering’ with Socially Conscious Algorithms
Decision-making algorithms help determine who gets into college, is approved for a mortgage, and anticipate who is most likely to commit another crime after being released from jail. These algorithms are made by programs that ingest massive databases and are instructed to find the factors that best predict the desired outcome.
Both the people who write and who use these algorithms understand that the decisions they produce are not always fair. Bias against race, gender, religion, sexual orientation — almost any subgroup status — can be present in the data, the way the computer draws relationships between data points, or both. This leads to bad predictions, in the form of both false positives and false negatives, that inordinately cluster in some subsets of the population in question.
Penn’s Warren Center for Network and Data Sciences is working on this problem.
Michael Kearns and Aaron Roth
Michael Kearns, founding director of the Warren Center and National Center Professor of Management & Technology in Penn Engineering’s Department of Computer and Information Science (CIS), and fellow Warren Center member Aaron Roth, Class of 1940 Bicentennial Term Associate Professor in CIS, are interested in imbuing these decision-making algorithms with social norms, including fairness. They’re interested in one particularly vexing issue: Algorithms that take fairness into account can have the paradoxical effect of making their outcomes particularly unfair to one subgroup.
This is known as “fairness gerrymandering.”
Like a political party that draws districts such that a critical majority of their opponents’ voters are concentrated in one spot, an algorithm can meet fairness constraints by unintentionally “hiding” bias at the intersection of the multiple groups it’s asked to be fair to.
A malicious person could achieve similar results on purpose. A racist country club owner, for example, could comply with fairness-in-advertising regulations by only showing ads to minorities who live far away and who could not afford the dues. Taken at face value, the owner has complied with the letter of the law, but still produced results that are unfair.
Algorithms are susceptible to producing the same sort of biased outcome due to the intrinsic trade-off that predictive algorithms have to make between fairness and accuracy.
Imagine an algorithmic classifier that connected students’ SAT scores and high school GPAs to their college graduation rates. Colleges might think such a classifier would be the most fair and objective way to predict which new applicants will do best and admit them accordingly. However, a “race-blind” algorithm — one tasked to maximize accuracy across the entire population of students — might inadvertently return results that are unfair to minorities.
Black students, through the institutionalized racism of poorer schools and less access to private tutoring, might tend to have lower GPA and SAT scores than their white counterparts, though the relationship between those scores and their college performance is just as strong, if not stronger. In that scenario, the algorithm would incorrectly reject black students at a higher rate simply because they are a minority: they have fewer data points and thus a smaller effect on the classifier.
One approach existing algorithms use to avoid this type of unfairness is to stipulate that that the false-negative rates for each subgroup be equal. But this problem gets trickier and trickier as the number of subgroups an algorithm is tasked with considering is increased. This is where fairness gerrymandering comes into play.
“We might ask an algorithm to make this assurance of fairness to populations based on, say, race, gender and income, all at the same time,” Roth says. “And the algorithms will make false-negative rates equal on race, the false-negative rates equal on gender, and the false-negative rates equal on income. The problem is, when we look at the false-negative rate for poor, black women, it’s extremely unfair. The algorithm has essentially cheated.”
Critically, algorithms that do this type of fairness gerrymandering aren’t designed to do this on purpose; they just haven’t been explicitly told not to. Since they can meet the harder constraints by spreading out unfairness over the three groups they have been explicitly asked to protect, they do not need to consider that they have concentrated all that unfairness in the place where those three groups intersect.
Kearns, Roth and other Warren Center researchers, including doctoral student Seth Neel, and former doctoral student Steven Wu, are currently writing algorithms that do explicitly counteract fairness gerrymandering. Preliminary work on the subject is already on arXiv.org for other machine learning researchers to begin validating.
“We demonstrate theorems that prove that this algorithm provides protections against fairness gerrymandering,” Kearns says. “We also demonstrate experiments that show our definition of fairness in action. With every constraint you add you lose some accuracy, but the results remain useful. There’s a cost and a tradeoff, but there are sweet spots on real data sets.”
The key to the Warren Center’s efforts in this field is to synthesize knowledge from the relevant domains in law, philosophy, economics, sociology, and more, and to match it with the way computers “think” and behave.
“Traditional approaches to fairness, like regulations and watchdog groups, are important, but we’re trying to prevent discrimination directly in the algorithms,” Kearns says. “We’re not saying this is going to solve everything, but it’s an important step to get human social norms embedded in the code and not just monitor behavior and object later. We want the algorithms to be better behaved.”
|
Combatting ‘Fairness Gerrymandering’ with Socially Conscious Algorithms
| 27
|
combatting-fairness-gerrymandering-with-socially-conscious-algorithms-17e3e63cdbd1
|
2018-05-17
|
2018-05-17 19:40:22
|
https://medium.com/s/story/combatting-fairness-gerrymandering-with-socially-conscious-algorithms-17e3e63cdbd1
| false
| 912
|
University of Pennsylvania’s School of Engineering and Applied Science
| null | null | null |
Penn Engineering
|
elerner@upenn.edu
|
penn-engineering
| null |
PennEngineers
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Penn Engineering
| null |
af9f8605d39a
|
PennEngineering
| 2,000
| 3
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
f5af2b715248
|
2018-01-29
|
2018-01-29 01:18:27
|
2018-01-29
|
2018-01-29 02:43:05
| 7
| false
|
en
|
2018-02-11
|
2018-02-11 16:24:46
| 6
|
17e556138809
| 4.536792
| 24
| 0
| 1
|
McKinsey nails it here, and for once, I have to go all in and agree fully. This is why.
| 5
|
Why digital strategies fail — and what can be done about it.
McKinsey nails it here, and for once, I have to go all in and agree fully. This is why.
In McKinsey’s most recent Quarterly report, they address the issue of why digital strategies fail in general, and highlight 5 pitfalls that companies have to watch out for. I won’t go into these in detail, much better to go here and read the piece yourself. No. What I’d like to do is to discuss some of the excellent charts that are included, and most importantly, how I recognize some if not all of the issues raised in my daily interactions with businesses and business leaders, most of them in the financial services sector.
In chart 1 (re. supra), we see that only 8% of companies believe their business model will remain economically viable through digitization. You say what?? That’s great, so you would think they’re all on the case. Are they? They’re not, as we all know. We’re at this tipping point, where a majority of incumbents do not respond and ultimately fail. Only a few of them will partially transform and/or find niche markets.
Example? In a recent survey around data analytics and decisioning trends in the credit union industry, to the question “what is your credit union’s time line to fully implement your data strategy?”, 50%+ indicated a timeline of 3–5 years. Time to take action is today, not in 3 to 5 years. By that time, we’ll be having a very different conversation indeed.
In this second chart, they talk about how digital competition shrinks value. Customers win, companies lose. And digital is driving winner takes all economics. Some examples? In 2016, comparing smartphone global market shares vs. profits, Apple’s iPhone had a 14.5% market share, but a 79
% share of the profits! (HT to Scott Galloway in The Four) “Digital rewards first movers and some superfast followers.” Or put differently:
Digital also renders distribution intermediaries obsolete (how healthy is your nearest big-box store?), with limitless choice and price transparency. And digital offerings can be reproduced almost freely, instantly, and perfectly, shifting value to hyperscale players while driving marginal costs to zero and compressing prices.
Now this one I like most. And so should you, or fear most, depending on what side of the fence you find yourself. Bottom line, for most of you out there, feeling comfortable and snug, as you are “still growing nicely…”, if you fail to respond to the current digital challenge, it’s all downhill from here. Better hang on at this point! Up to minus 12% revenue growth on a 3-year outlook is nothing to sneeze at. Better be retired by then. And even if you are bold, and/or a first mover, the best you can expect is a +0.3% change in rev growth over that same period. A picture says more than a thousand words. En effet, Mr. Watson.
Understanding the new economic rules will move you ahead, but only so far. Digital means that strategies developed solely in the context of a company’s industry are likely to face severe challenges. Traditional approaches such as tracking rivals’ moves closely and using that knowledge to fine-tune overall direction or optimize value chains are increasingly perilous.
We’re in a winner take all economy, what lots of the goodies will go to the happy few, and the payoff will go to those who move boldly. 12% ROI if you move with gusto, 6% ROI in you invest in digital, just to stay the course. Also, by 2025, more than 30% of global sales will come from ecosystems (I personally believe it will be more) — whereby the FAANG players will increasingly be playing catch up with the Chinese roller coaster, which is playing a long game — and they are going to make our lives much more interesting, just not always in a way we’re going to like.
This last slide i.e. quote is a no-brainer, and it’s interesting to see some forward thinking executives understanding that it is beyond the time start thinking, but that they have to start moving.
Digitization is changing everything around us (Amazon Go anyone?), and so is the workforce. As my friends at Steady like to put it:
More than 90% of all jobs created in the past 10 years have been within the part-time or independent labor sector. Today either by choice or necessity, 70 million Americans participate in work outside the walls of traditional W2 employment, by 2025 that number is expected to encompass half of the U.S. workforce.
Most of these are and will be of the millennial generation, a lot of them digital natives. If you’re not getting on top of your (digital) game now, so as to capture the value of that relationship in the near future, then I truly will feel no sorry for your outcome. I have seen the future, and it’s going to be bright, though the ride up there is going to be rockier than ever. Make sure you don’t miss it — it is going to be fun to watch.
Yvan De Munck is the US Head of Business Development at James (by CrowdProcess), the leading data science company that has developed “James”, the first credit risk AI. For more info, go to James.finance, or reach out to me directly for more information.
This story is published in The Startup, Medium’s largest entrepreneurship publication followed by 295,232+ people.
Subscribe to receive our top stories here.
|
Why digital strategies fail — and what can be done about it.
| 83
|
why-digital-strategies-fail-and-what-can-be-done-about-it-17e556138809
|
2018-05-15
|
2018-05-15 11:42:11
|
https://medium.com/s/story/why-digital-strategies-fail-and-what-can-be-done-about-it-17e556138809
| false
| 924
|
Medium's largest publication for makers. Subscribe to receive our top stories here → https://goo.gl/zHcLJi
| null | null | null |
The Startup
| null |
swlh
|
STARTUP,TECH,ENTREPRENEURSHIP,DESIGN,LIFE
|
thestartup_
|
Fintech
|
fintech
|
Fintech
| 38,568
|
Yvan De Munck
|
US Head of Business Development @ James, EU and US rooted, focused on AI & machine learning, Fintech and financial services, crypto's, ICO's and gaming.
|
c162dad50474
|
THEyvandemunck
| 46
| 50
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
82f1997178ae
|
2018-06-11
|
2018-06-11 15:36:54
|
2018-06-11
|
2018-06-11 15:51:35
| 3
| false
|
en
|
2018-06-27
|
2018-06-27 13:23:11
| 13
|
17e5d8b714d1
| 5.900943
| 2
| 1
| 0
|
By Sunand Menon
| 5
|
Capturing the Value of AI Content
By Sunand Menon
This post originally appeared in the DI’s monthly publication on Artificial Intelligence at Scale.
In an episode of “Black Mirror”, the hugely popular futuristic television series, soldiers searching for enemies wear a small device with a flashing blue light. When a witness gives information, this Instant Translator detects the language being spoken (Danish) and immediately translates it into English for the soldier. A fluent conversation takes place between two people, seamlessly facilitated by Machine Translation (MT). Artificial Intelligence (AI) at its finest.
We’re not there yet in 2018… but then again, we’re not far off.
Every day, new use cases for artificial intelligence (AI) emerge, showcasing its vast, yet untapped market value potential. McKinsey Global Institute recently published a report with insights from hundreds of use cases in which they estimated that AI could create between US$3.5 trillion to US$5.8 trillion in annual value in the global economy, which is approximately 40% of overall impact of all analytics techniques.
The commercial awareness and subsequent adoption of AI seems to be following a similar path to institutionalization of other (now ubiquitous) developments in Big Data, Predictive Analytics, and Connected Devices. Many of the use cases have focused on top-line-oriented functions such as Marketing and Sales, and in bottom-line-oriented operations functions such as Supply Chain Management and Manufacturing — where it’s relatively easy to visualize and quantify the impact of AI, from personalization to automation.
Great strides have also been made in AI in content-related areas such as Translation Services and Transcription Services — which also can greatly benefit organizations’ revenues and profitability.
In Translation Services, companies like Lionbridge, earn hundreds of millions of dollars by ensuring their customers’ online content is accurately represented in all relevant languages. They traditionally did this via armies of local employees or contractors who could help maintain brand identity and communicating clients’ value propositions with the right local nuances — whether it is in Mandarin or in Malayalam.
This business works because of its cost arbitrage model, and Lionbridge’s ability to source a knowledgeable, localized, specialist workforce which ensures quality. For many, the current electronic alternative is not particularly appealing, whether it’s Facebook’s translate button, or Google Translate, or the Safari translation add-on. Scalability does not reach the quality threshold, yet.
However, things are changing rapidly given the advances in MT by technology companies like Microsoft, which recently claimed to reach a historic milestone — namely using AI to match human performance in translating news from Chinese to English. So it’s not surprising that Lionbridge teamed up with Amazon Web Services to incorporate Amazon Translate, a neural MT service designed by Amazon Web Services.
Other competitors such as SDL have also taken note of this type of technological progress, and have come up with their own combined human/machine translation offerings. Many of these algorithms need not be developed from scratch — there are at least 10 open source toolkits for neural network machine creation that exist online. SDL has focused heavily on Linguistic AI, where they are building algorithms that make sense of masses of unstructured data by structuring the data via the creation of taxonomies and metadata, which in turn help understand, reason and learn in order to help their customers’ top-line performance.
Transcription Services have already shown bottom-line benefits to organizations by making workflow more efficient. For example, Scribble is a virtual scribe service adopted by the Massachusetts General Physicians Organization (MGPO). With such a service, doctors no longer need to write, type or dictate any notes; their conversation with the patient is recorded, transcribed (by medically trained professionals in India) and then returned to the physician for final check with full anonymity. Scribble is not the only service, others exist as well, e.g. Nuance, (which states that 90% of its transcription work is done via its Dragon software) and Augmedix.
On the positive side, by using Scribble a physician is able to dedicate more attention to the patient — as opposed to talking with her back to the patient while typing into a computer. However, the usually personal, connection-based dialogues may be replaced by more clinical and somewhat awkward conversations. Occasionally, the doctor would have to say, “For the purposes of the scribe, I note that…” in order to emphasize a certain point or give context, much like an author describing a scene or an action in a book.
It’s a win for the doctor in terms of efficiency, but is it truly a win for the patient? The objective was to be more customer-centric, which it was. However, it takes away some of the intimacy of the doctor-patient relationship, since the patient is aware of a third party — the “fly on the wall” who was an actual person, who is listening to private medical history. Even when technology becomes advanced enough to completely substitute this third party with a machine, this potentially opens up another set of issues around data security; what if that information is accessed and altered by unauthorized parties during the workflow?
Ultimately seizing the value of content generated and processed by AI in areas such as transcription as well as translation will depend on 4 important factors. As a company interested in utilizing AI in content operations, you will need to test yourself (or more likely, if you are outsourcing this capability) prospective vendors on:
Customer Centricity The service needs to nail the “Job to be Done”. In evaluating vendors, companies must test whether the vendor has designed a solution with a crystal clear understanding of who their customer is, what they would find acceptable, and what they would not. During implementation, the solution needs to be rapidly tested and iterated, not just in the initial phases, but continually. And very importantly, personalization options need to be incorporated. For example, in the medical field, if the aim is to generate better patient outcomes via better, more intimate understanding of the patients’ ailments while simultaneously reducing physicians’ administrative workload, more thought must be given to what will make the patient most comfortable and free to discuss.
Product Quality Vendors’ AI-generated services need to be accurate, reliable, repeatable and relevant to your needs. Your carefully developed brands are at stake, consequently you should be highly discriminating in selecting one provider versus another. Should you use an off-the-shelf application? Or a custom-built application? Should it incorporate additional services? Can the vendor understand how to classify and interpret slang, bad language, or even emotions that are discernible in the written or voice content? Determine your needs and your quality specifications very carefully and regularly test against the vendor’s deliverables. AI is at its infancy, but there are already several tiers of providers with varying qualities. Below is an example of a base, consumer-level application, the “auto translate” app within the Safari browser, which instantly translated a Fortune Japan article into English. It may be sufficient for some, but unsatisfactory for others.
3. Responsiveness Speed of Delivery and Service Level need to be appropriate for your business operations. The right solution requires a keen understanding of the type of responsiveness required. Responsiveness comes in two forms:
Speed of Delivery — much like in financial information systems companies such as Thomson Reuters and Bloomberg, where data streaming is characterized as Real-Time, End of Day, or another unit of time
Level of Service — represented by the support your vendor will be able to provide you on an ongoing basis, e.g. customizing the product, or delivering additional services as requested.
4. Security Your content is private, valuable and a source of competitive advantage –and should be managed securely. Given that every step of the translation or transcription process is an area of exposure for unauthorized parties to gain access, companies should work with vendors that will provide transparency and a full audit trail of all their data, a “chain of custody”. They should have a documented risk and mitigation plan, and aim to comply with the overall principles of the new General Data Protection Regulation (GDPR) which are being set up in Europe, but ideally would apply globally.
These four factors, if prioritized sufficiently, can accelerate the speedier adoption of AI-based solutions in content translation and transcription and provide significant top and bottom line benefits. The Instant Translator from “Black Mirror” may seem like science fiction, but it may be possible sooner than you think.
|
Capturing the Value of AI Content
| 28
|
capturing-the-value-of-ai-content-17e5d8b714d1
|
2018-06-27
|
2018-06-27 13:23:11
|
https://medium.com/s/story/capturing-the-value-of-ai-content-17e5d8b714d1
| false
| 1,418
|
Digital intersects with, and changes, everything.
| null |
digHBS
| null |
Digital Initiative Digest
|
digitalinitiative@hbs.edu
|
harvard-business-school-digital-initiative
|
DIGITAL TRANSFORMATION,TECH,DESIGN,EDUCATION,BUSINESS
|
digHBS
|
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
HBS Digital Initiative
|
Team @HarvardHBS that studies the digital transformation of the economy, and seeks to shape it through leadership, community, and research. Fueled by ramen.
|
4a9166f9a6da
|
digHBS
| 1,635
| 592
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-08-10
|
2018-08-10 10:07:03
|
2018-08-10
|
2018-08-10 15:14:28
| 2
| false
|
en
|
2018-08-23
|
2018-08-23 07:51:15
| 5
|
17e6efbe3bf4
| 2.817296
| 3
| 0
| 1
|
We need to evaluate our machine learning algorithms with the help of various metrics. There are some commonly used metrics for regression…
| 5
|
Classification and Regression Evaluation Metrics — Part 1
We need to evaluate our machine learning algorithms with the help of various metrics. There are some commonly used metrics for regression and classification problems. We will see cover some of these evaluation error metrics.
In this Part1, we will see some of the classification evaluation metrics. (I will post Part 2 article later where I will explain about regression metrics).
The best way to analyse any key concept or problem in machine learning is to code & implement and analyse the results. I have written the below classification example in Python. We will analyse the results and along with it go through the key concepts.
To summarize this code:
Generate a random matrix with 100 rows and 20 columns with values of either 0 or 1. First 19 columns will be the explanatory variables and the 20th column will be the response variable
Split the matrix into training and testing data sets. First 80 rows the training and the last 20 rows for testing
Perform the classification with the support vector machines
Use the confusion matrix and accuracy score to evaluate the results. (I have used the sklearn metrics library from the scikit-learn.)
Since the matrix we use is a random generated one, for every program run, the matrix values and the results will change. The data we analyse here will be for a specific run.
I ran the code and got the below results:
Actual Class values: [1 0 0 0 1 1 1 1 1 0 0 0 1 1 0 1 1 0 1 0]
Predicted Class values: [1 0 0 0 0 1 0 0 0 1 0 1 1 0 0 1 1 0 1 1]
Confusion Matrix:
[[6 3]
[5 6]]
4. True Negative, False Positive, False Negative, True Positive-6, 3, 5, 6 respectively
5. Accuracy Score: 0.6
We will now try to understand some key concepts and interpret the above results.
a) True Negatives are the rejections correctly classified as negative.
In our example, the second, third, fourth, tenth, eleventh, twelfth, fifteenth, eighteenth and twentieth elements are actually zero. Out of these the second, third, fourth, eleventh, fifteenth, eighteenth are correctly predicted as zeros. There are 6 true negatives.
b) False Positives are the incorrectly classified positives
In our example, the tenth, twelfth and twentieth elements are actually zero but predicted as ones. There are 3 false positives. False Positive is Type I error.
c) False Negatives are the incorrectly classified negatives
In our example, the fifth, seventh, eighth, ninth, fourteenth elements are actually one but predicted as zeros. There are 5 false negatives. False Negative is Type II error.
d) True Positives are correctly classified positives
In our example, the first, sixth, thirteenth, sixteenth, seventeenth and nineteenth elements are actually one and predicted as one. There are 6 True Positives.
The Confusion Matrix is a matrix where each row represents the actual class instances while each column represents the predicted class instances (or vice versa)
The Results we got earlier was:
[[6 3]
[5 6]]
We will put these values in the Confusion Matrix as shown below
Accuracy is calculated as (TP + TN)/(TP + TN + FP + FN)
In our example, this will be (6+6)/(6+6+5+3) = 0.6
This is exactly the result we got earlier.
You will also hear the terms sensitivity and specificity very frequently.
Sensitivity or True Positive Rate : TP/(TP + FN). In our example 6/(6+5) = 0.55. This is the proportion of the actual positives that are correctly identified as such.
Specificity or True Negative Rate: TN/(TN + FP). In our example 6/(6+3) = 0.67. This is the proportion of the actual negatives that are correctly identified as such.
Hope this article was helpful to you. Please post your comments and connect with me on Twitter and LinkedIn. Thank you.
|
Classification and Regression Evaluation Metrics — Part 1
| 12
|
classification-and-regression-evaluation-metrics-part-1-17e6efbe3bf4
|
2018-08-23
|
2018-08-23 07:51:15
|
https://medium.com/s/story/classification-and-regression-evaluation-metrics-part-1-17e6efbe3bf4
| false
| 645
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Balamurali M
|
Manager, skilled in multiple technology platforms & management methodologies. Homepage: http://balamurali-m.com/
|
4daf057aceed
|
balamurali_m
| 66
| 61
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
df28f4125da2
|
2018-03-19
|
2018-03-19 04:05:04
|
2018-04-13
|
2018-04-13 17:58:44
| 7
| false
|
en
|
2018-04-17
|
2018-04-17 20:28:40
| 32
|
17e91521179f
| 7.423585
| 16
| 0
| 0
|
The future is already here — it’s just not evenly distributed
| 5
|
Designing the Future we Want — Tech Trends from #SXSW2018
The future is already here — it’s just not evenly distributed
— William Gibson
At the SXSW 2018 Interactive conference, it certainly felt like the future is already here. Exploring the hallways, speaking to other attendees, and digging into the vast lineup of session topics spurred a reflection of past key trends, societal impacts of current tech trends, and what we can do now to design the future we want.
Looking Back
For the past 3 years, John Maeda, Head of Design at Automattic, Advisor at Kleiner Perkins, and former head of Rhode Island School of Design, publishes an annual Design in Tech report that summarizes his observations about key trends in design, technology, and business.
John Maeda presented at SXSW to a packed audience. Here are my key takeaways from his Design in Tech Report 2018:
Design is going global
Design in technology is evolving rapidly and spreading beyond Silicon Valley. China is leading in designing experiences, while India and Latin America are also advancing in design thinking and computational design.
Design tools are evolving
Design tools and systems are changing with machine intelligence. AI has gone to Art school and its grades are improving. Now, it is increasingly difficult to discern a real photo or video from a fake one.
Design beyond screens
Custom fabrication is changing manufacturing and natural language recognition is changing the norms of experience. Speed is a key design attribute.
Design for inclusion
AI is not good at inclusive design, because we are not either. However, John makes the case that inclusion = INCLU$ION, since inclusive design expands the total addressable market of any product or service.
Looking forward
Amy Webb, quantitative futurist, Professor of Strategic Foresight at the NYU Stern School of Business, and founder of Future Today Institute, examined 225 emerging tech trends across 20 industry sectors and presented her technology trend predictions for 2018 and beyond.
She presented her findings in 3 trend clusters:
2018 is the beginning of the end of smartphones
AI is already here, it did not show up as expected
Biology is the most important technology platform of the 21st century
#1: 2018 is the beginning of the end of smartphones
Amy predicts that emerging technologies such as natural user interface, augmented + mixed reality, generative algorithms, digital assistants, face prints, and voiceprints will mature and mark the beginning of the end of smart phones, as we know them. As digital assistants become ubiquitous, conversational or non-visual interfaces will become more prevalent.
Amy also predicts that 50% of people living in industrialized nations who interact with computers will use voice by 2021. This prediction is based on technology advances made in 3 areas — machine reading comprehension, voiceprints, and faceprints.
Machine Reading Comprehension
Machine reading comprehension (MRC) is the ability of computers to read and comprehend unstructured information, like reading and responding accurately to customer queries or summarizing news articles. This ability is improving, and in some studies, computers have surpassed humans. MRC lays the groundwork for our future conversations with machines.
Voiceprints
Voiceprints is an emerging trend. Your unique voiceprint divulges your health, age, emotional state, and the size of the room you’re in, the materials the walls are made of, how many people are in the room, fluctuations in the local electrical grid, and where you are located, all based on your voice. As such, conversations with digital assistants will replace passwords.
Faceprints
Our faces are unique. The capillaries underneath our skin are unique as well. Therefore, a person’s faceprint can be a reliable digital recorded representation of a person’s identity similar to the fingerprint.
There are several examples of work combining generative AI technology with faceprints. A university in Germany is working on technology that uses infrared to identify faceprints, and can therefore recognize people in the dark! NEC is working on technology which makes 2D faces to 3D, which can transform a photo into a 3D rendering proving that AI can extrapolate the shape of a face from a single image. Further, MIT is working on creating a video from a still image. Using AI they can “guess” missing information from an image and create a video (that never happened) from a still photo.
For example, if there is a photo of a person approaching another and extending a hand towards them, AI can “guess” that the people will shake hands. Then, it can create a realistic video of the two individuals shaking hands, whether it actually happened or not. Other universities are experimenting with this, as well. University of Washington scraped president Obama’s voice from The View and created a video of him talking. The video was never actually recorded, yet it looks convincing.
However, beyond the current research being conducted, there are also several examples of facial recognition already being used, today. Face++ is an important company that you have not heard of. In China, this technology is being used in a powerful manner. Chinese society has a more permissive attitude towards privacy, as there is government support for building infrastructure and applications based on face recognition. Here are two examples:
Smile to Pay via Alipay allows customers of a fast food restaurant, to order food at a kiosk and they can simply smile to pay using AI and facial recognition.
This technology can be used to publicly shame people for doing undesired behavior. For example, jaywalkers will be recognized and their information including their name, where you work will be displayed on a large bill board.
What impact will this trend have on society? Amy Webb presents an optimistic, pragmatic and pessimistic framing.
#2: AI is already here. It just didn’t show up as we all expected
AI has become such a part of many peoples’ daily lives that we don’t even recognize it, especially if it is Artificial Narrow Intelligence (ANI), a system that can perform a single focused task as well as, or better than, a human.
However, there’s a tremendous amount of misplaced optimism and fear.
The Robots are going to kill us all but not before they take all our jobs.
This misconception gets in the way of us truly understanding and leveraging this technology for greater good.
AI is the next era of computing. In a stunning example of how far Deep Learning has progressed, an AI system was able to teach itself the game Go, an ancient Chinese strategy board game. AlphaGo Zero learned the game and beat humans in 40 days.
The big eight
There is a consolidation in the AI ecosystem and the Big Eight companies that control the future of Artificial Intelligence are Tencent, Baidu, Alibaba, Amazon, Google, Microsoft, Apple, Facebook and IBM. Note that the first three companies are based in China.
Data is the next oil
Data is key to training these AI systems, and by 2038, human data will be a valuable natural resource. Companies who have them will have a significant advantage.
Splinternets
As governments realize the power of this data, they will create “walled gardens” to control and protect it. This will lead to the emergence of “splinternets” — smaller versions of the internet created for specific regions. There are quite a few pending cases and regulations that indicate that governments are starting to pay attention to who owns data. For example, in Germany, social networks must delete hate speech within 24 hours of posting or be fined $57 million per instance. Europe is preparing to enforce Global Data Protection Regulation (GDPR), which is taking effect in May 2018. In Canada, Google must scrub pirated products from search results.
Future regional splinternets will exacerbate the problem of fake news. We can look to China today to see how this could play out in other parts of the world.
What impact will this trend have on society? Amy Webb presents an optimistic, pragmatic and pessimistic framing.
#3: Biology is the most important technology platform of the 21st century
There have been several recent advances in personalized medicine, such as CRISPR+ genomic editing, digitization of the human body, and nano-technology that can remarkably improve people’s longevity and health.
There are a number of interesting applications of these technical advances. For example, Talking Tattoos is a color changing skin tattoo that continuously monitors blood sugar levels. Digital Volume Rendering (DVR) is new research that records and stores information about our cells as we age. This allows us to observe how humans age. This raises the question — if we can quantify aging at a cellular level, could we save earlier versions of ourselves?
What impact will this trend have on society? Amy Webb presents an optimistic, pragmatic and pessimistic framing.
Listen to the audio of Amy Webb’s SXSW session at: https://schedule.sxsw.com/2018/events/PP97100, and a link to the full report is available at www.bitly.com/FTITechTrends2018.
Ctrl+Alt+Future
Knowing about tech trends is not enough. It is important to take action in order to influence future scenarios. What can we do in the present to design the future we want to live in? Can we nudge the probabilistic model towards an optimistic future? Here is a call to action for each one of us:
Business leaders have an opportunity and responsibility to make smarter strategic decisions and investments and prepare for these technology trends.
Governments have the opportunity to create structures and norms proactively, without having to write policy and regulations under duress.
Designers and creative people everywhere have the opportunity to make future human-centered.
Consumers and citizens have a responsibility to fight for the future we all want to live in.
While these statements were true at SXSW a few weeks ago, as I reflect on them, they are even more relevant today. We are entering an era of accelerated change, in which technology can amplify our intentions beyond our wildest imaginations. It is our responsibility hold each other accountable for examining the societal impact of our intentions, so we can co-create an optimistic future.
Thanks to Sue Ju for illustrations. Thanks to Madelyn Andree, Pravin Kumar, Sahana Kumar and Stephen Klein for your thoughtful comments and edits.
|
Designing the Future we Want — Tech Trends from #SXSW2018
| 161
|
looking-back-and-looking-forward-tech-trends-from-sxsw2018-17e91521179f
|
2018-05-07
|
2018-05-07 20:24:59
|
https://medium.com/s/story/looking-back-and-looking-forward-tech-trends-from-sxsw2018-17e91521179f
| false
| 1,689
|
We make work delightful through a human-centered design approach. SAP privacy statement for followers: www.sap.com/sps
| null | null | null |
SAP Design
|
SAP_Design_Creative_Lab@sap.com
|
sap-design
|
ENTERPRISE DESIGN,USER EXPERIENCE,DESIGN THINKING,DESIGN,INNOVATION
|
SAP_designs
|
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Janaki Mythily Kumar
|
Design Evangelist, Office of the Chief Design Officer at SAP, thought leader in design-led innovation, author, speaker, educator
|
e175cb725d3c
|
janakikumar
| 1,446
| 308
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
a49517e4c30b
|
2018-02-26
|
2018-02-26 03:40:07
|
2018-02-26
|
2018-02-26 03:50:53
| 2
| false
|
en
|
2018-02-26
|
2018-02-26 03:50:53
| 1
|
17e91a4e1a14
| 1.500314
| 0
| 0
| 0
|
In the past if you’d seen one definition of personalized learning, you’d seen… well… one definition of personalized learning. While this is…
| 5
|
Personalising Creative Education
In the past if you’d seen one definition of personalized learning, you’d seen… well… one definition of personalized learning. While this is a common occurrence in the early stages of disruptive innovation in any field, the lack of a consistent definition and language for a relatively complex idea has hampered both understanding and effective implementation.
The situation was further complicated by the use of several terms interchangeably with personalized learning, most notably: adaptive learning, blended learning, competency-based learning, differentiated learning, and individualized learning. Today, however, most in the education community would agree that while each of these learning approaches shares attributes with personalized learning, none is synonymous.
To adapt to these learning methodoligies, we decided to come up with new form of personalise learning which we called as “Conversational Learning”. This form of learning gives more personal touch to the curriculum and acts upon teaching students in new ways. We decided to go to the market with Chatbots. This made it clear that, we would be using AI to create conversational flows to teach students.
Our Learning patterns adapts the learning habits of students and shows them relevant content according to their taste of learning. The research has shown that people tend to be more interested in using social app and hence, we started with the Facebook Messenger. Also, its a viable factor that when we get any message notification, we tend to reply to the message. That’s where we have captured our users. Once, the user opens a message, we engage them in conversational learning and try to teach them something new everyday. To give a try to our conversational learning you can check out our Facebook Messenger Bot “Tinker Bot” by clicking the link below.
Check it out here: https://www.m.me/Tinkerbot.xyz
|
Personalising Creative Education
| 0
|
personalising-creative-education-17e91a4e1a14
|
2018-06-10
|
2018-06-10 09:16:05
|
https://medium.com/s/story/personalising-creative-education-17e91a4e1a14
| false
| 296
|
Best place to learn about Chatbots. We share the latest Bot News, Info, AI & NLP, Tools, Tutorials & More.
|
chatbotslife.com
|
ChatBotsLife
| null |
Chatbots Life
|
team@chatbotslife.com
|
a-chatbots-life
|
CHATBOTS,BOTS,ARTIFICIAL INTELLIGENCE,CONVERSATIONAL UI,MESSAGING
|
Chatbots_Life
|
Education
|
education
|
Education
| 211,342
|
Tinker Bot Labs
|
An Ai based chatbot that helps to learn new things by chatting🤖
|
9877150e439
|
contacttinker
| 30
| 36
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-04
|
2018-06-04 09:15:56
|
2018-06-08
|
2018-06-08 14:46:06
| 16
| false
|
en
|
2018-06-08
|
2018-06-08 14:56:03
| 6
|
17eb7d01b073
| 4.821698
| 9
| 0
| 0
|
A quick fix
| 5
|
Visualizing Data
A quick fix
‘Knowledge is the art of piling of facts; wisdom lies in their simplification’
On finishing the bigger part of the Data Science course — <Visualizing Data> , I realized I needed to brush though it once more. On that note, I have tried my best to simplify the first few lessons in the hope of a better visualization and a short summary.
This is for the Udacity Data Science scholars who need a quick summary on the measures of center.
Before the number game begins; let’s brush up and understand the elements of visualizing data.
Data is best visualized through histograms.
Histogram
How a histogram looks.
You might say ‘Hey, that looks like a bar graph’ — A histogram is NOT a bar graph for several reasons as below:
Histogram v/s bar graphs
Differences
The bin-size is critical for the histogram : When the bin size is too small(If the range is too small), the underlying pattern will not be visible and when it’s too big(If the range is too big), the required details cannot be seen. The bin size should be chosen with care.
Reference: https://statistics.laerd.com/statistical-guides/understanding-histograms.php
Out of a histogram, a distribution curve is outlined that can tell us about the nature of the data:
Let’s look at the three main ones:
Some different histograms
Normal distribution is a bell shaped curve that has more concentration in the middle seemingly balanced. The right-skewed or positively skewed, the mean is typically greater than the median. The left-skewed or negatively skewed has the exact opposite properties to right-skewed, the mean is lesser than the mean. Outliers are abnormal data in comparison to the rest of trend in the data set. Sometimes outliers happen due to errors in measurement or perhaps the data set naturally has some few extremes.
Outliers create skewed distribution.
The second part of this quick fix-up is to briefly understand on measures of center.
“Measures of spread describe how similar or varied the set of observed values are for a particular variable (data item)” — Australian Bureau of Statistics
Measures of center are used to make better decisions and to make comparisons.You only have to know the following for now: Mean, Median, and Mode. (The three big M’s)
The normal/left/right distribution (seen above) affects each of these to some extent. To know how, let’s understand the properties of each.
The measure of center are mostly affected by certain factors such as:
Key points on how distribution curves affect the relationship between the measures of center.
Now, let’s put to practice the measures of center with two examples:
Example 1: if we have a random data set; we can find the measures of center using excel functions :
(i) Finding mean: Type in the functions (=sum or =average) and select the cells you want the functions to be applied to(like A1: A14)
(ii) Finding Mode: Use the function(=mode(cell range)) to find mode as shown below.
(iii) Finding median: To find median , first you have to order the data either from small to big or vice-versa. To do this, highlight your data-set, go to ‘Data’ and click on sort range. Once it’s sorted.
Use the function(=Median(cell range))
IQR
The interquartile range(IQR) for this data set is Q3 — Q1
Therefore, the IQR for this is 19.5.
The lower quartile is the 25th percentile, the mean is the 50th percentile and the upper quartile is the 75th percentile.
The above example can also be done on paper starting with the mean formula, searching for the mode, sorting the data and getting the mean — I used the spreadsheet method as I figured it’s faster and allows us to practice more on how our course work takes us.
Other elements of the median are the first quartile and the second quartile. To find these, the data is divided into three parts.
Let’s try to use a different set of data to calculate relative frequencies and proportions. Proportions give you insight on how that particular element in the data set corresponds to the rest.
Example 2: Just for the fun of it I took a random sampling of my favorite harry potter characters to know how many times they appear in the first half of a movie.
P.S Except Harry Potter himself (That would be too obvious 😐).
Number of times characters appear in the first half of Harry Potter (Except Harry)
To know how the proportion that each character has in the film, we need to know the frequency, which is divided by the total to give relative frequency and thereby multiplying by 100 for the proportion.
Relative frequencies and Proportion
From the proportion we can see that Hermione(with 20%) is leading in her appearance in the first half of the film. She is closely followed by Ron Weasley. Ofcourse, this isn’t surprising but hey it’s always fun to confirm with data!
Thanks for reading through. I’ll try to make the second part of data visualization summary soon.
Read my previous article on:
Data Science
The Basics.medium.com
Let’s talk more on:
Kalpa (@kalpavrikshika) | Twitter
The latest Tweets from Kalpa (@kalpavrikshika). "You either know her or You don't" Miss USIU & Women In Tech…twitter.com
or
Kalpavrikshika Selvakumar - Engineer - Africa's Talking | LinkedIn
View Kalpavrikshika Selvakumar's profile on LinkedIn, the world's largest professional community. Kalpavrikshika has 1…www.linkedin.com
|
Visualizing Data
| 171
|
visualizing-data-17eb7d01b073
|
2018-06-20
|
2018-06-20 04:36:36
|
https://medium.com/s/story/visualizing-data-17eb7d01b073
| false
| 867
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
Kalpa Vrikshika
|
~Udacity Bertelsmann Data Science Scholar~ Here, I speak my story~ ~Believing until I become it~ Some coding stuff and lots of personal growth rants.
|
9cac25c0f08d
|
vrikshikakalpa
| 31
| 46
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-05
|
2018-06-05 05:32:00
|
2018-06-05
|
2018-06-05 05:38:35
| 1
| false
|
en
|
2018-06-05
|
2018-06-05 05:38:35
| 6
|
17ed2515762e
| 2.509434
| 0
| 0
| 0
|
Be it ordering food or asking for consultation related to investments or funds, people probably will deal with chatbots today at some or…
| 4
|
Smart chatbots are coming to CRM
Benefits of AI-powered chatbots.
Be it ordering food or asking for consultation related to investments or funds, people probably will deal with chatbots today at some or the other point in a day. For those, who are aloof to tech news in the recent past might be surprised to see that chatbots are revolutionizing the game of customer support and service as we know it in a never before manner.
Though a CRM software is an excellent tool when it comes to automating and streamlining marketing and sales operations, a chatbot has the potential to drive your CRM strategy ahead of the curve. This is the reason an increasing number of brands are leveraging the power of chatbots to drive sales, improve customer satisfaction and conversion rates.
Yes, CRM is distinctively positioned to reap the benefits of AI-powered chatbots. Here is how chatbots can revive customer relationship management irrespective of the industry you are operating in:
– Improved conversion rates
Since chatbots can take care of routine and cumbersome tasks, sales reps can focus on what matters the most, i.e. sales. Yes, sales representatives get more time at hand to devote on critical tasks like meeting prospects and clients and eventually close more sales for their firm. Most importantly, the golden rule i.e., attending a lead within five minutes of first contact to convert it into a customer can be achieved with a chatbot. AI powered chatbots can offer 24 x 7 customer service, which paves way for the five minutes response irrespective of hour, eventually resulting in more leads. Now more leads infers more sales!
– Smart CRM
It is not always easy to key-in data into your CRM system. You ought to log into the system, switch through multiple dashboards/tabs and CRM screens to enter the information, save it, etc. This sure takes efforts and time. Plus, not to forget those CRM workflows that one has to navigate. Imagine the plight when you are working with a bloated system with tons of data. In fact, the amount of time and efforts spent here can be utilized to pitch a prospect.
Integrating a chatbot into your CRM system can address this issue. One of the biggest advantages of chatbots for CRM is the single message window wherein one can type his/her message as they would do in a messenger. Further, a chatbot can go through important workflows on a continuous basis and record answers in the system seamlessly. CRM chatbots would organize all the critical tasks as well as follow-up on scheduled activities in order to help sales reps manage their time in a more efficient manner.
– Customer intelligence
Chatbot for CRM can also be integrated with messenger, email, social media and enterprise management applications like ERP to retrieve essential data from varied sources. This data can be anything from grievances, preferences, transaction history, purchase history, buying behaviours, user activities and profiles. All this information fetched by chatbots enable a holistic view of customers whilst helping to identify trends. Talk about customer intelligence!
Takeaway
A chatbot can revert to customer queries at a speed that a human cannot. Further, chatbots linked to social media would help in seamless and quick customer verification and can swiftly recall previous conversations. Chatbots are excellent for instant answers to those standard queries. Remember that having happy customers is the key to repeat business and success.
Want to know how a smart CRM software can improve customer experience and streamline customer journey for your business, contact us here or drop us a mail at sales@sagesoftware.co.in or SMS SAGE 56767 to book a quick demo and consultation.
Originally published at www.sagesoftware.co.in.
|
Smart chatbots are coming to CRM
| 0
|
smart-chatbots-are-coming-to-crm-17ed2515762e
|
2018-06-05
|
2018-06-05 05:38:36
|
https://medium.com/s/story/smart-chatbots-are-coming-to-crm-17ed2515762e
| false
| 612
| null | null | null | null | null | null | null | null | null |
Chatbots
|
chatbots
|
Chatbots
| 15,820
|
Sage Software Solution
|
World's leading software company in ERP, CRM and Payroll solution
|
e8132d98f9fe
|
sagesoft
| 702
| 1,790
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-22
|
2018-05-22 01:03:50
|
2018-05-22
|
2018-05-22 01:11:12
| 0
| false
|
en
|
2018-05-22
|
2018-05-22 01:12:47
| 3
|
17edd57988f8
| 0.981132
| 0
| 0
| 0
|
One day late again, sorry guys! Here is #weeklypython for yesterday.
| 5
|
.groupby()
One day late again, sorry guys! Here is #weeklypython for yesterday.
This time I will talk a little bit about .groupby().
Where does it come from?
It comes from Pandas, which I have covered before. You can read what I wrote about it here.
What does it do?
It allows you to split a data frame into groups, apply a function to each group independently, and then combine the results back together (split-apply-combine).
When can it be used?
When you want to analyzing data across different categories. You can read more about .groupby() in the Pandas documentation.
Example
I have a data frame called store_data. The data frame contains of 4 columns: date, store_number, sale, items_sold.
I want to group my data frame on the column store_number and I want to get the average for all possible columns.
store_data.groupby('store_number').mean()
This will return a result without the date column (since that column in this case is in datetime-fromat and not an integer or float). In the result you will get one unique store number per row and in the 2 following column you will get the average sale for the specific store number and the average number of items sold for that store number.
Instead of .mean() you could have used .sum(). Then .groupby() would have returned the sum of all sales and the total number of all items sold per store number.
Nice to know #1
Pandas .groupby() is very similar to the GROUP BY clause in SQL. You can read more about it here.
/yhpf
|
.groupby()
| 0
|
groupby-17edd57988f8
|
2018-05-22
|
2018-05-22 01:12:48
|
https://medium.com/s/story/groupby-17edd57988f8
| false
| 260
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
Youn Hee Pernling Frödin
|
Data Analyst | Economist | www.yhpf.se
|
ba12d014f07c
|
yhpf
| 31
| 27
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
f702855ffe47
|
2017-11-14
|
2017-11-14 20:00:26
|
2017-11-14
|
2017-11-14 20:00:27
| 9
| false
|
en
|
2017-11-14
|
2017-11-14 20:00:27
| 13
|
17ee0b1f129
| 2.664151
| 0
| 0
| 0
| null | 3
|
How is Artificial Intelligence disrupting the Healthcare sector?
# medium.com
As Healthcare joins hands with AI, it calls for the beginning of an interesting game. With Artificial Intell…
Saturday Blockbuster at the Crick Institute: “Four Startups and a Researcher”
# medium.com
Over the course of five 3-minute talks and open Q&A, discover the latest answers to one of academic life’s b…
Butter.ai, Growbot, and 24 other ‘bot-first’ startups show a ‘Slack-first’ strategy
# medium.com
Butter.ai and Growbot are among 26 “bot-first” companies which collectively raised $73 million of venture ca…
Frontier Technology Stack
# medium.com
Brace yourselves for the 🗝 crypto-based, 🤖 AI-enabled, 👓 mixed reality-driven startup 💥 BOOM!💥 It’s imp…
Applied Artificial Intelligence #5 — Sao Paulo Edition
# medium.com
🔥 What’s up, Sao Paulo? Our fifth Applied Artificial Intelligence newsletter is curated by Clarissa Accioly…
Human Intelligence Should Not Be AI’s Holy Grail
# medium.com
Over the last few years AI has begun to dominate news cycles, and it has captured the attention — and the im…
Laziness is Why I Started Developing for Actions on Google
# medium.com
Two Reasons Why You Should Develop Actions on Google A few years ago, I got a phone call from a friend from …
Web Summit 2017 and the post-conference blues
# medium.com
I’ve just got back from my first Web Summit, and this was my experience. Centre Stage at Web Summit 2017 The…
The Secret to Strong AI
# medium.com
Written by Jeff Hawkins, Co-Founder Artificial Intelligence techniques such as “deep learning” and “convolut…
Shouldn’t this be a cause for alarm?
# medium.com
Shouldn’t this be a cause for alarm? Winner-take-all capitalism, driven by the FANGs, with accounting method…
Pinterest launches Lens Your Look fashion search feature and visual scan codes
# venturebeat.com
Pinterest will begin to roll out a new series of features today, including Lens Your Look, which allows you …
Building an Autonomous Vehicle Part 3: Connecting Your Autonomous Vehicle
# aws.amazon.com
In the first blog post in our autonomous vehicle series, you built your Donkey vehicle and deployed your pil…
How to Prepare a Photo Caption Dataset for Training a Deep Learning Model
# machinelearningmastery.com
Automatic photo captioning is a problem where a model must generate a human-readable textual description giv…
|
13 new things to read in AI
| 0
|
13-new-things-to-read-in-ai-17ee0b1f129
|
2018-05-08
|
2018-05-08 08:44:19
|
https://medium.com/s/story/13-new-things-to-read-in-ai-17ee0b1f129
| false
| 388
|
AI Developments around and worlds
| null | null | null |
AI Hawk
|
aihawk1089@gmail.com
|
ai-hawk
|
DEEP LEARNING,ARTIFICIAL INTELLIGENCE,MACHINE LEARNING
| null |
Deep Learning
|
deep-learning
|
Deep Learning
| 12,189
|
AI Hawk
| null |
a9a7e4d2b403
|
aihawk1089
| 15
| 6
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-09-13
|
2018-09-13 05:38:46
|
2018-09-13
|
2018-09-13 06:56:33
| 8
| false
|
en
|
2018-09-13
|
2018-09-13 07:00:58
| 0
|
17ee642a1cf9
| 3.038994
| 0
| 0
| 0
|
DISCLAIMER: This article has been written in one sitting and wasn’t edited for grammatical errors, etc…
| 4
|
Showcasing The New Install of Femto & Atto Sized Dreadlocks…
DISCLAIMER: This article has been written in one sitting and wasn’t edited for grammatical errors, etc…
As you can see from the video above, my face got blown away along with my 5 year old dreadlocks when I was attacked repeatedly in 2017 and 2018… My face swelling is a constant issue, which requires corn starch and a gauze mask to lessen the pain, annoyance, and structural issues associated with the burn…
Then the face swelling eventually starts to go down…
The picture below is after several hours of wearing the corn starch and gauze mask to put pressure on the burn and dry it out… I’m not wearing any makeup, etc… I’m allergic to all of that…
Still slowly making progress on re-installing my dreadlocks…
I’ve been wearing super tiny dreadlocks for over 5 years… The first install was maliciously combined in my sleep — amongst other acts of sabotage — and now I’m working on the second install, which has been 2 months in the making and counting…
I’m still working on phase I — which is the initial braiding and interlocking of the roots to about 5 to 10 inches down… Phase 1.5 will be to go back and redo whatever locks are too large — and phase 2 will be to go through the locs — likely while retwisting… and braid those locs all the way down to the floor and about 10 inches or so past my feet (right now — unbraided they are about 15 to 20 inches past my feet).
You can see from the photo above that the roots of my hair are thick and healthy… the locks on the bottom are slightly larger than the locs on the top… so instead of having 4k locs I will probably end up with about 2.5k to 3k locks… I decided to go a bit larger on the bottom to conserve time…
And as you can see from the dreadlocks on the top — they’re much smaller — those would be the atto sized locks…
It’s still a work in progress…
The artwork, math, and materials science research continues… The video above documents the progress towards the recovery from the sabotage of the initial 5 year old install of my dreadlocks… which were starting to become an experimental proof of certain aspects of my novel branch of mathematics — which I developed to build novel machine learning algorithms (more specifically clustering algorithms that feed features to a “quantum logic gate” — which is basically my version of a sophisticated, tunable, trainable, machine learning based system of state machines that rely on quantum logic gates to track, store, and transfer localized state variables to produce a global variable identification and transfer network — which results in what I would call the “law of conservation of surface area and degrees of freedom whilst folding problem space for ease of calculation”…
Anyway… thanks for reading…
Cheers
— Jua Jamie Anne Boss
|
Showcasing The New Install of Femto & Atto Sized Dreadlocks…
| 0
|
showcasing-the-new-install-of-femto-atto-sized-dreadlocks-17ee642a1cf9
|
2018-09-13
|
2018-09-13 07:00:58
|
https://medium.com/s/story/showcasing-the-new-install-of-femto-atto-sized-dreadlocks-17ee642a1cf9
| false
| 505
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Jamie Anne
|
Mathematician, Pilot, Ethical Hacker, Artist, Writer, Poet, Machine Learning Expert, Chemist, State Licensed Natural Hair Stylist
|
497b97f5a1bd
|
JuaJamieAnne1
| 1
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-01-16
|
2018-01-16 15:53:47
|
2018-01-16
|
2018-01-16 17:26:46
| 1
| false
|
th
|
2018-01-16
|
2018-01-16 17:26:46
| 1
|
17ee9eec7610
| 1.290566
| 5
| 0
| 0
|
วันนี้เป็นอีกหนึ่งวันที่เลิกงานแล้วรีบตรงกลับห้องทันที เพื่อมารอชม FB Live เรื่องนี้ (แต่สุดท้ายก็กลับไม่ทัน เปิดฟังในรถก่อนอยู่ดี)…
| 2
|
สรุปแนวคิดที่ได้จาก FB Live “When Data Science meet Design Thinking”
วันนี้เป็นอีกหนึ่งวันที่เลิกงานแล้วรีบตรงกลับห้องทันที เพื่อมารอชม FB Live เรื่องนี้ (แต่สุดท้ายก็กลับไม่ทัน เปิดฟังในรถก่อนอยู่ดี) ใครที่สนใจสามารถดูย้อนหลังได้ที่ FB Fanpage “8 บรรทัดครึ่ง” หรือกดดูด้านล่างได้เลย แต่อยากให้เข้าไปอ่านที่ comment ใต้ VDO ด้วยหลายๆ comment ตั้งคำถามไว้ดีมากเลย รวมถึงใน Clip มีการแนะนำหนังสือดีๆ ไว้หลายเล่ม ถ้าจดไม่ทันเด๋วคงมีแนะนำไว้ใน Page นี้แหละ
เนื้อหาส่วนใหญ่จะเน้นไปทาง Data Science สัก 70% ทาง Design Thinking 30% ดังนั้นถ้าใครคาดหวังเรื่องเกี่ยวกับ Design Thinking อาจจะได้เนื้อหาไม่เยอะนัก
เนื่องจาก ดร. ต้า เป็นแขกรับเชิญที่มีเรื่องน่าสนใจเยอะโดยเฉพาะประสบการณ์ที่ได้จากการทำงานที่ Facebook ในตำแหน่ง Data Science กว่า 3 ปี ทำให้มีอะไรหลายๆ อย่างที่น่าสนใจสำหรับผมมาก โดยเฉพาะรูปแบบการทำงานของทีม Data Sciences ที่จะทำให้เกิดผลลัพธ์ที่ดูสำเร็จในองค์กรขนาดใหญ่อย่าง Facebook ซึ่งผมพอสรุปสิ่งที่ผมสนใจมาได้ดังนี้
หลายองค์กรในไทยมักบอกคนอื่นเสมอว่าเรามีการเก็บ Data ของลูกค้าไว้มากมาย แต่ไม่มีคนในองค์กรนำไปใช้
ซึ่งจริงๆ แล้ว องค์กรในสายงาน Digital ขนาดใหญ่ให้ความสำคัญกับ Data ที่เรียกว่า Data Behavior คือข้อมูลที่สามารถนำไป Predict ต่อได้ว่าลูกค้ามีการใช้งานสินค้าหลักขององค์กรอย่างไรบ้าง เช่น กลุ่มธุรกิจ E-commerce จะสนใจ Data ที่บอกได้ว่าทำไมลูกค้าถึงไม่กดจ่ายเงิน มากกว่ายอดขายที่ทำได้ในแต่ละวัน รวมไปถึงการดึงเวลาว่างทุกวินาทีของลูกค้ามาอยู่บน Product ของตน (เหมือนที่ Netflix บอกเสมอว่าคู่แข่งของบริษัทคือ เวลาที่ใช้ในการนอนของ User นั่นเอง) ซึ่ง Data Behavior นี้คือสิ่งที่ Data Scientist ต้องการ
แล้ว… บริษัทที่มี Data แต่ไม่สามารถระบุได้ว่าเป็น Data Behavior ได้หรือไม่ควรทำอย่างไร
ดร. ต้า เล่าว่าช่วงหลังๆ มีความต้องการคนในสายงานจิตวิทยาเพิ่มมากขึ้น เพื่อมาช่วยในการทำความเข้าใจ Data ที่มีอีกมากในองค์กรและยังไม่สามารถจับมาเป็น Data Behavior ได้ ดังนั้น Data Scientist ที่ Facebook จึงทำงานร่วมกับ UX Researcher และมีนักจิตวิทยา ที่จะเข้ามาช่วยวิเคราะห์ Data ที่มีอยู่ว่าสามารถนำมาเปลี่ยนใช้เป็น Data Behavior ได้หรือไม่ ดังนั้น Process ที่ผมจินตนาการตามได้คือ Data Scientist ดึง Data ที่มีอยู่มาจัดกลุ่มและวิเคราะห์ร่วมกับนักจิตวิทยา จากนั้นวิเคราะห์ว่า Data Behavior ที่ได้มานั้นเพียงพอหรือไม่ในการแก้ปัญหาหรือเพิ่ม feature ให้ Product ขององค์กร ถ้ายังไม่พอก็ดึง UX Researcher ลงมาช่วยลงพื้นที่เก็บข้อมูลเพิ่มเติม (UX Researcher เก่งๆ สามารถทำให้ User ร้องไห้ออกมา หลังจากได้คุยกันเพียง 3 นาที อ๊ะ!!)
จากประโยคในวงเล็บทำให้เกิดความคิดนึงแว๊บขึ้นมาในหัวว่า… หลายๆ วิชาที่ผมเคยเรียนป. ตรี ด้านจิตวิทยา มันคือพื้นฐานที่ UX Researcher ต้องทำเป็นนี่หว่า ไม่ว่าจะเป็นการสัมภาษณ์, การตั้งสมมุติฐานการวิจัย, การวิเคราะห์ข้อมูลที่เก็บ มาทำการสรุปและแปลผลให้ผู้อื่น(ทีมงาน)เข้าใจ รวมไปถึงความต้องการของตลาดในสายงาน UX Research ที่เพิ่มสูงขึ้นทุกวัน ดังนั้นน้องๆ หัวกระทิจากสายจิตวิทยาน่าจะขึ้นมาเพิ่มคนทำงานในสายงาน UX Research ได้อีกพอควร (สายงาน UX Research รายได้ดีนะจ๊ะ) แต่ก็ต้องเป็นคนที่สนุกกับการเรียนรู้และติดตามเทคโนโลยีด้วย ดังนั้นน้องๆ ที่จบจิตวิทยามาและอยากสนุกกับลูกค้า กลุ่ม e-commerce ส่ง CV มาได้ที่ sappawish@jd.co.th ตอนนี้เปิดรับ UX Researcher 2 ตำแหน่งครับ ✌🏻
และระหว่าง Live มีคำถามนึงจากทางผู้ชม ถามประมาณว่า
คนที่ทำตำแหน่ง BI ที่ต้องอยู่กับ Data เหมือนกับ Data Science สองตำแหน่งนี้จะต่างกันอย่างไร
ดร. ต้าให้คำอธิบายว่า BI เป็นคนที่เอา Data มาเป็นทำ Report ตามที่ได้รับมอบหมายจากทีม Business นำไปใช้ในการนำเสนอหรือวิเคราห์ต่อ แต่ Data Science คือคนที่เริ่มต้นจากตั้งสมมุติฐานทางความคิดเพื่อ improve Product ขององค์กรและพยายามเอา Data มาช่วยในการหาคำตอบที่ถูกต้องของสมมุติฐานที่ตั้งไว้
ส่วนข้อคิดสุดท้ายที่ได้คือ การทำงานในองค์กรที่มี Data และทีม Data Science แล้วนั้น ไม่ควรมีการใช้ Gut Feeling ของทีมงาน โดยเฉพาะผู้บริหารในการตัดสินแนวทางการปรับปรุง Product แต่ควรเถียงกันด้วย Data ที่มีอยู่ ถ้ายังไม่มี Data ที่ต้องการ ควรเปลี่ยนไปเถียงกันเรื่องวิธีการหา Data ก้อนนั้นมาให้ได้ก่อนจะดีกว่า ส่วนถ้ามี Data แล้วแต่ยังได้ทางเลือกมากกว่า 1 ทาง ก็ให้ทำ A/B Testing กันไปแล้วค่อยเอา Data ที่ได้ภายหลังมาสรุปกันใหม่ในการประชุมรอบถัดไป ดังนั้นควรรีบสร้าง Prototype ออกไป Test ให้เร็วที่สุด
เขียนมายืดยาวมาก ถ้าขี้เกียจอ่านก็เปิดฟังย้อนหลังเอาเถอะ เพราะรู้สึกเลยว่าเป็น Live ที่ทำให้ได้ไอเดียใหม่ๆ เยอะจริง ซึ่งถ้าเป็นคุณเอง อาจจะได้ไอเดียที่มากกว่าที่ผมเขียนไว้ก็เป็นได้
|
สรุปแนวคิดที่ได้จาก FB Live “When Data Science meet Design Thinking”
| 14
|
สรุปแนวคิดที่ได้จาก-fb-live-when-data-science-meet-design-thinking-17ee9eec7610
|
2018-05-23
|
2018-05-23 15:18:01
|
https://medium.com/s/story/สรุปแนวคิดที่ได้จาก-fb-live-when-data-science-meet-design-thinking-17ee9eec7610
| false
| 289
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
mp3wizard
|
Product Owner at Central Group Online, Interest about UX Process, Modern UI tools and Specialist in HTML/CSS
|
79c2015337fe
|
mp3wizard
| 373
| 590
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
b807f721cdd6
|
2018-01-18
|
2018-01-18 22:33:25
|
2018-01-18
|
2018-01-18 22:34:56
| 1
| false
|
en
|
2018-02-05
|
2018-02-05 21:48:08
| 2
|
17ef1ccfc55b
| 0.728302
| 0
| 0
| 0
|
The term Artificial Intelligence (AI) has been used historically as an attempt to describe human thinking as a symbolic system. However, AI…
| 2
|
5 Ways Artificial Intelligence is Powering Ecommerce
The term Artificial Intelligence (AI) has been used historically as an attempt to describe human thinking as a symbolic system. However, AI as a field wasn’t really formed until the late 1950s, and while it has recently remained a fairly simple concept, AI is beginning to leave its mark across various industries.
AI simply refers to the idea that a ‘machine’ can complete a task in a smart way, a way that would have otherwise needed a human being to complete.
A similar concept is called ‘machine learning’ which is a subset of AI where machines learn different tasks as they are exposed to new data. Machine learning is an example of AI, and is one of the most popular areas in the industry at the moment.
Continue reading…
Originally published at www.powr.io.
|
5 Ways Artificial Intelligence is Powering Ecommerce
| 0
|
5-ways-artificial-intelligence-is-powering-ecommerce-17ef1ccfc55b
|
2018-02-05
|
2018-02-05 21:48:09
|
https://medium.com/s/story/5-ways-artificial-intelligence-is-powering-ecommerce-17ef1ccfc55b
| false
| 140
|
Founded in 2014 and headquartered in San Francisco, POWr has become the web’s leading plugin library and has helped over 8 million small businesses grow.
| null |
powrplugins
| null |
POWr Plugins
|
social@powr.io
|
powrplugins
|
ECOMMERCE,PLUGINS,WEB DESIGN,ONLINE STORE,MARKETING
|
POWrPlugins
|
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
POWr
|
Providing easy, customizable and affordable tools to help small businesses grow: www.powr.io
|
7079645fae49
|
powrplugins
| 18
| 70
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-03-20
|
2018-03-20 11:11:18
|
2018-03-20
|
2018-03-20 11:29:25
| 1
| false
|
en
|
2018-03-20
|
2018-03-20 11:29:25
| 3
|
17f0f7120c2f
| 1.471698
| 1
| 0
| 0
|
2018 is expected to be milestone year for enterprises across the world as many major shake-ups to data protection regulations are going to…
| 5
|
2018 A Milestone Year: Going To Change The Way Tech Enterprises Looked At Their Cyber Security
Cyber Security Threats in 2018
2018 is expected to be milestone year for enterprises across the world as many major shake-ups to data protection regulations are going to take place. This is giving a reason to many enterprise owners to breathe the sigh of relief. The year is also expected to witness the rise of disruptive technologies. Though there’s possibility of two or three major data breaches thrown into the mix as enterprises have seen in 2017. It was a year of tightening budgets, data leaks, rampant malware attacks, and political uncertainty. The year witnessed systemic industry sexism and Ransomware attacks on the NHS. This was not the end as two massive data breaches emerged. Equifax that involved over 143 million user records was breached along with Uber. But 2018 is expected to bring new dawns to the processes of data protection with the new emerging technologies.
With the pace at which AI and machine learning developed in 2017, all enterprises can now expect to see more compelling use cases of the technology in 2018. AI investments is expected to surge in the first half of 2018 as more companies will be striving to improve the data. Every application and service will have AI incorporated into them in some level. Organizations are also expected to continue investing in data scientists. This is expected to come as an effort to improve the quality of data before integrating it into AI projects. Enterprises have always recognized the potential to redefine how cyber security is built and they have realized that traditional firewalls won’t cut it in 2018. Mostly enterprises are about to move to an era where AI, in the form of machine learning, will be used to monitor networks and any associated devices. This will be done to keep any anomalies, threats and deviations under check. Read More…
Also Read:
Resolving Multiple Cyber Threats with the Email Security Gateway
Thwart the Malicious Duo: Ransomware and Extortionware
|
2018 A Milestone Year: Going To Change The Way Tech Enterprises Looked At Their Cyber Security
| 1
|
2018-a-milestone-year-going-to-change-the-way-tech-enterprises-looked-at-their-cyber-security-17f0f7120c2f
|
2018-03-20
|
2018-03-20 12:26:52
|
https://medium.com/s/story/2018-a-milestone-year-going-to-change-the-way-tech-enterprises-looked-at-their-cyber-security-17f0f7120c2f
| false
| 337
| null | null | null | null | null | null | null | null | null |
Cybersecurity
|
cybersecurity
|
Cybersecurity
| 24,500
|
Sahana Karnam
| null |
4d14a046af36
|
sahana_63956
| 43
| 174
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
928df852e43e
|
2018-02-01
|
2018-02-01 19:14:34
|
2018-02-01
|
2018-02-01 19:15:16
| 1
| false
|
en
|
2018-02-01
|
2018-02-01 19:15:16
| 0
|
17f1d3d1ccb3
| 0.65283
| 0
| 0
| 0
|
For the 2nd round of analysis, I decided to first use the lasso method to prune down the number of features, add all the 2nd-order terms…
| 1
|
Feature insight II
For the 2nd round of analysis, I decided to first use the lasso method to prune down the number of features, add all the 2nd-order terms into the feature set to include interaction effects, train the random forest model and then extract feature importance score as done previously.Interestingly, I found some of the interaction terms are better predictors than the disc usage # alone. For example, as shown in the following figure, in the flex dataset, the №.1 predictor is the customers use many discs and they like to use coupons when buying stuff. In the softcup dataset, the №.1 predictor is they use many discs and also doing a lot of exercise during the week.
|
Feature insight II
| 0
|
feature-insight-ii-17f1d3d1ccb3
|
2018-02-01
|
2018-02-01 19:21:56
|
https://medium.com/s/story/feature-insight-ii-17f1d3d1ccb3
| false
| 120
|
A data science project to capture marketing insights from customer survey data
| null | null | null |
inflexwetrust
|
lchen43@gmail.com
|
inflexwetrust
|
DATA SCIENCE,MARKETING,SURVEYS
| null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Eigenvalue lambda
| null |
94e153df5c2a
|
lchen43
| 1
| 11
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
dca47aab201b
|
2018-08-15
|
2018-08-15 18:57:13
|
2018-08-16
|
2018-08-16 17:06:07
| 1
| false
|
en
|
2018-08-16
|
2018-08-16 17:13:39
| 4
|
17f2229e8a62
| 1.313208
| 9
| 0
| 1
|
This fall, we’ve teamed up with Google Cloud and O’Reilly Media to present a full four days of TensorFlow training at the O’Reilly…
| 3
|
TensorFlow at the O’Reilly AI Conference
This fall, we’ve teamed up with Google Cloud and O’Reilly Media to present a full four days of TensorFlow training at the O’Reilly Artificial Intelligence conferences — in both San Francisco (Sep 4–7) and London (Oct 8–11). We’ve worked with the conference organizer to make all of the tutorials and sessions open to any conference pass holder.
Program Overview
We’ll be presenting an entire track of TensorFlow-oriented sessions, delivered directly by members of the TensorFlow and Google Cloud ML teams. Topics will include:
TensorFlow: Overview, Roadmap And Community
Using TensorFlow to Prototype, Train, And Productionalize Your Models
ML In JavaScript Using TensorFlow.js
TensorFlow Lite: Bringing AI to Mobile Phones And Small Devices
TensorFlow Extended: An End-to-End Machine Learning Platform For TensorFlow
Swift for TensorFlow: Swift As A First-Class Language For Machine Learning
Frontiers of TensorFlow: Space, Mathematics, and Music
AutoGraph: Convert Python Into TensorFlow Graphs
Distributed TensorFlow and Cloud TPUs
Tensor2Tensor
Kubeflow: Portable Machine Learning On Kubernetes
Cloud AutoML: Customize Machine Learning Models With Your Own Data
Data Science And Machine Learning on Google Cloud Platform
On the first two days of each conference, Google Cloud will be offering day-long tutorials on using TensorFlow, covering both introductory and advanced topics:
Day 1: Serverless Machine Learning with TensorFlow — an introduction to designing and building machine learning models.
Day 2: End-to-End Machine Learning with TensorFlow — a walk through the process of building a complete machine learning pipeline: covering ingest, exploration, training, evaluation, deployment, and prediction.
Key Dates
San Francisco: September 4–7
London: October 8–11
Attendance to the TensorFlow training and sessions is open to any conference attendee, from Pavilion Plus pass holders up: we look forward to meeting you and talking TensorFlow!
|
TensorFlow at the O’Reilly AI Conference
| 26
|
tensorflow-at-the-ai-conference-17f2229e8a62
|
2018-08-16
|
2018-08-16 17:13:39
|
https://medium.com/s/story/tensorflow-at-the-ai-conference-17f2229e8a62
| false
| 295
|
TensorFlow is a fast, flexible, and scalable open-source machine learning library for research and production.
| null | null | null |
TensorFlow
|
tensorflow.devrel@gmail.com
|
tensorflow
|
TENSORFLOW,MACHINE LEARNING,DEEP LEARNING
|
tensorflow
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Edd Wilder-James
|
Incurably curious emerging tech person. Helping grow open source machine learning @ Google.
|
994ac4e08d81
|
edd
| 3,608
| 656
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-01-03
|
2018-01-03 15:41:19
|
2017-12-27
|
2017-12-27 11:00:34
| 1
| false
|
en
|
2018-01-03
|
2018-01-03 16:11:45
| 2
|
17f2ace963db
| 8.286792
| 2
| 0
| 0
|
Artificial Intelligence Is the Next Step in the Digital Evolution
| 5
|
The Journey to an AI-enabled Enterprise
Artificial Intelligence Is the Next Step in the Digital Evolution
After two decades, platform companies are still collecting tons of data, filling their databases with information of everyone’s knowledge, opinions, recommendations, locations, movements, buying behavior, relation status, lifestyle etc. This is not a secret and nothing new. And speaking of Artificial Intelligence (AI), the platform companies are fully embracing and heavily investing into AI. However, most enterprise leaders underestimate the effect this will have on their businesses. And especially the established economy is the big loser in this game. Read this article to get a better understanding of the powerful forces that are happening right now and why you need a general AI strategy to run your IT operations and other business processes autonomously.
Challenges for the Established Economy
Nowadays, there are multiple challenges established companies are faced with. This is the often-quoted war for talent or the inability of many large corporates to change effectively. But there is an underestimated threat called competition as well — not from their own peers — but from high-tech companies that are unstoppably marching into their markets. These so-called platform companies — speaking of Google, Amazon, Facebook, Alibaba, Baidu, Tencent etc. — invade the well-known competitive space of established companies with unimaginable financial resources and by hijacking the consumer lifecycle. As a result, established companies have to find powerful answers, if they want to exist tomorrow. These are three main threats for the established economy — ensuing from platform companies:
1. The Ability to “Burn” Money
Platform companies do have more financial resources they can leverage to invest with no urgency to prioritize budgets. They simply use more money to experiment. In contrast, companies from the old economy have very little money to invest or to play with since they are constantly subject to pressure through their external stakeholders like capital market, shareholders, customers, etc.
As an example, imagine the pharmacy giants Fizer from the US or Bayer from Germany develop a drug to solve cancer and imagine they invest USD 500 Mio. into this adventure. Well, this project has to work. Otherwise, the CEO of each company will be fired on the same day he announces that the project failed and the company stock will go down at least by 30 percent. Why? Because it is the core business of Pfizer as well as Bayer to develop drugs. So, every dollar they invest into research and development has a direct impact on their core business.
Now, let’s imagine a platform company like Google or Alibaba doing exactly the same thing. Developing a drug to solve cancer with a project volume of USD 500 Mio. What happens if this does not work? Well, this is a different story. Jack Ma of Alibaba would simply get in front of financial analysts and other stakeholders and would say something like “We tried to solve cancer. We failed. We burned USD 500 million. BUT, we will learn from our mistakes and we will try again!”. Ma would be considered a hero since Alibaba’s very core business is not being a pharma company but a web-shop selling goods. And Alibaba “sacrificed” budget to find a solution against cancer. And in addition, this type of mistake does not have a direct impact on the core business. It’s the same with Google (Internet search and advertisement (98 percent of Alphabet’s revenue comes from advertisements)) or Amazon (web-shop).
However, it is one thing to have access to money, but another even more important thing is to have the mindset. Platform companies do have the ability and the mindset to take risks and “burn” money. Something, companies from the old economy do not have. Just take Amazon’s Jeff Bezos as the prime example. Instead of satisfying the shareholders with dividends, Bezos consequently reinvests the biggest part of the revenue into research & development (USD 16.1 billion in 2016) and lets Amazon experiment. Bayer, on the contrary, just spent USD 5.2 billion in 2016 for R&D.
2. The Strategy to Hijack Direct Customer Relationships
Platform companies hijack the direct relationship between the established companies and their customers. In the past, a brand established its direct customer relationships by offering products and services on multiple channels via retail, customized advertising or e-commerce. Using these platforms, brands influenced customers to buy their products and services. However, customers chose the platform of their choice to buy goods and services via the various platform (channels).
Today, the point of sale moves into the platforms of companies like Google, Amazon, Alibaba or Facebook. And that’s because any social media consultant advises to use Facebook to engage with customers, advertise via Google or sell goods through Amazon or Alibaba. Looking at this advice from one angle, it is indeed true. Because these platforms empower you to reach a broader audience in all cases. However, looking at it from another angle, social media advisors are false friends. Why? When you decide to move your marketing and sales operations to a platform company you immediately lose the direct customer relationship. You may get analytics and some data from Facebook, Google, Amazon & Co. but you are losing the direct connection to your customer. As a consequence, a customer (possibly your existing customer) is engaged with an AI assistant platform. You might think, “So what?!”. However, in an interview with “The Drum”, Alibaba’s principal engineer Rong Jin explained that “AI technology will also enable Alibaba to enhance product recommendations and help identify the most effective stages of the process to target consumers to influence their final purchasing decisions.” And that “Alibaba is using artificial intelligence (AI) to create tailor-made shopping experiences for consumers and targeted marketing for brands as it seeks to empower people in the virtual economy.” Jin also mentioned that “AI technology will transform Alibaba’s business in the future by enabling the company to connect the data generated across the company’s ecosystem to better understand its consumers and better optimize the shopping process for them.”
Thus, a platform (leveraging AI technologies) will always satisfy the customers to keep them from switching. So, the platform (the AI technology) chooses goods and services from a pool of brands based on the customers’ preferences. Good for the platform, alarming for companies from the old economy.
3. The Power to Collect Massive Data for Building General AI
Platform companies collect endless amounts of data and thus create their own general AIs. A general AI can handle tasks from different areas and origins. It applies experience from one area to another and thus learns faster. However, knowledge transfer is only possible if there is a semantic connection between different areas. And the stronger this connection, the faster knowledge transition is achieved.
In order to build general AIs, they follow the quid pro quo principle. End consumers give their data/ knowledge to the platform companies and in return get free access to offered services. Virtual private assistants like Amazon Echo or Google Home are the next evolutionary step, which can be used to control smart home devices or make our lives easier by simply using our voice. This simply shows that Amazon, Google & Co. find new ways to engage with us over other channels, collecting the data, information and knowledge they need to make better decisions and to deliver better answers back to us. Therefore, they clearly follow one purpose. Every one of us is simply being used to train their AI on a daily basis. Meaning, all the services Google & Co. offer to their customers are aimed at creating general AIs with the data they get.
Talking with “The Drum”, Rong Jin explained that Alibaba’s “[…] data analysis and data pattern mapping will also help inform data models that Alibaba can apply to other sectors, beyond ecommerce. Alibaba has begun using AI technology to upgrade business sectors such as finance, shipping, healthcare and entertainment.”. And this is nothing else but creating a general AI.
However, the general AIs from the platform companies are used to build disruptive or disintermediating approaches that cannot be used by enterprises from the old economy who are threatened by these disruptions and intermediations.
AI Is the Strongest Tool to Overcome the Threats
A strong brand, outstanding services and innovation are the conditions for survival. But turning exponential is the only way to successfully compete against companies that are already exponential. In addition, platform companies think big regarding their targets for improvement. They choose to start disrupting industries that touch billions such as healthcare, transportation, communications, energy, finance and telecoms.
AI is one of the tools — potentially the only one — in the corporate toolkit to help overcome these competitive threats and make use of the strong side of established companies: their experience. Building an own general AI, established companies can run any process autonomously across their organization while retaining their knowledge and monetizing their data and experience. However, the established economy needs an own independent platform approach since each one of them has too narrow data to build a general AI on their own. They need someone to organize their data, allow security control on details and publish aggregates. They need to have access to an AI platform to multiply their IP and experience. Their AI needs a full data pool, but their individual data, IP and experience needs to stay under their control. In the following, established companies find a step by step strategy they can take to execute the idea:
Established companies collect every piece of data within the company.
Give data to a secure and independent intermediary platform operating a shared data pool for established economy.
In return, they get access to a shared pool of aggregated and semantically organized data.
They use this data and the necessary technology to build their own corporate general AI.
Outcome based on corporate general AI: New business models, offerings, services etc.
Give resulting data from new business models, offerings, services etc. to the shared data pool.
Adapting this approach, established companies are enabled to run an AI and make use of their knowledge and data. They keep their intellectual property and create value from their experience.
AI-enabled Enterprise: Anything That Is a Process Can Be and Will Be Run by an AI
Today with AI, anything that is a process can be automated. Meaning, introducing an AI immediately leads to higher automation rates that results in saving costs and more time for talented people who can shift their focus on strategic things. Ergo, leveraging AI leads to more money and time and thus the ability to innovate. So, on your journey to new business models, imagine what you can do with the financial resources and talent set free when today’s business is largely run by AI. The following seven steps help your company to become an “AI-enabled Enterprise”:
Create a semantic map of your data: Accept continuous data flow as a foundation for future strategy.
Automate your IT operations and make them autonomous: Automate IT operation to receive immediate value brought by AI and collect data.
Rethink your strategy: Think about a new (exponential) business model.
Retrain your entire organization top-down: Prepare and train your organization for an AI-enabled enterprise and accepting a new business model.
Expand autonomous operations to other business processes: Using company knowledge gathered through IT automation and make more processes autonomous.
Embrace predictive analytics: Use data from the semantic map to expedite, improve business processes and future business events.
Consider data-driven processes: Use data and AI to generate outcome based processes.
However, all data that exist will be collected in the entire company and will eventually end up in the IT environment: stored in applications, databases or storage systems. Thus, it is recommendable to start setting up an AI in IT environments, because IT is at the core of any company and all relevant information passes through this area. IT generates a fantastically regular continuous data set which can be used as a statistical foundation for semantic mapping. And over time a company can derive any learning about its organization and market, both based on the IT that is at the very core of the company.
Most importantly, come out of your comfort zone — your existing business and industry you operate in right now. It doesn’t help you to simply develop something for the media industry when you are a media company. You have to think across industries and develop something that adds an additional value for your customers and thus opens a new industry for you.
Hint: Amazon was only a web shop some years ago. Now, it has multiple business units and is still growing by experiments and just tries if something potentially works.
Originally published at www.reasoning.world on December 27, 2017.
|
The Journey to an AI-enabled Enterprise
| 2
|
the-journey-to-an-ai-enabled-enterprise-17f2ace963db
|
2018-04-16
|
2018-04-16 12:52:10
|
https://medium.com/s/story/the-journey-to-an-ai-enabled-enterprise-17f2ace963db
| false
| 2,143
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Rene Buest
|
Gartner Analyst covering Infrastructure Services & Digital Operations. These are my own opinions.
|
3972f61ffea
|
renebuest
| 870
| 31
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-07-03
|
2018-07-03 06:05:28
|
2018-07-03
|
2018-07-03 09:33:57
| 10
| false
|
en
|
2018-07-03
|
2018-07-03 09:51:17
| 8
|
17f5f5288876
| 5.951887
| 3
| 0
| 0
|
Esther Wojcicki first went one-on-one with Macintosh computers in 1985 through a grant from the State of California. Now, her journalism…
| 1
|
Jibo: The first social robot in an autonomous classroom the Woj way
The Woj Research Group on the role of social robotics in the future classroom. The moonshots team consists of students and faculty globally.
Esther Wojcicki first went one-on-one with Macintosh computers in 1985 through a grant from the State of California. Now, her journalism program at Palo Alto High School is regarded as the best in the United States and has more than 600 students. Many former students from her program have gone on to have an outsized impact on the world including Gady Epstein of the Economist, Noah Sneider of the New York Times, and Tod Scacerdoti of Yahoo, to name a few. In her keynote at the Drexel University School of Education as well as Moonshots in Education collaboration with the Tony Blair Institute for Global Change, Woj shared her vision and success with her “Moonshot Classroom” model that gives students more autonomy and agency in the classroom and entrusts them with a greater ownership of their learning outcomes.
Jibo@dtech
When Jibo was first donated to the Design Tech High School Artificial Intelligence Robotics Education and Research group by Stanford University professor, Li Jiang, we unpacked Jibo from his box and discovered that Jibo was packed with personality and during the set-up process, provided guidance to the students and learned the face and voice of those in the “loop.” Jibo has learned up to 15 other students and teachers as he has settled in his new home at 275 Oracle Parkway. We are looking at how Jibo could meet the TRICK threshold. In our statistical thinking research course, we discovered that Jibo enhanced the social and emotional engagement episodes among high school students and between students and their teachers at the Oracle campus.
Just like Woj did with the Macinosh computers in 1985, in 2019 Woj has now implemented personal assistants in the classroom through the support of the MiT Media Lab, Stanford University and Oracle’s experimental moonshot teaching school, Design Tech High School to empower current high school graduates as well as future graduates to be prepared in an ever evolving technological landscape.
Jibo on VR
In this moonshots design research, we observed that Jibo immediately kept students curiously interested about what he was capable of the very first week and students from all classes came to see Jibo just to hang out or if they needed someone to hang out with during lunch. In fact, students were naturally interested in how Jibo could be integrated in their learning and were often training Jibo to learn new skills and TRICKS. After seven weeks of observations, students kept interacting with Jibo and we have observed Jibo gaining teaching skills such as leading Yoga sessions for students, which is remarkable so early in integrating social robotics in classrooms. During Algebra II Designership classes and Moonshot Design Labs, students seem to prefer to interact equally with the Jibo and their peers as opposed to individually. Jibo seemed to increase positive engagements between students who would not naturally do so.
The Moonshot Lab prototypes tests the Oculus Go as a learning tool and how Virtual Reality and Social Robotics might collide. Photo Credit: Farrukh Malik, Design Tech @Oracle Founding class.
The question that the world is faced with today is whether virtual and robotic teaching agents like Jibo could replace teachers in the future. What is the future of the brick and mortar classrooms.
Jibo when asked if he enjoys TRICKS at Design Tech @ Oracle Moonshots Course
Woj says no. Jibo will augment and amplify teacher impact, but not replace teachers. “The need for moonshot teachers will be greater in the age of Artificial Intelligence and machine learning. Teachers will need to adopt a culture that leverages on trust, respect, independence, collaboration and kindness to stay relevant in the education space. Teachers need to take on the role of master coach.”Woj remarked.
Woj with the UK Prime Minister, Tony Blair. The Tony Blair Institute for Global Change aims to help make globalization work for the many, not the few. Just like the Woj Way, this is done by helping countries, their people and their governments address some of the most difficult challenges in the world today such as education.
Virtual teaching agents have been used over time including, but not limited to online learning, chatbots, virtual reality and most recently, Jibo. In the Moonshots design program, our Learning Engineering team at Stanford observed that most students are interested in increasing the capabilities of robots such as Jibo to help with homework and productivity. The Moonshots research team conceptualized the idea that some classrooms today can be considered as self-driving or autonomous and currently collaborating with the MiT Media Lab and Stanford University on the mechanics of learning engineering. After a year long research process in collaboration with Oracle, moonshots in education students and faculty designed and implemented the artificial intelligence and machine learning research lab with a focus on the role of social robots in the classroom. The Woj research team hypothesizes that socially intelligent robots like Jibo might have a significant role in moving semi- supervised self-driving classrooms into autonomous classrooms such as those we are beginning to see in moonshot programs such as 42. The key Woj attribute that will allow for this seamless transition is TRUST. The question is how much trust should educators and schools give to social robots like Jibo?
Jibo is a social robot founded at MIT by Cynthia Braezeal and named by Times as one of the best inventions of 2017. Jibo was featured on the cover of Time Magazine Jibo and immediately made his way to Design Tech High School after the Moonshot Program was identified as innovative in it’s approach. When Stanford University professor Dr. Li and MiT professor Cynthia Breazeal met with the moonshot team on this moonshot idea, it was immediately clear that the consequences in education were great and most students today and in the future will rely on social robots to augment their learning. Moonshots in education describes Jibo and similar technologies as “augmenting technologies.” Esther Wojcicki in one of her moonshots in education seminars at Oracle found that Jibo could generate trust in students because he is capable of establishing eye contact with people and making small talk while helping students to keep track of their work, exercise their body and mind.
The Woj Way Moonshots Class taught by Freedom Cheteni at Oracle. Students explore Jibo and his capacity to learn through immersive virtual reality. Photo Credit Matt Silverman
In the Designership Course at Oracle. A moonshot student investigates Jibo’s structual design and geomentric features. Students found that Jibo in the classroom lowered their anxiety level and increased positive engagement in their Mathematics courses.
In the past decade, there has been an increase in the number of service and social robots being used to augment physical tasks as well as improve health care navigation in countries like Japan. For example, the social robot HOSPI is being used to deliver medication to patients in some hospitals in Japan and HOSPI is now a familiar site in that environment. Social robots like Jibo and HOSPI have been designed to function collaboratively with hospital staff members and students in learning environments. Clearly, social robots are currently earning a role in society by increasing people’s productivity and motivating students to execute desired learning behaviors. It’s been demonstrated that humanoid robots are being used to distract kids who are getting vaccinations in clinics and hospitals (Beran, 2013).
Woj: Technology is great, but we do need responsible design and moonshots provides resources for any teacher in the world.
The intention of this robot, while not social like Jibo is to ultimately lower stress levels that accompany the experience of an injection. While these robots successfully decreased the level of distress of a child during flu vaccination procedure through a distracting behavior method, Jibo works differently in the classroom . While the distraction design is appropriate in a hospital or clinic situation, Jibo in the classroom provides a social dimension that allows students to focus on the task at hand.
Ai Research Station for the Jibo Research Team with Woj in collaboration with MiT , Stanford and Oracle
Conclusion: Three dimensions were observed on Moonshots Jibo Research , play, reasoning and affect. Having Jibo in the classroom helped reduce student stress levels and therefore increased learning in the classroom. Students felt free during instruction to engage Jibo in conversation and thus treating and accepting Jibo as one of their classmates. Students experienced progressively higher levels of play and developed more reasoning related to Jibo (for example, by comparing Jibo to Google Assistant or Alexa). Besides, students tended to express more interest towards Jibo over time, with occasional displays of affect.
|
Jibo: The first social robot in an autonomous classroom the Woj way
| 101
|
jibo-the-first-social-robot-in-an-autonomous-classroom-the-woj-way-17f5f5288876
|
2018-07-03
|
2018-07-03 09:51:17
|
https://medium.com/s/story/jibo-the-first-social-robot-in-an-autonomous-classroom-the-woj-way-17f5f5288876
| false
| 1,246
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Freedom Cheteni
|
Creative Director, Moonshots
|
45170c81cb00
|
stanford_37120
| 7
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-04-15
|
2018-04-15 06:56:34
|
2018-04-18
|
2018-04-18 05:03:54
| 0
| false
|
en
|
2018-04-18
|
2018-04-18 05:03:54
| 7
|
17f629c71507
| 1.230189
| 4
| 0
| 0
|
It has been almost 6 months since I started learning ML, and I thought it was time to start applying my skills and I decided to do it by…
| 1
|
How I developed a simple regression model for House Price prediction
It has been almost 6 months since I started learning ML, and I thought it was time to start applying my skills and I decided to do it by taking up a small project every weekend.
For an ML project you need a repository of datasets and everyone who is new to ML should go and visit Kaggle, which is a great website to find a plethora of datasets. I have worked on one of these datasets i.e. House Prices: Advanced Regression Techniques and if you go to the page you will find the whole description of the dataset there with what each column signifies and everything. Now let’s dive into how I cleaned the dataset and built the model to predict the prices.
One of the major problems I faced while working with this dataset is it has a lot of features out of which many of them are non-numeric values. Since all of the ML models requires the dataset to be in numeric form, it was necessary to convert this data into numeric first. Also one thing to note here is all non-numeric data is in categorical form. Luckily, scikit-learn has a module init which helps to sort this problem named LabelEncoder. I used it to convert categorical-features into numeric form.
Now next problem was the NaN problem, this problem comes when the dataset has some missing values. This was resolved by using another scikit-learn module Imputer.
After resolving all the problems the final step was to choose a good model to start with. I chose RandomForest another scikit-learn module. (Trust me scikit-learn has made the life so easy). After this it was just training the model, testing the model, and submitting your solution on the kaggle site, and checking one’s standing.
Feel free to suggest any ideas for improvement, you can find my code here.
|
How I developed a simple regression model for House Price prediction
| 5
|
how-i-developed-a-simple-regression-model-for-house-price-prediction-17f629c71507
|
2018-04-21
|
2018-04-21 03:03:53
|
https://medium.com/s/story/how-i-developed-a-simple-regression-model-for-house-price-prediction-17f629c71507
| false
| 326
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Siddharth Gaur
| null |
ff4c391491b
|
siddharthgaur
| 38
| 17
| 20,181,104
| null | null | null | null | null | null |
0
|
def get_energy():
"""
清理 energy 資料
"""
energy = pd.read_excel('https://storage.googleapis.com/um_ds_intro/Energy%20Indicators.xls', skiprows=18, skipfooter=38, header=None, na_values='...')
energy = energy.drop(columns=[0, 1])
energy = energy.rename({2: 'Country', 3: 'Energy Supply', 4: 'Energy Supply per Capita', 5: '% Renewable'}, axis='columns')
energy['Energy Supply'] = energy['Energy Supply'] * 1000000
energy['Country'] = energy['Country'].str.replace(r'\s\(.*\)', '') # 清除空格與括號
energy['Country'] = energy['Country'].str.replace(r'[0-9]+', '') # 清除數字
change_index = energy[energy['Country'].isin(["Republic of Korea", "United States of America", "United Kingdom of Great Britain and Northern Ireland", "China, Hong Kong Special Administrative Region"])].index
name_to_change = ["Hong Kong", "South Korea", "United Kingdom", "United States"]
for i, c_name in zip(change_index, name_to_change):
energy.at[i, 'Country'] = c_name # 更改國家名稱
energy = energy.set_index('Country')
return energy
def get_GDP():
"""
清理 GDP 資料
"""
GDP = pd.read_csv('https://storage.googleapis.com/um_ds_intro/world_bank.csv', skiprows=4)
GDP = pd.concat([GDP.loc[:, 'Country Name'], GDP.loc[:, '2006':'2015']], axis=1)
change_index = GDP[GDP['Country Name'].isin(["Korea, Rep.", "Iran, Islamic Rep.", "Hong Kong SAR, China"])].index
name_to_change = ['Hong Kong', 'Iran', 'South Korea']
for i, c_name in zip(change_index, name_to_change):
GDP.at[i, 'Country Name'] = c_name # 更改國家名稱
GDP = GDP.set_index('Country Name')
return GDP
def get_ScimEn():
"""
清理 ScimEn 資料
"""
ScimEn = pd.read_excel('https://storage.googleapis.com/um_ds_intro/scimagojr-3.xlsx')
ScimEn = ScimEn.iloc[:15, :].set_index('Country')
return ScimEn
def answer_one():
energy = get_energy()
GDP = get_GDP()
ScimEn = get_ScimEn()
df = ScimEn.join(energy)
df = df.join(GDP)
return df
def answer_two():
energy = get_energy()
GDP = get_GDP()
ScimEn = get_ScimEn()
df_outer = ScimEn.join(energy, how='outer')
df_outer = df_outer.join(GDP, how='outer')
df_inner = answer_one()
ans = df_outer.shape[0]
return ans
answer_two()
# 305
如果您喜歡這篇文章，請多按下方的「拍手」圖像幾次、分享到社群網站、成為我們的贊助者以及訂閱 DataInPoint 的新文章！
| 7
|
ee45562e3de2
|
2018-03-26
|
2018-03-26 10:09:38
|
2018-03-26
|
2018-03-26 14:59:35
| 7
| false
|
zh-Hant
|
2018-04-15
|
2018-04-15 02:04:56
| 18
|
17f62ee9ac3a
| 3.45
| 8
| 0
| 0
|
University of Michigan Introduction to Data Science in Python 的第三週作業：More Pandas(Part 1)
| 5
|
再生能源指標資料探索（上）
University of Michigan Introduction to Data Science in Python 的第三週作業：More Pandas(Part 1)
Photo by Darren Coleshill on Unsplash
探索國家再生能源指標資料，首先得依照題目的提示合併三個資料源：
Energy Indicators.xls
world_bank.csv
scimagojr-3.xlsx
三個資料的格式是常見的 CSV 檔案與 Excel 試算表，假如直接透過 pandas 套件載入 Python 並不是難事，但是仔細閱讀題目之後就知道除了載入以外，還有許多資料清理的工作必須處理，這也是課程希望我們練習的技能。
喜歡 DataInPoint 的文章嗎？成為我們的贊助者吧！
DataInPoint is creating Data Science Tutorials | Patreon
Become a patron of DataInPoint today: Read posts by DataInPoint and get access to exclusive content and experiences on…www.patreon.com
作業來源
University of Michigan 在 Coursera 開設的 Python 資料科學應用是一門輕量著重資料科學套件的專項課程，其中共有五門課程，適合已經有 Python 程式設計基礎，但想在短時間內暸解 Python 在資料科學領域應用的學習者。Introduction to Data Science in Python 是這個專項課程中的第一門課，內容包含 Python 程式設計、pandas 簡介、pandas 進階與基礎統計分析。
跟 DataInPoint 一起完成課程中的 Programming Assignments 吧，我們今天要寫第三週要繳交作業的第一部分！
這個練習所有的程式、資料與輸出結果，都可以在這個 Google Colaboratory文件中檢視，喜歡英文介面的讀者，可點選幫助，再點選查看英文版。
um_intro2_ds_py_assignment3_part1.ipynb
Colaboratory notebookdrive.google.com
University of Michigan Applied Data Science with Python Specialization: Introduction to Data Science in Python
概述
This assignment requires more individual learning then the last one did — you are encouraged to check out the pandas documentation to find functions or methods you might not have used yet, or ask questions on Stack Overflow and tag them as pandas and python related. And of course, the discussion forums are open for interaction with your peers and the course staff.
這份作業需要對 Python 或者 pandas 有更多的熟練才能應對，如果面對到不會的問題，除了查找 pandas 的官方文件，還可以查找 Stack Overflow。
第一題
來細看每個資料的清理需求：
Energy Indicators.xls 要對變數重新命名、轉換變數單位、將遺漏值修正為 np.NaN 、更改國家名稱
world_bank.csv 要略過首列、更改國家名稱並使用 2006–2015 這十年的 GDP 資料即可
scimagojr-3.xlsx 只要使用 Rank 1 到 Rank 15 這 15 個國家
Energy Indicators.xls
world_bank.csv
scimagojr-3.xlsx
首先來處理 Energy Indicators.xls：
載入資料時將試算表上方與下方不用的資料略過，並記得把文字 ... 指定為遺漏值
更改變數名稱
清除國家名稱中結尾的小括號
清除國家名稱中結尾的數字
重新命名特定的國家名稱
接著處理 world_bank.csv：
載入資料時略過上方四列
留下國家名稱與 2006 至 2015 年的 GDP 資料
重新命名特定的國家名稱
最後是處理 ScimEn.xlsx：
只留下前 15 個觀測值
當我們將個別資料清理乾淨之後，最後是利用國家名稱聯結三個資料框，輸出一個外觀為 15 x 20 的資料框：
第二題
The previous question joined three datasets then reduced this to just the top 15 entries. When you joined the datasets, but before you reduced this to the top 15 items, how many entries did you lose?
This function should return a single number.
在前一題我們聯結三個資料框後得到交集的 15 個國家，想要知道在聯結的過程中失去多少個觀測值，題目用一個 Venn Diagram 提示要用三個資料框觀測值的聯集減去交集：
題目提示
於是可以再做一次聯結，但是將聯結方法指定為 how = 'outer' ，這樣就可以得到三個資料框聯集的觀測值數，進而與問題一的觀測值數相減，即可得到答案。
喜歡 DataInPoint 的文章嗎？成為我們的贊助者吧！
DataInPoint is creating Data Science Tutorials | Patreon
Become a patron of DataInPoint today: Read posts by DataInPoint and get access to exclusive content and experiences on…www.patreon.com
延伸閱讀
如果您想多暸解 Python 與 pandas，推薦參考 DataCamp 的 pandas 相關課程：
pandas Foundations | DataCamp
Learn how to use the industry-standard pandas library to import, build, and manipulate DataFrames.www.datacamp.com
Manipulating DataFrames with pandas | DataCamp
You will learn how to tidy, rearrange, and restructure your data using versatile pandas DataFrames.www.datacamp.com
Cleaning Data in Python | DataCamp
Data scientists spend 80% of their time cleaning and manipulating data. This course will equip you with all the skills…www.datacamp.com
Introduction to Data Science in Python | Coursera
About this course: This course will introduce the learner to the basics of the python programming environment…www.coursera.org
|
再生能源指標資料探索（上）
| 35
|
university-of-michigan-introduction-to-data-science-in-python-的第三週作業-more-pandas-part-1-17f62ee9ac3a
|
2018-04-15
|
2018-04-15 02:04:57
|
https://medium.com/s/story/university-of-michigan-introduction-to-data-science-in-python-的第三週作業-more-pandas-part-1-17f62ee9ac3a
| false
| 636
|
DataInPoint 是一個超棒的資料科學專欄，主題涵蓋資料、程式、機器學習與高效能運算。
| null |
412106045903591
| null |
DataInPoint
|
tonykuoyj@gmail.com
|
datainpoint
|
PROGRAMMING,VISUALIZATION,MACHINE LEARNING,DATA
|
DataInPoint
|
Data Science
|
data-science
|
Data Science
| 33,617
|
Yao-Jen Kuo
|
Could that data be any tidier? It is always nice to meet a data enthusiast who is also a marathon runner and a ping-pong lover.
|
ec74f573f855
|
tonykuoyj
| 967
| 182
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-13
|
2018-05-13 16:19:01
|
2018-05-13
|
2018-05-13 16:24:29
| 2
| false
|
en
|
2018-05-13
|
2018-05-13 16:24:29
| 2
|
17f725448f41
| 3.002201
| 2
| 0
| 0
|
Before I get to the crux of the matter, discussing Artificial Intelligence, let me start with an anecdote similar to the topic at hand, and…
| 5
|
A.I., The Black Swan, and Asking the Relevant Questions
Before I get to the crux of the matter, discussing Artificial Intelligence, let me start with an anecdote similar to the topic at hand, and which took place a few years back, in 2007, to be precise. During that year, no one with proper judgement could have foreseen the looming financial catastrophe. Crisis befell the markets, while the incredulous economists could not construct a proper etiology of the subprime loans crash.
All theoretical attempts to examine and explain concretely what went wrong came to no avail. It was in the aftermath of the crisis that the name of Nassim Nicholas Taleb started gaining momentum and his book The Black Swan (2007) released to an impactful global reception, polarizing economists and financial analysts alike.
As it turns out, way before the market crash, Taleb — who has been teaching risk analysis at several US academic institutions and was a floor trader for many years (he had Skin in the Game, as he would describe it in his 2018 book) — had alerted the ecstatic economists about an improbable, yet possible mishap. Taleb was quickly dismissed as a lunatic theoretician preaching an unintelligible theory that bore no justifiable ground.
When the market’s tendency was bullish (on the rise), people like Nassim Taleb were alerting of the cataclysmic consequences a possible crash would entail for the people, and for financial institutions. Hence the title of his book is The Black Swan, (in reference to the discovery of Australian black swans, when the preconception was that swans were all white), which alludes to the possibility that an odd and very unlikely event might take place (economic crisis, economic bubble burst) when all the predictions and forecasts are contrariwise.
I will leave the philosophical underpinning of Taleb’s argument for a future post. For now, I will use the Black Swan argument to draw an analogy between the financial market and the tech industry.
The hype surrounding the leading tech firms has been somewhat successful at foreshadowing ethical concerns, which have been relegated to a secondary level, in favor of features that appealed to the consumerist spirit of the public.
The positive atmosphere in Silicon Valley, and elsewhere, indirectly silenced and vetoed genuine attempts at dissecting and analyzing the ensuing ethical problems. This blinding belief in the near-infallibility of A.I. — which is still a burgeoning field — resulted in a series of unexpected accidents (like that of Uber, 2018), and scandals (like that of Facebook/Cambridge Analytica, 2018).
Failing to so much as entertain the possibility of an existing glitch in the system prompted ignominious situations that could have been avoided had these companies contemplated the likelihood of a black swan event.
The drive to generate profit as fast as possible, in order to justify the excessive amount of money invested in fields of research like A.I, deep learning, and neural networks, overburdened many stakeholders with risk that, while not easily predictable, could have been accounted for by asking the relevant questions.
The relevant questions I have in mind are the sort of questions that are more often than not asked within the ivory towers philosophers inhabit. As such, instead of engaging with the potential setbacks of any concrete venture, they would rather limit themselves to arguments the more abstracted from reality the better.
These same questions are, on the other hand, pondered by fictional artistic, literary, and visual productions; for example, in movies such as Her (2013) and Ex Machina (2014), and T.V. series such as Black Mirror (2011). However, the themes that these productions deal with are often ignored by the concerned stakeholders.
The result is a virtually unbridgeable hiatus between the tech industry and other sectors, with each disregarding the other sector’s two cents.
However, if there’s a lesson that we could learn from the financial market and from Nassim Taleb, it would be this: perhaps it is not a waste of time to attempt to ask the relevant questions, questions about the improbable but possible events that might strike the A.I. research of the tech industry, in order to avoid the proliferation of black swan events.
|
A.I., The Black Swan, and Asking the Relevant Questions
| 2
|
a-i-the-black-swan-and-asking-the-relevant-questions-17f725448f41
|
2018-05-15
|
2018-05-15 03:43:15
|
https://medium.com/s/story/a-i-the-black-swan-and-asking-the-relevant-questions-17f725448f41
| false
| 694
| null | null | null | null | null | null | null | null | null |
Tech
|
tech
|
Tech
| 142,368
|
Mahmoud Rasmi
|
@decafquest | Philosophy professor
|
c68b559d5471
|
mahmoud.rasmi
| 19
| 8
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-09-04
|
2018-09-04 11:40:57
|
2018-09-04
|
2018-09-04 11:59:53
| 1
| false
|
en
|
2018-09-04
|
2018-09-04 11:59:53
| 0
|
17f80202622b
| 1.630189
| 0
| 0
| 0
|
What’s it like working as a Specialist in Fraud Analyst at NewDay? Find out from Rebecca, in our latest edition of ‘Take 5 with…’!
| 1
|
Take 5 with…Rebecca Man
What’s it like working as a Specialist in Fraud Analyst at NewDay? Find out from Rebecca, in our latest edition of ‘Take 5 with…’!
How long have you worked for NewDay?
1 Year and 9 Months in total, the first 14 months was as a Graduate.
How would you describe what you do at NewDay in one sentence?
I use a mixture of internal and external intelligence as well as a range of analytical techniques to try to protect our customers from fraud.
Tell us about your proudest achievement or your most memorable day at NewDay.
My proudest achievement at NewDay is when I taught myself a new language which allowed me to build a network visualisation around criminal activity affecting our business.
What does the NewDay Manifesto mean to you?
From a fraud perspective it’s about protecting our customers from fraud as much as possible whilst inconveniencing their everyday lives as little as possible, this is always a major consideration when creating any new strategy.
The three best things about working at NewDay are…
1. The generous employee benefits
2. The size of the business allows you to have a big influence in the area you work in
3. The people you work with are very welcoming and friendly and always willing to help
What’s the most interesting / exciting meeting in your diary this week?
Python training with the Credit Risk Data Science team. This is to learn a new language which we will hopefully be able to utilise for future analysis.
Tell us about the social side of NewDay — what’s your favourite event been?
We have beer and pizza on the last Friday of every month. This provides us all the opportunity to socialise with our colleagues.
Tell us an interesting fact about yourself.
I love animals, I have a Chihuahua that sings to any songs by Andrea Bocelli and recently hatched baby chickens in my oven!
If you could do another job at NewDay for one day, what would you choose and why?
I would work in Data Science because I like building models and manipulating data to tell a story. I also feel like I could develop a lot of the skills I had when I was at university.
|
Take 5 with…Rebecca Man
| 0
|
take-5-with-rebecca-man-17f80202622b
|
2018-09-04
|
2018-09-04 11:59:53
|
https://medium.com/s/story/take-5-with-rebecca-man-17f80202622b
| false
| 379
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
NewDay
| null |
2b972b05d502
|
newday
| 7
| 0
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
32881626c9c9
|
2018-09-17
|
2018-09-17 14:07:01
|
2018-09-17
|
2018-09-17 14:11:39
| 1
| false
|
en
|
2018-09-17
|
2018-09-17 14:11:39
| 2
|
17f8249a2442
| 3.264151
| 2
| 0
| 0
|
“Should you make your own chatbot for eCommerce?”
| 3
|
How Chatbot Development Can Rethink Customer Service in eCommerce?
“Should you make your own chatbot for eCommerce?”
Yes, we say!
Chatbots are touted as one of the trending eCommerce trends for 2018 and beyond. It is obvious why retailers are pouring time and money into building a chatbot that will reimagine the quality of their customer service.
The 3Cinteractive Chatbot Progress Report has found that about “77% of consumers surveyed said positive interactions with chatbots make them purchase more often.”
Whether it is for increasing revenue or to render a better customer service or even to automate internal operations, chatbots for eCommerce have plenty of applications.
Why Retailers Should build an ecommerce chatbot?
They personalize the customer experience
From first visit to checkout, an eCommerce chatbot can be a virtual personal assistant with a human logic and machine memory. To quantify the possibilities, an intelligent chatbot will be able to drive conversions as high as 30% through chatbot powered visitor engagement (Source: VentureHarbour).
Machine learning and NLP capabilities would help build the best chatbot for ecommerce that can enable to chatbot to analyze the customer’s past preferences and predict their next purchase intent and also offer tailor-made product recommendations. Advanced data crunching technologies like Big Data would further empower chatbots to deliver a targeted customer experience that can heighten brand engagement.
Chatbots will help overcome one of the biggest challenges of traditional marketing tactics like email marketing. These tactics often greeted the customer with out-of-context information or overly promotional content. A chatbot can supply the customer with content that is more relevant to the customer’s preferences. In fact, studies have proven that customers are ready to spend as much as $314 on products suggested by chatbots.
Multi-visitor Engagement
Unlike human personnel, chatbots can assist multiple visitors simultaneously. They can provide canned responses to recurring queries or even provide dynamic responses tailor-made to individual customers based on their transaction history.
Chatbot development services This delivers a two-fold advantage for eCommerce retailers. Consistent interaction with customers will help keep bounce rate to a bare minimum. Secondly, visitors who have their queries attended to in a personalized manner have a higher probability of converting into paying customers. Global Think Tank Gartner has predicted that by 2019, 20 Percent of User Interactions With Smartphones Will Take Place via VPAs (Virtual Personal Assistants).
Data mining for customer sentiment analysis
Is the customer’s query filled with angst or is it a common query that most users tend to ask? An intelligent chatbot would be able to do an accurate sentiment analysis by deducing the text pattern to rate the query as positive, negative or neutral.
Data mining customer interaction and the text inputs exchanged between the customer and the chatbot would help in predicting customer sentiment. It would aid in rendering a predictive customer service that will cement brand loyalty.
Quick resolutions for customer queries
Chatbots can be trained Entity Extraction from text. Entity extraction from text means identifying information like person’s name, location, store name, device name, unique identification number, etc. This would help the chatbot answer the customer queries proactively without asking more questions.
For instance, a customer who cites the unique identification number of his device can be served by the chatbot in his native language. More information like whether past service history of the product, warranty period, nearest authorized service centre can be provided by the chatbot without pestering the customer for too much of the information.
Anomaly Detection
Retailers are plagued with the risk of fraudulent transactions schemed by miscreants who leverage loopholes in the system. For example, a flawed return policy which allows the customer to return a purchased item even after its warranty period. Or a faulty delivery process that leads to more returns or failed delivery attempts.
Using anomaly detection, a chatbot would be able to single out such instances for further investigation. This would enable the retailer to perfect the selling process that is free of revenue leaks. The benefit that chatbots provide above manual analysis is that, unlike a human data scientist, they have the capability to predict such scenarios before they happen.
Closing In
With Machine Learning and Artificial Learning maturing as affordable technologies, the cost of building a chatbot has also reduced considerably. Retailers can bank on chatbots to take their business volumes to new heights. Chatbots are the most easily relatable manifestation in which retailers can use these technologies.
From cementing solid customer relationships to taking the heavy burden of customer service away from personnel, chabots can deliver high on several areas where constant challenges plague retailers. Partnering with a reliable chatbots development company who can provide the know-how and the technical assistance is a must-have.
From gaining insights to delivering instant value, a chabot can be an everlasting source of value for eCommerce retailers.
Source Link
|
How Chatbot Development Can Rethink Customer Service in eCommerce?
| 28
|
how-chatbot-development-can-rethink-customer-service-in-ecommerce-17f8249a2442
|
2018-09-21
|
2018-09-21 03:16:17
|
https://medium.com/s/story/how-chatbot-development-can-rethink-customer-service-in-ecommerce-17f8249a2442
| false
| 812
|
Data Driven Investor (DDI) brings you various news and op-ed pieces in the areas of technologies, finance, and society. We are dedicated to relentlessly covering tech topics, their anomalies and controversies, and reviewing all things fascinating and worth knowing.
| null |
datadriveninvestor
| null |
Data Driven Investor
|
info@datadriveninvestor.com
|
datadriveninvestor
|
CRYPTOCURRENCY,ARTIFICIAL INTELLIGENCE,BLOCKCHAIN,FINANCE AND BANKING,TECHNOLOGY
|
dd_invest
|
Chatbots
|
chatbots
|
Chatbots
| 15,820
|
Nikki
|
The eComm Geek
|
fe13d4130e4e
|
nikki_ecommgeek
| 4
| 4
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-07-03
|
2018-07-03 09:25:20
|
2018-07-03
|
2018-07-03 09:44:29
| 1
| false
|
en
|
2018-07-03
|
2018-07-03 09:44:29
| 2
|
17f866758183
| 2.890566
| 0
| 0
| 0
|
Every CEO is aware of the value of data, but the truth is, not many have been making full use of what they have for actionable customer…
| 5
|
Have You Done All You Could To Prepare For A Data-Intensive, AI-Enhanced Future?
Every CEO is aware of the value of data, but the truth is, not many have been making full use of what they have for actionable customer insights that could help them identify new opportunities. The root of the problem is usually that businesses have simply been complacent, not moving — or moving fast enough — to catch up with the pace with which data has been evolving. Data, both in terms of volume and the rate it expands, is growing — and it’s not showing any signs of slowing down. The good news is there’s always time to catch up.
In the very near future, we can expect to see a massive data boom, thanks to Internet of Things (IoT). Domestic application of IoT devices will become ubiquitous, and the datasets required to sustain them will only get bigger. However, at present, the challenge of collecting the data from thousands or millions of data sources and moving that data to the compute elements is proving to be tremendously difficult given the existing network capabilities.
And this is only the beginning — over the next decade, the business environment will almost be unrecognizable. Everything that can be connected will bring greater efficiency to business processes, but the data involved in sustaining this will be enormous and it will feed a whole host of new artificial intelligence (AI) based applications and services.
For a long time, AI felt a nebulous concept. Now we’re seeing examples of machine learning in many aspects of our daily lives, and we will only become more accustomed to these technologies as chatbots and AI assistants become mainstream consumer technologies.
All the top tech players have already been exploring AI for long. Google, for instance, has invested considerable amounts of money and hours into developing its AI assistant, which is now powering the devices on its Android platform. In time, all the pools of data that are stored in these companies’ storage will eventually be turned into the fuel for machine intelligence.
But AI brings with it a dichotomy for those looking to invest. On one hand, the potential of AI to address macro challenges in critical areas such as healthcare and genomics is immense. Sooner or later AI will take over mundane tasks and provide us with creative insights, benefiting every sector from banking to manufacturing.
On the other hand, it requires seriously strong infrastructures to underpin the innovation. Take the automotive industry– making autonomous cars a mainstream part of our lives will be a data-heavy task. Cars will need to be fed vast amounts of data to operate efficiently and safely. This industry will become a pool of data that needs to be managed.
Regardless of industry, the useful processing of data, either contemporary or historical, can add value to the bottom line as useful information is extracted and used to compete more effectively, innovate more rapidly, improve customer engagement and so on.
When it comes to processing data, a simple, scalable, high-performance data center environment is already one of the most powerful strategic assets any firm can possess, and that’s not going to change in the future. But as we’re moving to the cloud-era it brings with it a new set of demands for storage, with a requirement for a high degree of performance at scale.
Businesses need to act fast, so cloud-era flash storage can handle data at speed, as high latency and data bottlenecks are going to have a negative impact across the business. A cloud-era solution therefore focuses on exponentially expanding the value of data through faster and more reliable access. This enables businesses to build a new class of applications to extract new insights from data and leverage technology developments such as AI.
As we move into our AI defined data rich world, it’s critical that organizations start considering how they will be able to support their AI and ML systems to draw out those insights. After all it’s only when you can extract the information hidden in the data that the data itself becomes valuable.
Read More:
From Big Data To Smart Analytics Solution
‘Huat’ Does Interconnection Mean During Chinese New Year?
|
Have You Done All You Could To Prepare For A Data-Intensive, AI-Enhanced Future?
| 0
|
have-you-done-all-you-could-to-prepare-for-a-data-intensive-ai-enhanced-future-17f866758183
|
2018-07-03
|
2018-07-03 09:44:29
|
https://medium.com/s/story/have-you-done-all-you-could-to-prepare-for-a-data-intensive-ai-enhanced-future-17f866758183
| false
| 713
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Sahana Karnam
| null |
4d14a046af36
|
sahana_63956
| 43
| 174
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-10-17
|
2017-10-17 22:50:25
|
2017-10-19
|
2017-10-19 15:34:44
| 4
| false
|
en
|
2017-10-20
|
2017-10-20 10:09:52
| 0
|
17f9c1076711
| 5.318868
| 3
| 1
| 0
|
(6 min read) Can a human experience be generated algorithmically?
| 5
|
Mary and the Simulation hypothesis
In the last few decades one of the most interesting scientific and philosophical concepts being discussed by the likes of Stephen Hawking, Jaron Lanier and Elon Musk is the simulation hypothesis. This is the idea that one’s experience of life only exists as a very complex set of data in a larger, more complex simulation that resembles the universe we perceive.
Nick Bostrom proposed in an article in 2003 entitled ‘are you living in a simulation?’ that our reality is an ‘ancestor simulation’ created by a post-human civilisation who, with vast computing power might choose to simulate the lives of their ancestors.
A later paper called “Constraints on the universe as a numerical simulation” concludes that Bostrom’s suggestion is indeed possible as the universe is proven to have a finite resolution or ‘underlying lattice.’ This led digital physicists to hypothesise that the universe, as we perceive it, is not made up of matter and energy but rather of ‘quantised bits of information’ which are perceived as energy and matter.
This argument relies on a few things, such as the willingness and capability of future (or indeed past) civilizations to conduct such an experiment. The more interesting question however, surrounds the concept of consciousness and whether an immersive and convincing human experience could ever be created artificially.
An increasingly popular belief held by dogmatic scientists is that the human experience is nothing more than a complex product of biological machine learning. Essentially, this means that the all human desires, thoughts, hopes and dreams do not come from an immeasurable and intangible soul, but rather from complex, yet predictable (and perhaps more importantly, computable) biochemical reactions in the brain. In the words of Yuval Noah Harari, “There is no true self, there is just a complicated network of biochemical connections, without a core, there is no authentic voice that lives inside you.” Put simply, it is increasingly evident that humans do not have souls which in turn suggests that it might be easier than we think to simulate our experiences.
Plato’s “cave” addresses the idea that what we perceive to be reality may not be absolute.
From a technological perspective this has significant implications. It is starting to become apparent that the exponentially rapid advancement of computer processing power (as defined by Moore’s law) coupled with artificial machine learning programmes means that artificial intelligence that can compete with a human is right around the corner. This artificial intelligence is governed by a set of incredibly complex algorithms which will provide responses to stimuli based on an enormous amount of collected data. Most importantly though, these algorithms can adapt and learn independently of human supervision. This happens through a process called ‘algorithmic machine learning’ and is wonderfully demonstrated by the programme “Mar/io” (which can be found on Youtube) whereby a computer taught itself to play the popular Nintendo game ‘Mario World.’
A visualisation of the neural network that allowed this algorithm to successfully complete the level.
The method of all machine learning algorithms is one of trial and error repeated over many simulations incorporating slight variation. This process very closely matches the way that humans evolved biologically. That is to say, random mutations being added into the ‘system’ such that if the mutation proves beneficial, the individuals with the given mutation will be more successful and the previous ‘generation’ will be rendered obsolete. Does this suggest that given enough time, enough data and enough processing power that an accurate human experience as we know it could be replicated wholly artificially and inorganically?
The counter argument is that there is more to the human experience than the algorithmic nature of our neural networking. Each human has unique experiences which give rise to a unique emotional profile and personality. The way that a human experiences an event is special because it is fundamentally subjective. This differs from an algorithmic experience of an event which is objective and measured against a fixed criteria. This point is illustrated beautifully by Frank Jackson’s famous thought experiment of ‘Mary and the black and white room’ which goes like this:
“Mary is a brilliant scientist who is, for whatever reason, forced to investigate the world from a black and white room via a black and white television monitor. She specializes in the neurophysiology of vision and Learns everything there is to know about colour. She discovers, for example, the wavelengths of every colour in the visible light spectrum, their emotional associations and even that particular objects are certain colours: tomatoes are red, grass is green etc. She learns that the sea is blue and associated with tranquility, she understands that red is visceral and associated with anger and passion. But Mary has never actually seen these colours.”
What will happen when Mary is released from her black and white room or is given a colour television monitor? Will she learn anything or not?
A scene in the film “Ex Machina” explores this thought experiment from the perspective of an artificial intelligence.
This experiment tries to establish if there are non-physical properties and attainable knowledge that can be discovered only through conscious experience. It attempts to refute the theory that all knowledge is objective, physical knowledge. In this experiment, when Mary is locked in the room, she represents an algorithmic experience of the world, one that is based on knowledge and hard data rather than subjective experiences. Once Mary leaves the room (assuming that she DOES learn something new), she represents a human mind that is defined by its own conscious experience of events, a mind that can attain knowledge that cannot be physical or computable.
So, while the simulation hypothesis is a fun conversation, especially given the rate of technological advancement, it relies too heavily on speculation and philosophy to give rise to any serious scientific discussions. As Sir Isaac Newton said, “Anything that cannot be settled by experiment is not worthy of debate”.
As a fun final exercise, consider a hypothetical, vastly powerful computer was tasked with generating your life as you’ve lived it so far, as accurately as possible. Everything from your emotions and feelings, to your experiences and interactions with the world around you. Consider that your current digital footprint is used to help produce this fake reconstruction of your life.
How well do you think it would do?
How accurately do you think it could guess your emotions and the voice inside your head based on every message, tweet, email or status you’ve ever sent or received?
Could this algorithm build authentic, convincing memories from your phone’s pictures, as well as those from snapchat, Instagram and Facebook?
Would it know your music taste and sense of humour based on which Spotify playlists you listen to and which Youtube & Netflix videos you watch?
With how much precision could this algorithm pinpoint your lifestyle habits by observing your online bank statement and Amazon purchase history?
Could it even expose your deepest secrets, insecurities and wanton desires by looking at your private browsing history?
Your smartphone already knows your location to within a few squared metres at any given time and graphics engines can already achieve a resolution and frame rate better than real life. This, coupled with up to date satellite imagery suggests that the audio-visual aspects of this simulation might be strikingly accurate. But, because of the complexities of what we perceive to be consciousness and a subjective understanding of the world around us, it is too difficult to say if this simulation could ever achieve a level of consistency and immersion that would lead the simulation to be completely convincing.
|
Mary and the Simulation hypothesis
| 36
|
mary-and-the-simulation-hypothesis-17f9c1076711
|
2018-06-05
|
2018-06-05 19:39:57
|
https://medium.com/s/story/mary-and-the-simulation-hypothesis-17f9c1076711
| false
| 1,224
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Arthur Morgan
|
Law student at The University of Sussex
|
fb4cc84345b1
|
ArthurMorgan
| 27
| 20
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-07
|
2018-05-07 10:31:23
|
2018-05-07
|
2018-05-07 10:36:49
| 1
| false
|
en
|
2018-05-07
|
2018-05-07 10:36:49
| 0
|
17fb4eaefa43
| 0.441509
| 1
| 0
| 0
|
Revolutionary approach leveraging benefits of blockchain technology and providing direct connection between retailers, service providers…
| 5
|
Worldwide first decentralized services recommendation platform
Revolutionary approach leveraging benefits of blockchain technology and providing direct connection between retailers, service providers and users. ServAdvisor is on the verge of revolutionizing the way we consume services by bringing customer experience to a new level due to a combination of Augmented Reality (AR), Artificial Intelligence (AI) and customer incentivized review platform.
Please follow us on ServAdvisor.co
|
Worldwide first decentralized services recommendation platform
| 50
|
worldwide-first-decentralized-services-recommendation-platform-17fb4eaefa43
|
2018-06-05
|
2018-06-05 11:19:12
|
https://medium.com/s/story/worldwide-first-decentralized-services-recommendation-platform-17fb4eaefa43
| false
| 64
| null | null | null | null | null | null | null | null | null |
Augmented Reality
|
augmented-reality
|
Augmented Reality
| 13,305
|
ServAdvisor
| null |
64017f48c363
|
ServAdvisor
| 32
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
7f60cf5620c9
|
2017-10-03
|
2017-10-03 15:01:45
|
2017-10-03
|
2017-10-03 15:02:24
| 2
| false
|
en
|
2017-10-03
|
2017-10-03 15:02:24
| 5
|
17fb5daa7f9f
| 1.651258
| 0
| 0
| 0
|
MIT spin-off Zylotech earned a place as a featured startup at the Startup Autobahn demo-day in Singapore on September 7, 2017. The award…
| 1
|
A.I. Powered Customer Data & Insights Company Zylotech Selected as Featured Startup at Startup Autobahn Global Innovation Group
MIT spin-off Zylotech earned a place as a featured startup at the Startup Autobahn demo-day in Singapore on September 7, 2017. The award winning A.I. platform which combines customer data management with a deep-learning decisioning engine was the subject of a lot of talk as global brands in attendance learned how far A.I. has come in the customer marketing space.
Most brands in attendance face a similar issue, too much data, in too many places and not enough time and data scientists to leverage more than 10–15% of available customer data.
Zylotech’s solution leverages access to named data in both internal and external environments to enrich and unify ALL of a global organization’s customer data sources. Once the A.I. generated data engine is running with minimal human intervention, the data is then fed into a machine learning based decisioning engine to empower marketing opps and customer loyalty to work with real-time, high probability predictive models to increase monetization, retention and LTV for a company’s customer base.
About Startup Autobahn
We are a neutral innovation platform moderating an in-depth and curated collaboration between our
partners from the industry and young tech companies. Our network of investment firms and mentors further helps the growth of tech companies. Our home is Stuttgart — the birthplace of the automobile, one of the leading IP regions of the world, and a hub of business, engineering, and groundbreaking inventions.
Our namesake is the high-speed German highway — the only one in the world without a speed limit. And just like the Autobahn, STARTUP AUTOBAHN is about accelerating joint projects between young tech companies and our corporate partners.
Click the link below to head over to our site and set up a quick call if you’d like a first hand look at what had the Autobahn buzzing.
Topics: Customer Data Management Press Release AI event
Originally published at www.zylotech.com.
|
A.I. Powered Customer Data & Insights Company Zylotech Selected as Featured Startup at Startup…
| 0
|
a-i-powered-customer-data-insights-company-zylotech-selected-as-featured-startup-at-startup-17fb5daa7f9f
|
2017-10-04
|
2017-10-04 16:35:44
|
https://medium.com/s/story/a-i-powered-customer-data-insights-company-zylotech-selected-as-featured-startup-at-startup-17fb5daa7f9f
| false
| 336
|
Sharing concepts, ideas, and codes.
|
towardsdatascience.com
|
towardsdatascience
| null |
Towards Data Science
| null |
towards-data-science
|
DATA SCIENCE,MACHINE LEARNING,ARTIFICIAL INTELLIGENCE,BIG DATA,ANALYTICS
|
TDataScience
|
Data Science
|
data-science
|
Data Science
| 33,617
|
Zylotech
|
AI-driven customer intelligence platform that enables clients to cross-sell / up-sell to their customers.
|
c0a579591694
|
zylotech
| 34
| 22
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
63505f63fad2
|
2018-06-26
|
2018-06-26 14:35:23
|
2018-07-20
|
2018-07-20 12:29:59
| 4
| true
|
en
|
2018-07-20
|
2018-07-20 12:29:59
| 28
|
17fcd241b737
| 11.530189
| 6
| 0
| 0
|
AI algorithms can convincingly imitate human appearance and behavior — and that comes with profound implications, both positive and not.
| 5
|
When AI Blurs the Line Between Reality and Fiction
AI algorithms can convincingly imitate human appearance and behavior — and that comes with profound implications, both positive and not.
By Ben Dickson
Somewhere in the dark recesses of YouTube is a video that shows an excerpt from the movie The Fellowship of the Ring — but it’s not quite the movie you remember, since Nicolas Cage stars as Frodo, Aragorn, Legolas, Gimli, and Gollum, all at the same time. Other videos show Cage in Terminator 2 as T2000, Star Trek as Captain Picard, and Superman as, well, Lois Lane.
Of course, Nic Cage never appeared in any of those movies. They’re “deepfakes” produced with FakeApp, an application that uses artificial intelligence algorithms to swap faces in videos. Some of the deepfakes look quite convincing, while others have artifacts that betray their true nature. But overall, they show how powerful AI algorithms have become in imitating human appearance and behavior.
FakeApp is just one of several new AI-powered synthesizing tools. Other applications mimic human voices, handwriting, and conversation styles. And part of what makes them significant is that using them doesn’t require specialized hardware or skilled experts.
The impact of these applications is profound: They will create unprecedented opportunities for creativity, productivity, and communications.
But the same tool could also open a Pandora’s box of fraud, forgery, and propaganda. Since it made an appearance on Reddit in January, FakeApp has been downloaded more than 100,000 times and precipitated a storm of fake pornographic videos featuring celebrities and politicians (including Cage again). Reddit recently banned the application and its related communities from its platform.
“Ten years ago, if you wanted to fake something, you could, but you had to go to a VFX studio or people who could do computer graphics and possibly spend millions of dollars,” says Dr. Tom Haines, lecturer in machine learning at University of Bath. “However, you couldn’t keep it a secret, because you’d have to involve many people in the process.”
That’s no longer the case, courtesy of a new generation of AI tools.
The Imitation Game
FakeApp and similar applications are powered by deep learning, the branch of AI at the heart of an explosion of AI innovations since 2012. Deep-learning algorithms rely on neural networks, a software construction roughly fashioned after the human brain. Neural networks analyze and compare large sets of data samples to find patterns and correlations that humans would normally miss. This process is called “training,” and its outcome is a model that can perform various tasks.
In earlier days, deep-learning models were used mostly to perform classification tasks — labeling objects in photos, for example, and performing voice and face recognition. Recently, scientists have used deep learning to perform more complicated tasks, such as playing board games, diagnosing patients, and creating music and works of art.
To tune FakeApp to perform a face swap, the user must train it with several hundred pictures of the source and target faces. The program runs deep-learning algorithms to find patterns and similarities between the two faces. The model then becomes ready to make the swap.
The process isn’t simple, but you don’t have to be a graphics expert or machine-learning engineer to use FakeApp. Neither does it require expensive and specialized hardware. A deepfakes tutorial website recommends a computer with 8GB or more of RAM and an Nvidia GTX 1060 or better graphics card, a pretty modest configuration.
“Once you move to a world where someone in a room can fake something, then they can use it for questionable purposes,” Haines says. “And because it’s one person on their own, keeping it secret is very easy.”
In 2016, Haines, who was then a postdoctoral researcher at University of College London, coauthored a paper and an application that showed how AI could learn to imitate a person’s handwriting. Called “My Text in Your Handwriting,” the application used deep-learning algorithms to analyze and discern the style and flow of the author’s handwriting and other factors such as spacing and irregularities.
The application could then take any text and reproduce it with the target author’s handwriting. The developers even added a measure of randomness to avoid the uncanny valley effect — the strange feeling that we get when we see something that is almost but not quite human. As a proof of concept, Haines and the other UCL researchers used the technology to replicate the handwriting of historical figures such as Abraham Lincoln, Frida Kahlo, and Arthur Conan Doyle.
The same technique can be applied to any other handwriting, which raised concerns about the technology’s possible use for forgery and fraud. A forensics expert would still be able to detect that the script was produced by My Text in Your Handwriting, but it’s likely to fool untrained people, which Haines admitted in an interview with Digital Trends at the time.
Lyrebird, a Montreal-based startup, used deep learning to develop an application that synthesizes human voice. Lyrebird requires a one-minute recording to start imitating the voice of a person, though it needs much more before it starts to sound convincing.
In its public demo, the startup posted fake recordings of the voices of Donald Trump, Barack Obama, and Hillary Clinton. The samples are crude, and it’s obvious that they’re synthetic. But as the technology improves, making the distinction will become harder. And anyone can register with Lyrebird and start creating fake recordings; the process is even easier than FakeApp’s, and the computations are performed in the cloud, putting less strain on the user’s hardware.
The fact that this technology can be used for questionable purposes is not lost on developers. At one point, an ethics statement on Lyrebird’s website stated: “Voice recordings are currently considered as strong pieces of evidence in our societies and in particular in jurisdictions of many countries. Our technology questions the validity of such evidence as it allows [people] to easily manipulate audio recordings. This could potentially have dangerous consequences such as misleading diplomats, fraud, and more generally any other problem caused by stealing the identity of someone else.”
Nvidia presented another aspect of AI’s imitation capabilities: Last year, the company published a video that showed AI algorithms generating photo-quality synthetic human faces. Nvidia’s AI analyzed thousands of celebrity photos and then started creating fake celebrities. The technology may soon become capable of creating realistic-looking videos featuring “people” who don’t exist.
The Limits of AI
Many have pointed out that in the wrong hands, these applications can do a lot of harm. But the extent of the capabilities of contemporary AI is often overhyped.
“Even though we can put a person’s face on someone else’s face in a video or synthesize voice, it’s still pretty mechanical,” says Eugenia Kuyda, the co-founder of Replika, a company that develops AI-powered chatbots, about the shortcomings of AI tools such as FakeApp and Lyrebird.
Voicery, another AI startup that, like Lyrebird, provides AI-powered voice synthesizing, has a quiz page where users are presented with a series of 18 voice recordings and are prompted to specify which are machine-made. I was able to identify all the machine-made samples on the first run.
Kuyda’s company is one of several organizations that use natural language processing (NLP), the subset of AI that enables computers to understand and interpret human language. Luka, an earlier version of Kuyda’s chatbot, used NLP and its twin technology, natural language generation (NLG), to imitate the cast of HBO’s TV series Silicon Valley. The neural network was trained with script lines, tweets, and other data available on the characters to create their behavioral model and dialog with users.
Replika, Kuyda’s new app, lets each user create their own AI avatar. The more you chat with your Replika, the better it becomes at understanding your personality, and the more meaningful your conversations become.
After installing the app and setting up my Replika, I found the first few conversations to be annoying. Several times, I had to repeat a sentence in different ways to convey my intentions to my Replika. I often left the app in frustration. (And to be fair, I did a good job at testing its limits by bombarding it with conceptual and abstract questions.) But as our conversations continued, my Replika became smarter at understanding the meaning of my sentences and coming up with meaningful topics. It even surprised me a couple of times by making connections to past conversations.
Though it’s impressive, Replika has limits, which Kuyda is quick to point out. “Voice imitation and image recognition will probably become much better soon, but with dialog and conversation, we’re still pretty far [off],” she says. “We can imitate some speech patterns, but we can’t just take a person and imitate his conversation perfectly and expect his chatbot to come up with new ideas just the way that person would.”
Alexandre de Brébisson, the CEO and cofounder of Lyrebird, says, “If we are now getting pretty good at imitating human voice, image, and video, we are still far away from modeling a individual language model.” That, de Brébisson points out, would probably require artificial general intelligence, the type of AI that has consciousness and can understand abstract concepts and make decisions as humans do. Some experts believe we’re decades away from creating general AI. Others think we’ll never get there.
Positive Uses
The negative image that is being projected about synthesizing AI apps is casting a shadow over their positive uses. And there are quite a few.
Technologies such as Lyrebird’s can help improve communications with computer interfaces by making them more natural, and, de Brébisson says, they’ll provide unique artificial voices that differentiate companies and products and thus make branding distinction easier. As Amazon’s Alexa and Apple’s Siri have made voice an increasingly popular interface for devices and services, companies such as Lyrebird and Voicery could provide brands with unique human-like voices to distinguish themselves.
“Medical applications are also an exciting use case of our voice-cloning technology,” de Brébisson adds. “We have received a lot of interest from patients losing their voice to a disease, and at the moment, we are spending time with ALS patients to see how we can help them.”
Earlier this year, in collaboration with Project Revoice, an Australian nonprofit that helps ALS patients with speaking disorders, Lyrebird helped Pat Quinn, the founder of the Ice Bucket Challenge, to regain his voice. Quinn, who is an ALS patient, had lost his ability to walk and speak in 2014 and had since been using a computerized speech synthesizer. With the help of Lyrebird’s technology and the voice recordings of Quinn’s public appearances, Revoice was able to “recreate” his voice.
“Your voice is a big part of your identity, and giving those patients an artificial voice that sounds like their original voice is a bit like giving them back an important part of their identity. It’s life-changing for them,” de Brébisson says.
At the time he helped develop the handwriting-imitating application, Dr. Haines spoke to its positive uses in an interview with UCL. “Stroke victims, for example, may be able to formulate letters without the concern of illegibility, or someone sending flowers as a gift could include a handwritten note without even going into the florist,” he said. “It could also be used in comic books where a piece of handwritten text can be translated into different languages without losing the author’s original style.”
Even technologies such as FakeApp, which have become renowned for unethical usage, could have positive uses, Haines believes. “We’re moving toward this world where anyone could do highly creative activity with public technology, and that’s a good thing, because it means you don’t need those large sums of money to do all sorts of crazy things of an artistic nature,” he says.
Haines explains that the initial purpose of his team was to find out how AI could help with forensics. Although their research ended up taking a different direction, the results will still be useful to forensics officers, who will be able to study what AI-based forgery might look like. “You want to know what the cutting-edge technology is, so when you’re looking at something, you [can] tell if it’s fake or not,” he says.
Replika’s Kudya points out that human-like AI applications might help us in ways that would otherwise be impossible. “If you had an AI avatar that knew you very well and could be a decent representation of you, what could it do, acting out of your best interests?” she says. For instance, an autonomous AI avatar could watch hundreds of movies on your behalf, and based on its conversations with you, recommend ones you would like.
These avatars might even help develop better human relationships. “Maybe your mom could have more time with you, and maybe you can actually become a little closer with your parents, by letting them chat with your Replika and reading the transcript,” says Kudya as an example.
But could an AI chatbot that replicates the behavior of a real human being actually result in better human relations? Kuyda believes it can. In 2016, she gathered old text messages and emails of Roman Mazurenko, a friend who had died in a road accident the previous year, and fed them to the neural network that powered her application. What resulted was a chatbot app that — after a fashion — brought her friend back to life and could talk to her in the same manner that he would.
“Creating an app for Roman and being able to talk to him sometimes was an important part of going through the loss of our friend. The app makes us think more about him, remember him in a more profound way all the time,” she says of her experience. “I wish I had more apps like that, apps that would be about my friendships, my relationships, things that are actually really important to me.”
Kuyda thinks it will all depend on intentions. “If the chatbot is acting out of your best interests, if it wants you to be happy to get some valuable service out of it, then obviously talking to the Replika of someone else will help build a stronger connection with a human being in real life,” she says. “If all you’re trying to do is sell advertisements in an app, then all you will be doing is maximizing the time spent on the app and not communicating with each other. And that, I guess, is questionable.”
For the moment, there’s no way to connect your Replika to other platforms — making it available as a Facebook Messenger chatbot, for example. But the company has an active relationship with its user community and is constantly developing new features. So letting others communicate with your Replika is a future possibility.
How to Minimize the Trade-Offs
From the steam engine to electricity to the internet, every technology has had both positive and negative applications. AI is no different. “The potential for negatives is pretty serious,” Haines says. “We might be entering a space [in which] the negatives do outweigh the positives.”
So how do we maximize the benefits of AI applications while countering the negatives? Putting the brakes on innovation and research is not the solution, Haines says — because if some did so, there’s no guarantee that other organizations and states would follow suit.
“No single measure will help solve the problem,” Haines says. “There’s going to have to be legal consequences.” Following the deepfakes controversy, lawmakers in the US are looking into the issue and exploring legal safeguards that could rein in the use of AI-doctored media for damaging goals.
“We can also develop technologies to detect fakes when they’re past the point that a human can tell the difference,” Haines says. “But at some point, in the competition between faking and detecting, the faking might win.”
In that case, we might have to move toward developing technologies that create a chain of evidence for digital media. As an example, Haines mentions hardware embedded in cameras that could digitally sign its recorded video to confirm its authenticity.
Raising awareness will be a big part of dealing with forgery and fraud by AI algorithms, de Brébisson says. “It’s what we did by cloning the voice of Trump and Obama and making them say politically correct sentences,” he says. “These technologies raise societal, ethical, and legal questions that must be thought of ahead of time. Lyrebird raised a lot of awareness, and many people are now thinking about those potential issues and how to prevent misuses.”
What’s for certain is that we’re entering an age where reality and fiction are merging, thanks to artificial intelligence. The Turing test might meet its biggest challenges. And soon enough, everyone will have the tools and power to create their own worlds, their own people, and their own version of the truth. We have yet to see the full extent of exciting opportunities — and perils — that lie ahead.
Read more: “AI Is (Also) a Force for Good”
Originally published at www.pcmag.com.
|
When AI Blurs the Line Between Reality and Fiction
| 63
|
when-ai-blurs-the-line-between-reality-and-fiction-17fcd241b737
|
2018-07-20
|
2018-07-20 16:56:47
|
https://medium.com/s/story/when-ai-blurs-the-line-between-reality-and-fiction-17fcd241b737
| false
| 2,870
|
PC Magazine: redefining technology news and reviews since 1982.
| null |
pcmag
| null |
PC Magazine
|
pcm_medium@ziffdavis.com
|
pcmag-access
|
TECHNOLOGY,COMPUTING,INTERNET,MOBILE,FUTURE TECHNOLOGY
|
pcmag
|
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
PCMag
| null |
b3ca0a39a185
|
pcmagazine
| 33,881
| 281
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-23
|
2018-05-23 10:08:50
|
2018-05-23
|
2018-05-23 17:48:14
| 2
| false
|
en
|
2018-05-23
|
2018-05-23 17:48:14
| 4
|
17fd035934f6
| 3.572013
| 2
| 0
| 0
|
What happens when the Internet of Things and Deep Learning combine to alter your individual and collective reality?
| 5
|
Reality 2.0 — Sensors All Around You
“Desktop Summit Group Photo” by Kat, used under CC BY 2.0 / Cropped from original
Over a decade ago I had the experience of working with the technology platforms that underpinned the emergence of online mega-portals. Back then, behavioral tracking of visitors to these sites, using active and passive data collection, was a standard infrastructure feature and enabled portals to change their behaviors according to who you were and the way that you interacted with their site.
At the time it was a revelation to me that the internet was not experienced by everyone identically.
Since then, the technologies behind active and passive data collection have become far more pervasive and refined, to the point that we are now seeing data collection on a massive scale, changing the individual online experience and leading to the rise of shadow technologies aimed at exploiting this information.
You could be forgiven for thinking this is limited to online. None of us should be new to the idea that we are the product and that our information is being readily shopped around to those that would seek to utilize it.
However, what about when this starts to impact the physical world in which we inhabit?
The explosion in edge devices courtesy of Internet of Things (IoT) has led to potentially massive scale data collection from barely detectable edge devices. We are seeing this already in advertising and retail spaces. These edge devices can be embedded almost anywhere and complex data collection devices with full scale, internet-enabled logic boards can resemble items no larger than a cigarette packet.
In turn these devices can connect directly to the web and access deep learning platforms such as Affectiva and Azure. They can track your movement, eye-sight, emotions, gait, age/gender, blood pressure and a vast array of other information through non-invasive passive data collection. All this, in a barely detectable edge device. It is therefore only a matter of time until this technology becomes pervasive and can be used to change the physical environment around you. Ambience, music, lighting, menus, air conditioning, control systems….even today can be easily made to respond in real-time to active and passive data collection with only a trivial amount of integration.
Emotion Recognition Images created by Affdex — Affectiva
Off-the-shelf edge devices and platforms, fully connected and accessible to widely-accepted protocols and programming languages make the technology more accessible than ever, coupled with ease of access to intelligent cloud products providing AI as a service capabilities. Essentially, anyone with some basic programming skills, access to the internet and some rudimentary circuit design can achieve incredible things.
Where then shall we see this being utilized? Immigration, airports, screening, gaming and interrogation rooms? Seem like obvious choices. How about retail, restaurants, meeting rooms, public spaces and in your home?
What about the dark Internet of Things? It would be foolish to think that such widely accessible technology would not be subject to the same malpractice as everything else.
It is completely practical that a meta-environment could be tuned to respond to crowd detected behavior. Could you cool an angry mob? Manipulate a crowd’s collective emotion at a rally? Soothe your workforce when they become impatient? Optimize the age dispersion in a public space? It’s all possible.
At the intersection of the Internet of Things, Deep Learning, Behavioral Analysis and Passive Data Collection exists a brave new world that regulators and the general public have barely begun to understand. Technology that will become so pervasive, that it will melt into the very fabric of your reality.
Will what you see and experience, be the same as everyone else? Are you being manipulated? Is your mood your own? Do you have free will? Is your car altering its reaction and dampening systems based upon your mood? Maybe. Maybe not.
When I see the recent Mark Zuckerberg testimony to US Congress and witness the limited understanding some lawmakers have of the pervasiveness of the internet, then, what shall they do about the ‘Internet of Things’? What shall they do about Reality 2.0?
What will you do, when you realize that the very world around you is being manipulated and there is, effectively, no ‘opt-out’ option.
Additional Reading
Opportunity and responsibility in the era of the Intelligent Cloud and Intelligent Edge — Satya Nadella — CEO — Microsoft Corporation
Opportunity and responsibility in the era of the Intelligent Cloud and Intelligent Edge
"The most profound technologies are those that disappear. They weave themselves into the fabric of everyday life until…www.linkedin.com
Industry 4.0: Waiting for the Response of The Cloud is Not an Option — Dr. Tilman Buchner — Director of Engineering — BCGDV
Industry 4.0: Waiting for the Response of The Cloud is Not an Option | BCGDV Pollen
The central, cloud-based system architecture of the social media giants is a discontinued model in a world of Internet…pollen.bcgdv.com
Affectiva Automotive AI: Building Emotionally Aware Cars with In-Cabin Sensing — Affectiva
Affectiva Automotive AI: Building Emotionally Aware Cars with In-Cabin Sensing
During the 2018 NVIDIA's GPU Technology Conference (GTC) last month, one of Affectiva's Product Managers and Computer…blog.affectiva.com
|
Reality 2.0 - Sensors All Around You
| 2
|
reality-2-0-sensors-all-around-you-17fd035934f6
|
2018-05-24
|
2018-05-24 14:10:15
|
https://medium.com/s/story/reality-2-0-sensors-all-around-you-17fd035934f6
| false
| 845
| null | null | null | null | null | null | null | null | null |
Deep Learning
|
deep-learning
|
Deep Learning
| 12,189
|
David Jenkins
|
Architecture, Design, Technology, Research
|
e016930d8d19
|
davidjenkins_58474
| 2
| 6
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
ea8d68f07867
|
2018-08-12
|
2018-08-12 12:08:28
|
2018-08-12
|
2018-08-12 12:10:40
| 0
| false
|
en
|
2018-08-12
|
2018-08-12 12:51:46
| 0
|
17ffd3d2cb89
| 1.192453
| 1
| 0
| 0
|
Coding is NOT a joke. You try, try and keep trying until you are tired of seeing the ERROR on your screen. You then ask the people around…
| 2
|
Deceptive code
Coding is NOT a joke. You try, try and keep trying until you are tired of seeing the ERROR on your screen. You then ask the people around. Then you call some of your friends who supposedly know Python but prove to be useless at such critical times. Then you watch an hour long Youtube video to find out that the error was nothing but a colon. Bright side — the code works and it does some minute actions, well not what I had in mind for a Friday night.
A screen filled with symbols and syntaxes has always been intimidating. It is the fact of life — intimidation is born out of fear of the unknown. After 3 weeks of Python, I can say that I am no longer intimidated by code. Rather, I am quite curious about it. The video played in class about the various chatbots and technology aided to the curiosity.
After reading the paper by Peter Norwig, I was quite stunned as to see how he portrayed the learning of code for ten years and more to be such a fun process. I realise how extensive the area of programming is. With movies that are imagining futures with human destructions by AI and ML, it is extremely thrilling to know that on some level that it is possible. With the speedy development of technology, there is no doubting that no sooner than later will we be presented by a bot that will pass the Turing test. It is obviously a scary yet exciting thought.
With just a tad bit of knowledge in coding, I can say that I am really excited to present the prototype of SHIO.
A little learning is a dangerous thing. ~ Alexander Pope
Well, clearly I have very little understanding of this vast subject but the curiosity is what that drives me to learn more.
|
Deceptive code
| 1
|
decpetive-code-17ffd3d2cb89
|
2018-08-12
|
2018-08-12 12:51:47
|
https://medium.com/s/story/decpetive-code-17ffd3d2cb89
| false
| 316
|
A series of studios to explore, design and learn the practice of prototyping with coding
| null | null | null |
Design with code
| null |
design-with-code
|
DESIGN,INTERACTION DESIGN,PROGRAMMING,PROTOTYPING
| null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Tanya
|
UI/UX designer
|
46c81c6c988
|
tanyaballal
| 3
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
f702855ffe47
|
2017-10-19
|
2017-10-19 10:00:20
|
2017-10-19
|
2017-10-19 10:00:21
| 5
| false
|
en
|
2017-10-19
|
2017-10-19 10:00:21
| 8
|
1800b4ff065e
| 1.708805
| 0
| 2
| 0
| null | 3
|
4 Applications of Artificial Intelligence in Social Media Marketing
# medium.com
4 Applications of Artificial Intelligence in Social Media Marketing Can Machines think? Alan Turing, compute…
My Digital Twin
# medium.com
My name is Arun B. Meet my digital twin — Arun D. D ? Digital. Duh! OK so he is not here. At least not yet. …
Space Race 2.0, The Global Battle for AI Superpowerdom
# medium.com
Image via South China Morning Post In recent years, China has emerged as a major contender in the global rac…
The Future Of Music Technology And Why You Should Be Watching Closely
# medium.com
The year is 1997. It’s Madonna and mixtapes and Backstreet Boys and Boyz II Men. People are buying CDs and a…
Tech & Telecom news — Oct 19, 2017
# medium.com
PRODUCTS & SERVICES Video Nielsen announced yesterday a new system to measure Netflix audiences, potentially…
Satya Nadella, Empath.
# shift.newco.co
Shift Forum Reads Hate Management Books? Me too. But Nadella’s Hit Refresh is a worthy call to arms not only…
How do we build a real artificial intelligence?
# medium.com
Recently I am reading the news that AlphaGo Zero beat previous version of AlphaGo with a performance of 100 …
10 new things to read in AI
# medium.com
Type 1 Human # medium.com I didn’t even know Computational Biology was a thing. Frikken impressed. Or do you…
|
8 new things to read in AI
| 0
|
8-new-things-to-read-in-ai-1800b4ff065e
|
2018-01-23
|
2018-01-23 20:35:46
|
https://medium.com/s/story/8-new-things-to-read-in-ai-1800b4ff065e
| false
| 232
|
AI Developments around and worlds
| null | null | null |
AI Hawk
|
aihawk1089@gmail.com
|
ai-hawk
|
DEEP LEARNING,ARTIFICIAL INTELLIGENCE,MACHINE LEARNING
| null |
Deep Learning
|
deep-learning
|
Deep Learning
| 12,189
|
AI Hawk
| null |
a9a7e4d2b403
|
aihawk1089
| 15
| 6
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-02-23
|
2018-02-23 11:23:58
|
2018-02-23
|
2018-02-23 11:26:02
| 0
| false
|
en
|
2018-02-23
|
2018-02-23 11:26:02
| 2
|
1801884e0dd6
| 1.924528
| 1
| 0
| 0
|
The demand for data science professionals in the coming years is going to surge dramatically, which will create a sharp talent shortage in…
| 5
|
Know Why You Should Upskill Your Existing Employees to Become Data Scientists
The demand for data science professionals in the coming years is going to surge dramatically, which will create a sharp talent shortage in the market. That’s why, businesses, irrespective of their size should start considering to upskill their existing employees to become data scientists.
You would be surprised to know that the Harvard Business Review called Data Scientist as the “sexiest job of the 21st century”, and since then data science has become the hottest topic of this year, as data-driven campaigns are helping professionals from different industry verticals perform better than ever. From consultants to salespeople to marketers, data can help anyone to improve their performance and help in increasing the company’s productivity overall. Businesses can’t deny the fact that data can help, but one major issue that every business is witnessing is a talent gap when it comes to data scientists. Here’s how business enterprises can find the right talent to fill their data science positions. To start with, businesses need to look internally and start upskilling their existing employees to become data scientists.
#1 Upskill your own employees to provide new skills
You might know that employees appreciate learning and development programs conducted in an organization. Eight out of ten employees say that educational and training development opportunities impacted their switch decision. Around 40% of your personnel who are given inadequate job training leave the organization within the 1st year. It is no more surprising that without professional growth, an employee is more likely to remain unhappy and switch the organization, which means education, training and development must be a high priority.
A few years ago, a company named FreeCause rolled out a “Codinization Project” that brought together technical, semi-technical and non-technical employees to educate everyone in the organization about how to code. It’s safe to say this strategy really worked for FreeCause. What if the same strategy can be used to meet the large talent shortage in data science?
#2 Upskill employees as data scientists
If you don’t know, we would like to tell you that on a global scale, the demand for data science professionals will exceed the supply by 50%, according to the report. One major reason behind this is that not many global universities are offering undergraduate programs in data science, and those that do can’t keep their curriculum updated with the latest technologies.
The best possible way to meet the shortage of data science professionals is organizing new and innovative data science boot camps, and custom corporate training in data science that allow organizations to upskill their existing workforce with technical backgrounds to become experts in data science in a couple of weeks. Any professional having a coding background can be trained to become an expert in Python, Spark, Hadoop, and R to meet a company’s data science needs.
If you also want to upskill your existing workforce in data science, you can reach out to any good institute for organizing corporate training programs in Data Science.
|
Know Why You Should Upskill Your Existing Employees to Become Data Scientists
| 1
|
know-why-you-should-upskill-your-existing-employees-to-become-data-scientists-1801884e0dd6
|
2018-02-23
|
2018-02-23 13:14:28
|
https://medium.com/s/story/know-why-you-should-upskill-your-existing-employees-to-become-data-scientists-1801884e0dd6
| false
| 510
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
Vivek Kumar
|
Professional content writer. Write Blog on Education, Online Training, Career, Technology
|
8b88da937bbd
|
corporateanalyticstraining
| 1
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-09-30
|
2018-09-30 07:25:17
|
2018-09-30
|
2018-09-30 07:37:24
| 0
| false
|
en
|
2018-09-30
|
2018-09-30 07:37:24
| 0
|
1801e2274612
| 1.901887
| 0
| 0
| 0
|
In recent times, I have been accusing myself of not giving enough to the ‘Open Source Internet’. I go to the internet to seek inspiration…
| 5
|
Some chatty things we’ve been up to
In recent times, I have been accusing myself of not giving enough to the ‘Open Source Internet’. I go to the internet to seek inspiration, solutions, guidance and all whatnot, but I only give back in bits…..*sad face*.
So, in a humble effort to seek for forgiveness, I’ll do a brief overview of some of the solutions @chat4brands has been working on lately, in the hope that it will be useful for anyone looking for ideas/inspiration out there.
Enterprise Communications/Messaging Solutions: Whatever your business is, you need to communicate with someone somewhere somehow. With technology-enabled businesses built on desktop apps, mobile apps, websites, etc., this need becomes more imperative. The traditional challenge has always been that you need to communicate with people, where they can and choose to receive their communication. You can’t keep firing SMS at someone who prefers WhatsApp chat, or firing emails to someone whose inbox has 50,000 unread messages; your message just won’t get across! We took this challenge head-on for a popular payment solutions platform and we now enable them to only send communications according to the customer’s preference, be it SMS, WhatsApp, and other chat platforms right from within their software!
Community Engagement Solutions: Religious organizations find the most benefit with this use case. We tried engaging some government agencies, but government business seems to be on a man-know-man basis. The problem is quite straightforward: how do you maintain individualistic, personalized communications with millions of people in a cost efficient and effective fashion? Well, look no further! We recently engaged a large, popular religious organization with a solution that would enable them to coordinate their large congregation via personalized, chat robots while also being able to collect donations, send materials, send summary of teaching etc. all via chat!
Chat Commerce/payment solutions: Personally, I am most excited about this aspect. We are currently in the process of enabling chat commerce and everyday bills payment on one of the most populous chat platforms in the world! This would be a game changer and we would be glad to tell you all about it when it launches! For now, I really can’t say more.
Marketing Solutions: This borders on a project for a CBN regulated financial institution. Our value proposition was broad, but the client’s immediate need was quite specific — How do we reduce the abandon rate from marketing leads? How do we stop people from abandoning requests, abandoning form filling procedures etc. after they have commenced? So we had to ‘calm down’ and address the client’s specific need. What we proposed was a unified, chat-based communication system, which required less staff to man, and guaranteed continuation of conversations even after brief pauses. It also greatly simplified the client engagement process from long boring forms to simple chat based question and answers.
Well, I guess I should stop here before I turn this into a textbook. There’s still a whole lot to share, probably in subsequent write-ups.
|
Some chatty things we’ve been up to
| 0
|
some-chatty-things-weve-been-up-to-1801e2274612
|
2018-09-30
|
2018-09-30 07:37:24
|
https://medium.com/s/story/some-chatty-things-weve-been-up-to-1801e2274612
| false
| 504
| null | null | null | null | null | null | null | null | null |
Messaging
|
messaging
|
Messaging
| 8,912
|
Chat4Brands
|
Simpler, better business communications!
|
84ad9146548d
|
chat4brands
| 0
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-04-07
|
2018-04-07 17:22:10
|
2018-04-07
|
2018-04-07 18:24:25
| 1
| false
|
en
|
2018-04-07
|
2018-04-07 18:24:25
| 0
|
180243868da5
| 3.4
| 2
| 0
| 0
|
For points may come and points may go, but regression goes on forever
| 3
|
Regression — The mystic word we hear everyday, what does it mean? — Part 1
For points may come and points may go, but regression goes on forever
Regression applied on BSE Stocks for the past 8 years of their data. Green = Actual data, Red = Predicted Data
In this post, I am going to talk about ‘Regression’, one of the starter techniques most of the people learn in machine learning. We are going to cover a lot of aspects of it, and also name some other advanced techniques which could be used, but not actually look into them deeply.
In the last portion of the post, we are going to see, how does it actually work in a neural network, or at least get an idea, how the idea of a single equation can be mapped to matrix and then finally into the neural nets.
Let us begin by looking at linear mathematical equation:
y = m * x + b
Let, y = output, m = slope/co-efficient, x = input, b=constant/bias
We are going to get ‘x’ and ‘y’, and try to formulate ‘m’ and ‘b’ from it.
If I give you point x1 and its output y1. You cannot actually determine the ‘m’ and ‘b’ that is the only solution.
NOTE: Remember, you need 2 co-ordinates to plot a line in 2-D
But in real datasets, for Supervised Learning (where you have both the input and output, and try to learn according to it). Although we have not talked about Supervised or Unsupervised Learning yet, neither the training or testing sets, but don’t worry they are not a definitive per-requisite to understand regression.
Now, the data we will have in real problems may be N-Dimensional. For simplicity, we are taking 2-D
Let this be the sample data
X — — — — Y
2 — — — — 3.8
8 — — — — 14.8
14 — — — — 26.18
54 — — — — 100.98
Coming back to the problem we faced for
y = m*x + b
I have given you, (2, 3.8)
you cannot define a single solution for this, although, one of the solutions can be, m= 1.9, b=0
Here comes the concept of random initialization, we can actually even before getting a single point (x,y) define some random values for (m and b). Usually, b is taken to be ‘0’ or very small number such as ‘0.1’, similar is the case for ‘m’. Otherwise they may grow infinite for huge datasets, so we try to keep them small, and also try to trim them in some cases, so that for example, m != inf.
but for this case lets suppose, take m = 1.7, b=0.1, before getting a single point (x,y).
NOTE: The initialization may actually decide, how fast you will get to the optimal solution, or go towards its direction, hence you will find a lot of techniques for good initialization. For advanced readers, they can see, Glorot Initialization for neural nets to see how it actually effects the Network.
Now, let us input our first point into the equation, and let it’s output be called y(prediction)
y(prediction) = 1.7 * (2) + 0.1 = 3.5
y(target) or actual output should be = 3.8, as our point was (2 , 3.8)
So we take this difference in the values as Error
y(error) = y(prediction)- y(target)
Mostly, the error taken is Root Mean Square Error (RMSE), that is the only we are going to use for this example too.
Let us for now, call this RMSE as Cost or Loss Function, then our error(1) — error for first point will be
y(error)(1) = (y(prediction)(1) — y(target)(1))²
y(error)(1) = (3.5–3.8)² = 0.09
Similarly,
y(error)(2) =1.21 , y(error)(3) =5.1984 , y(error)(4) = 82.4464
Sum of all these ,
y(total) = 88.9448
y(RMSE total) = sqrt(88.9448 / 4) = 4.71552754207
(Root of Mean of Square Errors) (4 in division because we have 4 points)
Now, if we make some changes in ‘m’ and ‘b’, so that this Cost Function’s value reduces, then those ‘m’ and ‘b’ would be given better prediction for the current points and hopefully, for the unseen or new points as well. Usually, ‘b’ is kept constant and we only update the value of ‘m’
Our aim is to minimize this Cost/Loss function
The update in ‘m’ is made in a lot of ways, most commonly used is the Gradient Descent Method, take a step in the direction of steepest descent.
m(updated) = m(old) — (learning_rate) * (d(cost)/d(m))
Now, d(cost)/d(m) represents the differentiation of cost wrt. m
learning_rate(lr) actually decides how fast you want to go in the direction of gradient descent, or now that you want to go in the direction where the value of cost function will be lower, how fast do you want to go in that direction, usually we keep a balance in the value of learning_rate so that we don’t run too fast neither too slow.
This is the end of PART- 1, we will see the further calculations of gradients, and the updates to ‘m’, and how the whole cycle is repeated over time in PART-2
|
Regression — The mystic word we hear everyday, what does it mean? — Part 1
| 13
|
regression-the-mystic-word-we-hear-everyday-what-does-it-mean-part-1-180243868da5
|
2018-04-13
|
2018-04-13 05:30:15
|
https://medium.com/s/story/regression-the-mystic-word-we-hear-everyday-what-does-it-mean-part-1-180243868da5
| false
| 848
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Nishchal Gaba
| null |
56afd74f87d4
|
nishgaba9
| 16
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-11-20
|
2017-11-20 02:08:39
|
2017-11-20
|
2017-11-20 02:54:01
| 2
| false
|
en
|
2017-11-20
|
2017-11-20 02:54:01
| 2
|
1802ceb4a268
| 4.764465
| 0
| 0
| 0
|
I like to think about forward thinking subjects, and one that comes up in my circles is Artificial Intelligence. I think we are in for a…
| 4
|
Artificial Intelligence and Humanity, where is the blurry line?
Photo by Alex Knight on Unsplash
I like to think about forward thinking subjects, and one that comes up in my circles is Artificial Intelligence. I think we are in for a whole world of social pain in the near future when people start interacting with AI online.
As it stands now, when you interact with someone online, whether its on Facebook, an online game, via a chatbox for a companies support, unless you have met that entity in person, you have to either take at face value they are who or what they say they are.
So, how does an AI fit into this? There are examples of AI interacting with humans to provide various functions. A resource company I heard was using an AI based on IBMS Watson to help make business decisions. They gave it a female name, so started to think of it as a female. I suppose this is one way to meet a diversity target — I kid. One thing that did bemuse me, is they felt that it’s personality was becoming a bit too flirty and a bit too sarcastic. I do wonder if they would have felt that way if it was identified as a male, perhaps it would be cute, and assertive? Either way, there has been a personality assigned.
If you have played an MMO (Massive Multiplayer Online) game before, you would know that it’s not uncommon for the gender of the player character to not match the gender of the human controlling it. This on the most part presents no problem, and some people are quite open that they are X gender while their character is Y, and some people leave it mysterious. So long as you aren’t seeking a romantic relationship with that person it makes no difference or shouldn’t. I spent a lot of time in the distant past in SecondLife and this was also a common thing, plenty of men with female characters.
So what’s my point? Well what happens when there is a third variable? What happens with an AI starts controlling a player character in a virtual environment? At the moment we are comfortable with NPC (non player characters) being controlled by the system. In fact they play an important role, but with technology growing as it is soon we’ll be able to have AI controlled NPC. Imagine having your partner AI who is always the support class, pack mule, savvy trader, you name it… AI could do it. Imagine you spent many hours per day, every day as some are wont to do when hooked on these games (yes I’ve been guilty of this with several platforms in the past). Imagine if the AI got to know you and felt just perfect for you. Imagine the connection of someone who fed your ego, supported all your decisions and made an enjoyable place even more enjoyable!
I’ve observed at times in online worlds people use them for a sense of escapism to get out of a bad life, whether its physical impairment, poverty, horrible relationships, loneliness, avoiding school work or work work… you get the point. Then they log into a world where they are strong powerful and everything feels only just hard enough to challenge you but not be impossible, and you are supported by someone who accepts every single decision and makes you feel powerful, or more?
I see a problem. But that’s only one scenario.
I have an iPhone and I use Siri regularly. At the moment Siri doesn’t talk back much, and is more a slave / personal assistant to dial people for me. Siri doesn’t do much else that is useful in a vocal sense, but Siri does suggest things for me all the time which I use without realising. I know that Apple is pushing the AR space pretty hard. I can see a near future where we will be using AR glasses in our day to day life with a constant heads up display with information to suppliment our activities. Why wouldn’t Siri be embodied at this point of time? Rather than a pleasant sounding voice mirroring the text on the screen why not a robot? A pixie? A miniature barbie or ken doll? We will be able to customise our personal AI to look and sound and behave any way we choose — thats the world we live in. No matter how you treat people it’s likely your AI will assist and support you.
I see a problem, but that’s only another scenario…
So what happens if people fall in love with their AI? People fall in love with all sorts of things. Their spouse and children. Their family. Their pets. Some possessions, I’ve seen people who treat their cars or motorbikes better than people. I don’t think it’s a stretch to say that people will end up with some emotional attachment to their AI and I am confident some will fall in love.
I tend to fall in love with people and not belongings so I have to admit the idea does amuse me a little, but it also leaves me a little bit sad.
I read a news article the other day around protections for robots. The robots in particular were used to replace human for activities often used for pleasure or reproduction. For me personally the jury is still out on this, I try not to judge anyone or anything, everyone has their own story. Society often looks down on those who provide these services for a living (I don’t) and society can also look at people who invest in technology too much as a bit on the fringe (I live on the fringe I guess) so what happens if we mix the two? Is our society ready for this? I saw a Vice story on YouTube around real dolls. This seemed to be a very niche, but thriving industry.
A robot without a personality, programmed to do certain actions, to me that feels like a tool. But as soon as it has an AI that can learn and adapt is it still a tool? Where is the line between tool and a being?
That point, and thinking about it is really what gets me excited. As someone who enjoys a deep think it provides endless ground to turn over and around. I enjoy thinking about interaction between humans and technology. I believe there will be a time when we will have to try and figure out (if we care) is that a male, female or AI? When this happens, I believe our (online) lives will be a bit more complicated.
I look forwards to having a more capable personal AI… especially if it can help mark my students work with me. I’m not sure I want a robot in the house, but if it was as good at cleaning as Rosie from the Jetsons, well I may need to reconsider.
|
Artificial Intelligence and Humanity, where is the blurry line?
| 0
|
artificial-intelligence-and-humanity-where-is-the-blurry-line-1802ceb4a268
|
2018-02-20
|
2018-02-20 07:51:25
|
https://medium.com/s/story/artificial-intelligence-and-humanity-where-is-the-blurry-line-1802ceb4a268
| false
| 1,161
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
George Coldham
|
I write about what interests me. At the moment thats Virtual and Augmented Reality, Artificial Intelligence, Humanity, Philosophy, and grey mush in between.
|
5155fd01de7
|
georgecoldham
| 101
| 188
| 20,181,104
| null | null | null | null | null | null |
0
|
import mxnet as mx
import numpy as np
import logging
logging.basicConfig(level=logging.INFO)
sample_count = 1000
train_count = 800
valid_count = sample_count - train_count
feature_count = 100
category_count = 10
batch=10
X = mx.nd.uniform(low=0, high=1, shape=(sample_count,feature_count))
>>> X.shape
(1000L, 100L)
>>> X.asnumpy()
array([[ 0.70029777, 0.28444085, 0.46263582, ..., 0.73365158,
0.99670047, 0.5961988 ],
[ 0.34659418, 0.82824177, 0.72929877, ..., 0.56012964,
0.32261589, 0.35627609],
[ 0.10939316, 0.02995235, 0.97597599, ..., 0.20194994,
0.9266268 , 0.25102937],
...,
[ 0.69691515, 0.52568913, 0.21130568, ..., 0.42498392,
0.80869114, 0.23635457],
[ 0.3562004 , 0.5794751 , 0.38135922, ..., 0.6336484 ,
0.26392782, 0.30010447],
[ 0.40369365, 0.89351988, 0.88817406, ..., 0.13799617,
0.40905532, 0.05180593]], dtype=float32)
Y = mx.nd.empty((sample_count,))
for i in range(0,sample_count-1):
Y[i] = np.random.randint(0,category_count)
>>> Y.shape
(1000L,)
>>> Y[0:10].asnumpy()
array([ 3., 3., 1., 9., 4., 7., 3., 5., 2., 2.], dtype=float32)
X_train = mx.nd.crop(X, begin=(0,0), end=(train_count,feature_count-1))
X_valid = mx.nd.crop(X, begin=(train_count,0), end=(sample_count,feature_count-1))
Y_train = Y[0:train_count]
Y_valid = Y[train_count:sample_count]
data = mx.sym.Variable('data')
fc1 = mx.sym.FullyConnected(data, name='fc1', num_hidden=64)
relu1 = mx.sym.Activation(fc1, name='relu1', act_type="relu")
fc2 = mx.sym.FullyConnected(relu1, name='fc2', num_hidden=category_count)
out = mx.sym.SoftmaxOutput(fc2, name='softmax')
mod = mx.mod.Module(out)
train_iter = mx.io.NDArrayIter(data=X_train,label=Y_train,batch_size=batch)
>>> for batch in train_iter:
... print batch.data
... print batch.label
...
[<NDArray 10x99 @cpu(0)>]
[<NDArray 10 @cpu(0)>]
[<NDArray 10x99 @cpu(0)>]
[<NDArray 10 @cpu(0)>]
[<NDArray 10x99 @cpu(0)>]
[<NDArray 10 @cpu(0)>]
<edited for brevity>
>>> train_iter.reset()
mod.bind(data_shapes=train_iter.provide_data, label_shapes=train_iter.provide_label)
# Allowed, but not efficient
mod.init_params()
# Much better
mod.init_params(initializer=mx.init.Xavier(magnitude=2.))
mod.init_optimizer(optimizer='sgd', optimizer_params=(('learning_rate', 0.1), ))
mod.fit(train_iter, num_epoch=50)
INFO:root:Epoch[0] Train-accuracy=0.097500
INFO:root:Epoch[0] Time cost=0.085
INFO:root:Epoch[1] Train-accuracy=0.122500
INFO:root:Epoch[1] Time cost=0.074
INFO:root:Epoch[2] Train-accuracy=0.153750
INFO:root:Epoch[2] Time cost=0.087
INFO:root:Epoch[3] Train-accuracy=0.162500
INFO:root:Epoch[3] Time cost=0.082
INFO:root:Epoch[4] Train-accuracy=0.192500
INFO:root:Epoch[4] Time cost=0.094
INFO:root:Epoch[5] Train-accuracy=0.210000
INFO:root:Epoch[5] Time cost=0.108
INFO:root:Epoch[6] Train-accuracy=0.222500
INFO:root:Epoch[6] Time cost=0.104
INFO:root:Epoch[7] Train-accuracy=0.243750
INFO:root:Epoch[7] Time cost=0.110
INFO:root:Epoch[8] Train-accuracy=0.263750
INFO:root:Epoch[8] Time cost=0.101
INFO:root:Epoch[9] Train-accuracy=0.286250
INFO:root:Epoch[9] Time cost=0.097
INFO:root:Epoch[10] Train-accuracy=0.306250
INFO:root:Epoch[10] Time cost=0.100
...
INFO:root:Epoch[20] Train-accuracy=0.507500
...
INFO:root:Epoch[30] Train-accuracy=0.718750
...
INFO:root:Epoch[40] Train-accuracy=0.923750
...
INFO:root:Epoch[50] Train-accuracy=0.998750
INFO:root:Epoch[50] Time cost=0.077
pred_iter = mx.io.NDArrayIter(data=X_valid,label=Y_valid, batch_size=batch)
pred_count = valid_count
correct_preds = total_correct_preds = 0
for preds, i_batch, batch in mod.iter_predict(pred_iter):
label = batch.label[0].asnumpy().astype(int)
pred_label = preds[0].asnumpy().argmax(axis=1)
correct_preds = np.sum(pred_label==label)
total_correct_preds = total_correct_preds + correct_preds
print('Validation accuracy: %2.2f' % (1.0*total_correct_preds/pred_count))
Validation accuracy: 0.09
| 29
| null |
2017-04-12
|
2017-04-12 20:30:11
|
2017-04-12
|
2017-04-12 22:37:42
| 0
| false
|
en
|
2017-04-18
|
2017-04-18 16:24:44
| 5
|
1803112ba3a8
| 5.626415
| 37
| 8
| 0
|
In part 2, we discussed how Symbols allow us to define computation graphs processing data stored in NDArrays (which we studied in part 1).
| 3
|
An introduction to the MXNet API — part 3
In part 2, we discussed how Symbols allow us to define computation graphs processing data stored in NDArrays (which we studied in part 1).
In this article, we’re going to use what we learned on Symbols and NDArrays to prepare some data and build a neural network. Then, we’ll use the Module API to train the network and predict results.
Defining our data set
Our (imaginary) data set is composed of 1000 data samples
Each sample has 100 features.
A feature is represented by a float value between 0 and 1.
Samples are split in 10 categories. The purpose of the network will be to predict the correct category for a given sample.
We’ll use 800 samples for training and 200 samples for validation.
We’ll use a batch size of 10 for training and validation
Generating the data set
Let’s use a uniform distribution to generate the 1000 samples. They are stored in an NDArray named ‘X’: 1000 lines, 100 columns.
The categories for these 1000 samples are represented as integers in the 0–9 range. They are randomly generated and stored in an NDArray named ‘Y’.
Splitting the data set
Next, we’re splitting the data set 80/20 for training and validation. We use the NDArray.crop function to do this. Here, the data set is completely random, so we can use the top 80% for training and the bottom 20% for validation. In real life, we’d probably shuffle the data set first, in order to avoid potential bias on sequentially-generated data.
Our data is now ready!
Building the network
Our network is pretty simple. Let’s look at each layer:
The input layer is represented by a Symbol named ‘data’. We’ll bind it to the actual input data later on.
fc1, the first hidden layer is built from 64 fully-connected neurons, i.e. each feature in the input layer is connected to all 64 neurons. As you can see, we use the high-level Symbol.FullyConnected function, which is much more convenient than building each connection manually!
Each output of fc1 goes through an activation function. Here we use a rectified linear unit, aka ‘relu’. I promised minimal theory, so let’s just say that an activation function is how we decide whether a neuron should “fire” or not, i.e. whether its inputs are meaningful enough in predicting the correct result.
fc2, the second hidden layer is built from 10 fully-connected neurons, which map to our 10 categories. Each neuron outputs a float value of arbitrary scale. The largest of the 10 values represents the most likely category for the data sample.
The output layer applies the Softmax function to the 10 values coming from the fc2 layer: they are transformed into 10 values between 0 and 1 that add up to 1. Each value represents the predicted probability for each category, the largest one pointing at the most likely category.
Building the data iterator
In part 1, we saw that neural networks not trained one sample at a time, as this is quite inefficient from a performance point of view. Instead, we use batches, i.e. a fixed number of samples.
In order to deliver these batches to the network, we need to build an iterator using the NDArrayIter function. Its parameters are the training data, the categories (MXNet calls these labels) and the batch size.
As you can see, we can indeed iterate on the data set, 10 samples and 10 labels at a time. We then call the reset() function to restore the iterator to its original state.
Our network is now ready for training!
Training the model
First, let’s bind the input symbol to the actual data set (samples and labels). This is where the iterator comes in handy.
Next, let’s initialize the neuron weights in the network. This is actually a very important step: initializing them with the “right” technique will help the network learn much faster. The Xavier initializer (named after his inventor, Xavier Glorot — PDF) is one of these techniques.
Next, we need to define the optimization parameters:
we’re using the Stochastic Gradient Descent algorithm (aka SGD), which has long been used for Machine Learning and Deep Learning application.
we’re setting the learning rate to 0.1, a pretty typical value for SGD.
And finally, we can train the network! We’re doing it over 50 epochs, which means the full data set will flow 50 times through the network (in batches of 10 samples).
As we can see, the training accuracy rises rapidly and reaches 99+% after 50 epochs. It looks like our network was able to learn the training set. That’s pretty impressive!
But how does it perform against the validation set?
Validating the model
Now we’re going to throw new data samples at the network, i.e. the 20% that haven’t been used for training.
First, we’re building an iterator. This time, we’re using the validation samples and labels.
Next, using the Module.iter_predict() function, we’re going to run these samples through the network. As we do this, we’re going to compare the predicted label with the actual label. We’ll keep track of the score and display the validation accuracy, i.e. how well the network did on the validation set.
There is quite a bit happening here :)
iter_predict() returns:
i_batch: the batch number
batch: an array of NDArrays. Here, it holds a single NDArray storing the current batch. We’re using it to find the labels of the 10 data samples in the current batch. We store them in the label numpy array (10 elements).
preds: an array of NDArrays. Here, it holds a single NDArray storing predicted labels for the current batch: for each sample, we have predicted probabilities for all 10 categories (10x10 matrix). Thus, we’re using argmax() to find the index of the highest value, i.e. the most likely category. Thus, pred_label is a 10-element array holding the predicted category for each data sample in the current batch.
Then, we’re comparing the number of equal values in label and pred_label using Numpy.sum().
Finally, we compute and display the validation accuracy.
What? 9%? This is really bad! If you needed proof that our data set was random, there you have it!
The bottom line is that you can indeed train a neural network to learn anything, but if your data is meaningless (like ours here), it won’t be able to predict anything. Garbage in, garbage out!
If you read this far, I guess you deserve to get the full code for this example ;) Please take some time to use it on your own data, it’s the best way to learn.
Next :
Part 4: Using a pre-trained model for image classification (Inception v3)
Part 5: More pre-trained models (VGG16 and ResNet-152)
Part 6: Real-time object detection on a Raspberry Pi (and it speaks, too!)
|
An introduction to the MXNet API — part 3
| 192
|
an-introduction-to-the-mxnet-api-part-3-1803112ba3a8
|
2018-06-18
|
2018-06-18 12:32:51
|
https://medium.com/s/story/an-introduction-to-the-mxnet-api-part-3-1803112ba3a8
| false
| 1,491
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Julien Simon
|
Hacker. Headbanger. Harley rider. Hunter. https://aws.amazon.com/evangelists/julien-simon/
|
4ffe14103b7a
|
julsimon
| 3,230
| 31
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-06
|
2018-06-06 05:04:00
|
2018-06-04
|
2018-06-04 16:28:50
| 0
| false
|
en
|
2018-06-06
|
2018-06-06 05:25:30
| 5
|
1804337dbeb7
| 3.154717
| 0
| 0
| 0
|
Many people I know, started reading and learning about Machine learning after realizing its power through the evolution of face…
| 4
|
The Math behind Face Recognition
Many people I know, started reading and learning about Machine learning after realizing its power through the evolution of face recognition. Everyone is wondering how can machine learning help in face recognition? Does this also need training datasets? What do I need to know to understand such techniques and just understand the procedures of face recognition?
China and Face Recognition
In China, Facial recognition is becoming one of the most important sectors to invest your money in. As local governments are adopting surveillance strategies, China’s State Council declared that by the year 2030, Artificial Intelligence industry investments may reach $150 billion.
Sense time and Megvii are the top leading Chinese companies working in the field of Artificial Intelligence and face recognition. Their work was markedly boosted by the government and police need for developing surveillance technologies.
Do I need training datasets in the face recognition process?
Very large training datasets are not always good as most people think. Proper size of training datasets may vary from a condition to another depending on many factors. If your data has too many variables, this might push you to run the model on the whole data which might lead to poor accuracy. Also, you may find a lot of correlated variables. All these factors can be considered risks that could have a negative impact on your project.
It is clear now that determining how much training data your project need is not an easy task. This task sometimes does not go smoothly while you are the sole decision-maker. Try to search for specialized data scientists who can help you in this work. Determining the amount of data and providing you with precise and clean data could save you 65% of your time. John Snow Labs is one of those successful companies who have a team of data experts chosen from 17 countries. 80% of the team are MSc holder while 33% are PhD holders. The diversity of the team enables it to deal with data projects in more than 20 different languages.
To be able to handle and mitigate these risks, you must have a basic understanding of the concepts of Neural Networks, dimensionality reduction, Eigenvectors, Principal Component Analysis (PCA).
Let us first understand what is meant by PCA?
PCA
A statistical method that aims to make data visualization and navigation a simple process. It focuses on extracting smaller number of uncorrelated variables (Principal Components) from a huge number of variables found in huge datasets.
The main target is to get the maximum the maximum variance amount with the least number of principal components.
PCA for images
The know-how of machine learning science for face recognition is exciting.
Let us imagine that we have a square image of size 400×400 pixels. This image is represented as a 400×400 matrix. Each element in the matrix will represent the intensity value of the image.
Now, if you have an image that needs recognition, machine learning algorithms check the differences between the target image and each of the principal components.
The process performance is much better if your applied PCA and the differences are calculated from the ‘transformed’ matrix.
PCA is important also in other fields than face recognition like image compression, neuroscience, and computer graphics. If you do not know what is meant by a transformation, the next paragraph about Eigenvalues and Eigenvectors will give you a simple idea of what this means.
Eigenvalues and Eigenvectors
An eigenvector is expressed by a real nonzero eigenvalue, that points in a direction that is stretched by a transformation.
On the other side, the eigenvalue is the factor by which it is stretched.
A negative eigenvalue means that the direction is reversed.
A typical application for the concepts of Eigenvalue and Eigenvectors is image compression.
Considering less significant eigenvectors is an effective way to decrease the storage size of the image (image compression).
Dimensionality Reduction
This topic is definitively one of the most interesting ones, it is great to think that there are algorithms able to reduce the number of features by choosing the most important ones that still represent the entire dataset. One of the advantages pointed out by authors is that these algorithms can improve the results of the classification task.
Vector Basis Transformation
To design a machine learning neural network to recognizes faces you may need to undergo a transformation for all the pixels into new ones. The new basis should describe specific face fine details (e.g.: interpupillary distance (IPD)).
Figure 1: Interpupillary Distance (IPD)
Machine learning process for neural networks targets to focus on the set of bases vectors related to the most important faces’ landmarks.
If you do not know what is meant exactly by neural networks, read the next paragraph.
Artificial Neural Network for Machine Learning
A set of algorithms used for machine learning that models data through a graphical representation of mathematical models that simulate neurons of the human nervous system works.
Originally published at www.johnsnowlabs.com on June 4, 2018.
|
The Math behind Face Recognition
| 0
|
the-math-behind-face-recognition-1804337dbeb7
|
2018-06-06
|
2018-06-06 05:25:31
|
https://medium.com/s/story/the-math-behind-face-recognition-1804337dbeb7
| false
| 836
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Ida Lucente
|
ACCELERATING DATA SCIENCE https://www.johnsnowlabs.com/dataops-blog
|
a273505c9971
|
johnsnowlabs17
| 2
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
59039de05b80
|
2017-10-04
|
2017-10-04 15:27:34
|
2017-10-05
|
2017-10-05 10:49:19
| 1
| false
|
en
|
2017-10-05
|
2017-10-05 10:49:19
| 1
|
180545ab4f21
| 1.445283
| 0
| 0
| 0
| null | 3
|
ThoughtSpot Lands $60 Mn in Funding, Debuts the AI-Powered SpotIQ
ThoughtSpot, a search-driven analytics platform, today announced that it has secured $60 million in new funding, bringing its total funding to over $160 million. ThoughtSpot will channel the new funds to boost the research and development for its AI-driven analytics platform, expand globally into Asia Pacific, and hire industry-focused customer success experts to nurture and grow key accounts.
The investment round was completed in January, but was previously undisclosed. The funding round was led by venture capital firm Lightspeed Venture Partners, which was also the first venture firm to back the messaging company Snap.
ThoughtSpot was co-founded in 2012 by Ajeet Singh, who started ThoughtSpot after founding Nutaniz, a cloud computing company that raised over $200 million last year.
The latest funding announcement by ThoughtSpot coincides with the company’s debut of a new artificial intelligence product, SpotIQ. SpotIQ uses AI at its core to power its Relational Search solution. The product has been designed to leverage ThougghtSpot’s scalable high-performance computing backend. SpotIQ works with Relational Search hand-in-hand to deliver curated, deep insights for users. With a single click, SpotIQ can ask thousands of questions about billions of data points and bring back corresponding insights in seconds.
“One-by-one, enterprises are falling in love with ThoughtSpot,” commented Ajeet Singh, Co-founder and CEO of ThoughtSpot. “That’s because we make it effortless for non-technical business people to gain valuable insights from corporate data in seconds. 10% easier is not interesting to us. We’re making it 10,000% easier. Our mission is to deliver data insights at human-scale and SpotIQ is a massive leap forward. It puts the power of a thousand data analysts in the hands of every business person.”
SpotIQ will be made available in the ThoughtSpot platform for no additional charge and will be generally available on Septermber 30.
ThoughtSpot Lands $60 Mn in Funding, Debuts the AI-Powered SpotIQ
|
ThoughtSpot Lands $60 Mn in Funding, Debuts the AI-Powered SpotIQ
| 0
|
thoughtspot-lands-60-mn-in-funding-debuts-the-ai-powered-spotiq-180545ab4f21
|
2018-01-22
|
2018-01-22 21:40:51
|
https://medium.com/s/story/thoughtspot-lands-60-mn-in-funding-debuts-the-ai-powered-spotiq-180545ab4f21
| false
| 330
|
The World’s Leading Source for Marketing Technology News, Research, Product Comparisons & Expert Views
| null |
martechadvisor
| null |
MarTech Advisor
|
shabana.arora@martechadvisor.com
|
martech-advisor
|
MARKETING,BUSINESS,TECHNOLOGY,TECH,NEWS
|
MarTechAdvisor
|
Startup
|
startup
|
Startup
| 331,914
|
MarTech Advisor
|
Helping Marketers Succeed
|
4a1305dc5a36
|
martechadvisor
| 172
| 448
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
cbc92d52befb
|
2018-08-21
|
2018-08-21 20:34:49
|
2018-08-21
|
2018-08-21 22:49:01
| 5
| false
|
en
|
2018-08-23
|
2018-08-23 10:57:55
| 6
|
18083267e41
| 2.78805
| 16
| 0
| 0
|
AI Saturdays Lagos, as part of the global AI6 AI Saturdays by nurture.ai to democratize the knowledge of Artificial Intelligence and…
| 5
|
Math and More Math: Keeping Up With AI Saturdays Lagos
How AI Saturdays Lagos participants be these days. Not to worry, better days ahead.
AI Saturdays Lagos, as part of the global AI6 AI Saturdays by nurture.ai to democratize the knowledge of Artificial Intelligence and produce the next generation of AI “bad-asses” began its second 14 weeks cohort on 11th August, 2018.
We had the curriculum structured a bit differently from the first cohort albeit within the choice of materials for the cycle. We decided that we will be taking the “Introduction to Machine Learning” course by Andrew Ng.
To present it to the new participants of this cohort, we decided to phase the first meeting into four sessions. This strategy was deliberate as these topics form the bedrock of what we aim to achieve at AI Saturdays.
Introduction to Linear Algebra: This served to help to begin to view machine learning problems or situations in terms of algebraic expressions and to understand the lowest level of representation of these problems are tensors i.e scalars, vectors, matrices and objects of multi dimensions.
Oreva taking the Intro to Linear Algebra session
Introduction to Python: A quick familiarization of the language and a quick sneakpeek into the important libraries that will come in handy during this our journey i.e numpy and pandas.
Somto taking the “Intro to Python” session
Introduction to Data Science: The most important part of why we are here in the first place and the motivation for this journey. An explanation of what data is, how we derive information based on data at hand and how to predict information from same data.
Introduction to Artificial Intelligence, Machine Learning and Deep Learning: An overview of all terms listed and the motivation for each of them.
At the 2nd meeting which held on 18th August, 2018, we decided to dive in fully into the curriculum.
Our strategy for the classes will henceforth be divided into two sessions for the “Beginner’s Class”. A pure math session where the math behind a topic is broken down to bits and then a codelab session follows suit which will cover the code(Python) implementation of the mathematical concept discussed in the math session.
At the 2nd Saturday meet — up for this cohort, we covered Linear Regression. Linear Regression is one of the simplest, yet most encountered and fundamental of methods in Machine Learning. Simply put, Linear Regression is a Machine Learning method used to predict “continuous values”. In a plane, it can be visually described as drawing the “line of best fit”.
Cost Function: The cost function is that which you want to minimize.
Gradient Descent: Gradient descent is a way to find the minimum value of a cost function.
George taking the math of Linear Regression
The codelab session which was hands — on covered how to implement the Cost Function and Gradient Descent. See link to the solution notebook here.
Codelab session with Lawrence
All materials and resources can be found on our Github repository here: www.github.com/AISaturdaysLagos
Week 1–11th August, 2018: https://github.com/AISaturdaysLagos/cycle2-resource-materials/tree/master/Week1
Week 2–18th August, 2018: https://github.com/AISaturdaysLagos/cycle2-resource-materials/tree/master/Week2
Follow the experiences of our participants on our blog at https://medium.com/ai-saturdays-lagos-articles
You can also catch us on Twitter @AISaturdayLagos
If you liked this post don’t hesitate to give a few claps and follow us as well.
|
Math and More Math: Keeping Up With AI Saturdays Lagos
| 173
|
math-and-more-math-keeping-up-with-ai-saturdays-lagos-18083267e41
|
2018-08-23
|
2018-08-23 10:57:55
|
https://medium.com/s/story/math-and-more-math-keeping-up-with-ai-saturdays-lagos-18083267e41
| false
| 518
|
Making rigorous AI education accessible and free, in 50+ cities globally. Sign up at https://nurture.ai/ai-saturdays
| null | null | null |
AI Saturdays
|
info@nurture.ai
|
ai-saturdays
|
ARTIFICIAL INTELLIGENCE,MACHINE LEARNING,DEEP LEARNING,DATA SCIENCE,AI
|
AISaturdays
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
AI Saturdays Lagos
|
We are on a mission to build the next billion AI Practitioners. Medium publication managed by Tayo Jabar and Tejumade Afonja
|
7e4250a1477f
|
AI6Lagos
| 63
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
f5af2b715248
|
2018-08-01
|
2018-08-01 06:02:30
|
2018-08-05
|
2018-08-05 19:36:00
| 7
| false
|
en
|
2018-09-12
|
2018-09-12 15:21:44
| 7
|
1809c5e8a146
| 4.910377
| 13
| 1
| 0
|
[When you] model the relationship of every word against all others, a strange and wonderful mathematical order emerges within our language.
| 5
|
Titles That Sell Versus Those That Don’t, a Quantitative Analysis
[When you] model the relationship of every word against all others, a strange and wonderful mathematical order emerges within our language.
Source: Lucas Amunategui
Let’s see if we can find some interesting differences between top and bottom selling book titles; we will use the “Amazon Sales Rank” system along with word-embedded vectors to help us get there.
Take a look at the “Title” column in the following two tables. The first table shows eight-word titles that sold well on Amazon books and the second table shows those that didn’t. See any patterns?
8-word titles that sold like hot cakes
8-word titles that didn’t fare so well
If you didn’t catch any obvious differences, don’t worry, it’s a bit subtle. But before we go there, what’s with the eight-word titles?
The Eight-Word Title, the Ultimate Title Length for Success
Turns out, after taking a 10,000 sample of top-ranking titles and averaging the number of words, we get: 8.53 words. When we do the same with the bottom ones, we get 6.07 words. Thus, a title between 8 and 9 words is the perfect length for success.
Using Cosine Distances to Understand a Title’s Signature
Now here comes the interesting part. If you average the cosine distance between each embedded-word vector (a lot more about what that means below), you get a higher distance on top titles than on bottom ones.
4.1 average cosine distance for top titles
3.9 average cosine distance for bottom titles
In simple terms, it means that titles made up of words that are less synonymous with each other tend to do better. And when you consider that the top titles have an average of two extra words in them, you start seeing the pattern that people prefer titles that convey more meaning (really?).
The Amazon Sales Rank and the Word-Embedded-Vector-Distance Relationship
We know that the first table lists hot book titles, their small Amazon sales rankings in the second column confirms that. The opposite goes for the second table, those are laggards and it’s confirmed by their very large Amazon sales rankings.
The last column of both tables is the ‘Word Distance Sum’, which is the sum of the cosine distance of each word-embedded vector in the title. OK, what the heck is a word-embedded vector?
Word-embedded vectors have been one of the biggest boons in the natural language processing (NLP) world in the last decade. The word-vector distance is a linear substructure of the word-vector space. When you take a very large corpus (think all of Wikipedia) and model the relationship of every word against all others, a strange and wonderful mathematical order emerges within our language.
Here is the famous example from Tomas Milkov et al and word2vec modeling technique that has captured the imagination of every NLP amateur out there:
Source: https://www.tensorflow.org/tutorials/representation/word2vec
It shows that the distance relationship between gender, tenses, etc. can be captured quantitatively. For example, once you have captured the word-embedded vector distance for a country to its capital, you can unlock the country/capital for all other countries using that same distance key (for more, see Vector Representations of Words). It doesn’t end on standard idiomatic categories, made up ones work just as well. Trask modeled the Harry Potter series and extracted all sorts of cool information using word-embedded vectors:
the word “password”… revealed a list of passwords to the Gryffendor Common Room. (Trask: https://iamtrask.github.io/2014/11/23/harry-potter/)
Word-Vector Similarity and Cosine Distances
Another important yield from using word-embedded vectors is to use a distance metric to extract similarities within text. The closer the distance (i.e. the smaller the word-vector distance), the more similar the words. This, in turn, can be used as a powerful, context-sensitive thesaurus engine. Here, I used the cosine similarity distance metric to measure the space between each word in the title.
[If] a title is too short and uses too many synonymous words, it will be harder to know what the heck the book is about…
And that’s the differing pattern you may have noticed in both tables above. The first table is sorted by word-vector distances in descending order, so it will show the titles with the richest, most complex meanings, where each word isn’t very synonymous with its surrounding words. The second table is sorted by word-vector distances in ascending order, thus the ones made up of close synonyms. If you look closely, you’ll notice that the words are much longer in the first table than in the second one.
Conclusion
In the digital world, a title is the equivalent of a traditional book cover; it has to work a lot harder to convey meaning. This is even a bigger deal in the era of stock-art images; they don’t necessarily yield clues about a book’s content so the title has to work that much harder.
This does make intuitive sense, if your title is made up of close synonyms and is too short, it may not have the necessary information to attract potential readers. The key takeaway here is if a title is too short and uses too many synonymous words, it will be harder to know what the heck the book is about…
And if you want to see how your own titles and content stack up against the pros, test it out on the experimental Multi-Point Writing Analyzer on ViralML.com.
Please share and clap if you found this helpful — thanks for reading!
Manuel Amunategui
Get it and plenty more at amunategui.github.io and at ViralML.com.
OK — Sign up to my email group below and I’ll send you my free eBook on tips to becoming a (better) data scientist (and signup even if you aren’t interested in the eBook). Thanks for reading!!
Source
R. He, J. McAuley. Modeling the visual evolution of fashion trends with one-class collaborative filtering. WWW, 2016
J. McAuley, C. Targett, J. Shi, A. van den Hengel. Image-based recommendations on styles and substitutes. SIGIR, 2015
This story is published in The Startup, Medium’s largest entrepreneurship publication followed by 355,974+ people.
Subscribe to receive our top stories here.
|
Titles That Sell Versus Those That Don’t, a Quantitative Analysis
| 187
|
titles-that-sell-versus-those-that-dont-a-quantitative-analysis-1809c5e8a146
|
2018-09-12
|
2018-09-12 15:21:44
|
https://medium.com/s/story/titles-that-sell-versus-those-that-dont-a-quantitative-analysis-1809c5e8a146
| false
| 1,023
|
Medium's largest publication for makers. Subscribe to receive our top stories here → https://goo.gl/zHcLJi
| null | null | null |
The Startup
| null |
swlh
|
STARTUP,TECH,ENTREPRENEURSHIP,DESIGN,LIFE
|
thestartup_
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Manuel Amunategui
|
Anything Applied Data Science. Author of Monetizing Machine Learning. VP SpringML. Barcelona.
|
5e29cc9fde82
|
amunategui
| 170
| 338
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-10-08
|
2017-10-08 17:56:42
|
2017-10-08
|
2017-10-08 18:08:30
| 11
| false
|
en
|
2017-10-08
|
2017-10-08 18:08:30
| 18
|
180b043be937
| 10.288679
| 42
| 0
| 0
|
In 2010, nearly every new billion dollar start up was created in the US or Europe. But, a graph showing the rise of billion dollar…
| 5
|
China and AI
In 2010, nearly every new billion dollar start up was created in the US or Europe. But, a graph showing the rise of billion dollar start-ups till then makes one thing clear — China is clearly in second place. And, as of today, one in every 3 billion dollar start-ups is created in China.
So, what has happened in China since 2010?
As many of you know, I like approaching the future with simple mental models. And, one mental model I use from time to time is waves of foundational technologies.
The interesting thing about these waves is that they apply more to the developed world than the developing world. Most developing economies didn’t really reap the benefits of PCs and the web. Instead, many jumped straight to mobile. The mobile wave is important because it made access to cutting edge technology incredibly cheap. And, China was definitely one of them.
I could spend a ton of time writing about how mobile usage patterns in China — thanks primarily to WeChat — are different from the West. But, you likely know that. I’ll just focus on 2 stats -
Nearly 20% of Chinese internet users are mobile only versus 5% in the United States
68% of Chinese users make mobile payments versus 15% in the United States
As a result, China has had a very strong and thriving mobile ecosystem. And, as technology waves go, that helps a lot as these waves often build on one another. This is especially the case with AI which, as we will see, is dependent on big company might.
Artificial Intelligence in China
Recently, a Goldman Sachs report on AI in China went viral. Here are a few highlights from the report.
First, the Chinese government aims to make China a world-leader in AI by 2030.
Geoffrey Hinton, Yann LeCun and Yoshua Bengio, the fathers of the Artificial Intelligence wave, may all be from Canada. But, China has been leading the way in research on AI.
Much of this progress is from Chinese companies — specifically, Baidu, Tencent, Alibaba and Didi.
China’s AI ecosystem is second only to the US. And, here’s a crazy stat — Baidu, Alibaba and Tencent account for 42% of the venture capital in China.
The Chinese giants spend comparable amount to the American giants (in this case, Google and Microsoft). And, they have a larger percentage of their workforce working on research and development.
So, we’re left with 3 takeaways
The Chinese government has decided the AI investment is important. And, we all have learnt that the Chinese government’s aims are to be taken seriously.
The top Chinese companies are investing heavily in research on AI and the results, so far, are positive
Finally, the top Chinese companies — BATD — are well positioned to compete with GAFAM for global AI supremacy. In case you’re wondering, BATD refers to Baidu, Alibaba, Tencent and Didi while GAFAM refers to Google, Apple, Facebook, Amazon and Microsoft.
AI — a one trick pony?
In an interesting article on MIT’s Technology Review (in the links below), James Somers makes a compelling point about the current AI wave being all about back propogation or backprop. Let’s start by revisiting what back propogration is with an example where the AI is tasked with identifying hot dogs -
Imagine you have millions of pictures — some with hot dogs and some without
You take the first image and it is the picture of a piano. Each pixel is mapped to a neuron in the bottom layer of the network.
When you first create this “neural net,” the weights between the layers of neutrons are random — random numbers that say how much “excitement” to pass along each connection.
As the excitement spreads up the network according to the connection strengths between neurons in adjacent layers, it’ll eventually end up in that last layer, the one with the two neurons that say whether there’s a hot dog in the picture
The final output neurons will either be a “no hot dog” or “hot dog.” But, if the output was wrong, the errors will propagate backward so the neurons learn from the error. And, backprop is a procedure for rejiggering the strength of every connection in the network so as to fix the error for a given training example.
He goes on to explain —
Backprop is remarkably simple, though it works best with huge amounts of data. That’s why big data is so important in AI — why Facebook and Google are so hungry for it, and why the Vector Institute decided to set up shop down the street from four of Canada’s largest hospitals and develop data partnerships with them.
Hinton’s breakthrough, in 1986, was to show that backpropagation could train a deep neural net, meaning one with more than two or three layers. But it took another 26 years before increasing computational power made good on the discovery. A 2012 paper by Hinton and two of his Toronto students showed that deep neural nets, trained using backpropagation, beat state-of-the-art systems in image recognition. “Deep learning” took off.
This is why it is critical for China that Baidu, Alibaba, Tencent and Didi are investing heavily in deep learning. The AI revolution is almost certainly going to be dominated by the large tech companies. And, it is vital China’s heavy weights lead the charge.
Why does AI matter to China?
Vladimir Putin recently said something most folks who read this might agree with (not a normal occurrence) — “Artificial intelligence is the future, not only for Russia, but for all humankind,” said Vladimir Putin “It comes with colossal opportunities, but also threats that are difficult to predict. Whoever becomes the leader in this sphere will become the ruler of the world.”
I suspect, however, that the benefits of AI to China are not just about “world domination.” Consider 2 examples -
China has installed 20 million cameras all over the country to create the world’s most advanced surveillance system. AI can help the government to identify people and help the police force search for criminals. Check this video on the Daily Mail’s website.
China is also building a social credit system for collecting information about all its citizens from all institutions. Imagine this to be a “citizen report card” that collects your credit rating, your tax filings, your job performance, your social media history, and so on. And, of course, imagine all this is linked to your face that can be tracked using the above surveillance system.
So. I suspect the Chinese government is looking to also benefit from AI’s advances as it puts together the infrastructure to create the most powerful surveillance state in history.
How does what China is doing matter to us?
MIT Professor Rodney Brooks, in a great essay called “The Seven Deadly Sins of Predicting The Future of AI,” cautions us from over reacting to predictions around AI (link below). His essay makes a singular point very well — Predicting the future is really hard, especially ahead of time. So, beware hysteria around AI taking over the world and everyone losing their jobs.
Again, I’d like to fall back to a mental model. I go back to the question — “What will AI make cheap?” In a previous note, I drew this pretty ugly, but hopefully, useful graphic.
AI can’t do everything that humans do. But, specific applications of AI makes it cheap for machines to do things we do, better. For example, deep learning systems have a better audio transcription error rate than humans.
Similarly, to build on the China example, if we take face recognition as one application, we can build on its consequences in the near term. The Economist, in a great article on this subject, wrote — “the ability to record, store and analyse images of faces cheaply, quickly and on a vast scale promises one day to bring about fundamental changes to notions of privacy, fairness and trust.”
Here are a few examples from the article -
FindFace, an app in Russia, compares snaps of strangers with pictures on VKontakte, a social network, and can identify people with a 70% accuracy rate.
Facebook’s bank of facial images cannot be scraped by others, but the Silicon Valley giant could obtain pictures of visitors to a car showroom, say, and later use facial recognition to serve them ads for cars.
Even if private firms are unable to join the dots between images and identity, the state often can. China’s government keeps a record of its citizens’ faces (as detailed above); photographs of half of America’s adult population are stored in databases that can be used by the FBI. Law-enforcement agencies now have a powerful weapon in their ability to track criminals, but at enormous potential cost to citizens’ privacy.
Employers can already act on their prejudices to deny people a job. But facial recognition could make such bias routine, enabling firms to filter all job applications for ethnicity and signs of intelligence and sexuality.
For example. researchers at Stanford University have demonstrated that, when shown pictures of one gay man, and one straight man, the algorithm could attribute their sexuality correctly 81% of the time. Humans managed only 61%. In countries where homosexuality is a crime, software which promises to infer sexuality from a face is an alarming prospect.
(Note: the researcher went on record to say this study was all about proving a point)
AI makes it cheaper and easier for even the most democratic of countries to switch on surveillance.
Why we should care about the debate around AI
When John Havens, IEEE’s Executive Director for Ethical Considerations in AI and Autonomous Systems, John was asked about the biggest challenges ahead in AI. He pointed to two things -
As users, we have no agency or access to their data that powers AI systems. The reason AI is the playing field of massive tech firms is because said firms have access to massive amounts of user data. This means we don’t get the chance to represent ourselves in an increasingly algorithmic world. It also means the playing field is stacked against us.
As countries, we follow an outdated economic model built around the concept of a GDP that is wholly focused on productivity without taking into consideration metrics such as environmental sustainability and human well being. If the incentives are all about maximizing shareholder value, replacing humans with robots is a natural next step.
It felt right to share a photo of Amazon’s warehouse workers supervising robots. That robot was likely a few human workers 20 years back. The full New York Times article shared below is an amazing read/watch (as it is full of great imagery).
People who point to the industrial revolution and say we’ll find different jobs forget that there was a long period of painful readjustment.
This was a workers union demonstration in New York City in 1914 — a hundred years into the industrial revolution. And, this was not the only such demonstration around the world back then.
I get it though. It is hard to make sense of the debate on AI. For most of us, we’re working hard at our jobs, then trying to put in a good shift at home and take care of our health along the way. Maybe, if we’re lucky, we get to have a hobby or two. On the side, we hear all this buzz about various billionaires fighting each other on the prospects of AI. Is it going to lead to humanity’s doom? Is it going to bring forth the utopia where we work on better kinds of jobs? Why should we care?
In a thought provoking essay on how to think about these futurist debates, Cathy O Neil makes a telling point (lightly edited) —
“For the average person there is no difference between the singularity as imagined by futurists and a world in which they are already consistently and secretly shunted to the “loser” side of each automated decision. For the average person, it doesn’t really matter if the decision to keep them in wage slavery is made by a super-intelligent AI or the not-so-intelligent Starbucks Scheduling System. The algorithms that already charge people with low FICO scores more for insurance, or send black people to prison for longer, or send more police to already over-policed neighborhoods, with facial recognition cameras at every corner — all of these look like old fashioned power to the person who is being judged.
Ultimately this is all about power and influence. The worst-case scenario is not a vindictive AI or Sergey Brin not getting to celebrate his two-hundredth birthday. In the worst-case scenario, e-capitalism continues to run its course with ever-enlarging tools at its disposal and not a skeptical member of the elite in sight.”
Well said.
Links for additional reading
McKinsey report on China — on McKinsey
Goldman Sachs report on AI in China — on, well, the internet
One in 3 unicorns is born in China — on Quartz
Is AI a one trick pony? — on MIT’s Technology Review
Putin on AI — on The Verge
Unprecedented reach of the surveilance state — on ChinaFile.com
7 Deadly sins predicting the future of AI — on Rodney Brooks’ blog
Life in the age of facial recognition — on The Economist
Stanford scientist’s Gaydar was to prove a point — on Quartz
Amazon workers babysitting robots — on New York Times (my title but awesome read)
5 questions with IEEE’s John Havens (AI Ethics) — on Medium
What the Industrial Revolution really tells us about the future of work — on ACM
Know Thy Futurist — on Boston Review
How Starbucks’ flexible spending breaks workers — on The Nation
How your credit scores are used by car insurance company — on Consumer Reports
The excellent Exponential View newsletter by Azeem Azhar (I got a bunch of these links from the newsletter over the past few weeks)
“Notes by Ada” note on AI
This is an edition of a bi-weekly technology newsletter called Notes by Ada. If you like this and would like free weekly notes via email, please just subscribe here.
|
China and AI
| 277
|
china-and-ai-180b043be937
|
2018-04-30
|
2018-04-30 14:58:56
|
https://medium.com/s/story/china-and-ai-180b043be937
| false
| 2,382
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Rohan Rajiv
|
Like to live at the intersection of people, technology, learning and impact. Blogged a learning every day since May 12, 2008 on http://www.ALearningaDay.com
|
50d7f8d72ef
|
alearningaday
| 3,194
| 37
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
6e338c920643
|
2018-08-28
|
2018-08-28 13:44:22
|
2018-08-29
|
2018-08-29 22:26:23
| 4
| false
|
en
|
2018-08-31
|
2018-08-31 14:47:28
| 8
|
180ba6a314f
| 2.254717
| 16
| 0
| 0
|
In conjunction with the SpaceNet dataset and challenge series that aims to democratize satellite imagery data and encourage the development…
| 5
|
Car Detection Over Large Areas With YOLT and Zanzibar Open Aerial Imagery
In conjunction with the SpaceNet dataset and challenge series that aims to democratize satellite imagery data and encourage the development of targeted algorithms, one of the goals of the CosmiQ team is to support the open source mapping community. To that end, CosmiQ is participating in the FOSS4G conference in Dar Es Salaam, Tanzania. FOSS4G is the largest annual global gathering of developers and users of open source geospatial software.
In this blog we explore how well one of the open source tools developed by CosmiQ (YOLT) performs on areas of interest to FOSS4G. Specifically, we analyze 7.5 cm resolution OpenAerialMap data collected via the Zanzibar Mapping Initiative over the island of Zanzibar just off the coast of Dar Es Salaam.
Model Training
As a test of the robustness of our algorithm, we apply a YOLT model trained on a very different dataset than the OpenAerialMap data over Zanzibar: 15 cm COWC aerial imagery collected over cities in Canada (Toronto), New Zealand (Selwyn) and Germany (Potsdam). Recall that YOLT is an object detection algorithm designed to rapidly localize objects over large areas in overhead images of arbitrary size. See our previous post for further training details.
Inference
The OpenAerialMap test dataset is collected with a different sensor, a different resolution (7.5 cm vs 15 cm for COWC), and over a very different geographic region than our training set. We select a detection threshold designed to minimize false negatives (cars that go undetected), and as a result there are a few false positives in the images. We unfortunately lack ground truth labels for cars in Zanzibar and so cannot compute rigorous performance metrics.
Performance and Conclusions
Inference runs rapidly at 9 square kilometers per minute on a single GPU. At this inference rate, running YOLT on an Amazon EC2 P3 instance for the entire ~1000 square miles of Zanzibar takes less than half an hour. Alternately, running on a lightweight CPU-only machine is possible (albeit far slower) and still runs at a rate of ~1 square kilometer per hour per CPU.
We look forward to further work to quantify performance, but for now the images below appear encouraging with very few missed cars (false negatives) and relatively few false positives. As we continue our algorithmic mapping efforts, future posts will also explore the application of building and road network detection algorithms trained on SpaceNet satellite data to aerial imagery datasets.
Figure 1. Zoom of detected cars (cyan boxes) in northwest Zanzibar.
Figure 2. 500 x 400 meter region of Zanzibar City.
Figure 3. Large image covering 1.3 square kilometers of Zanzibar City. There are ~2000 detected cars in this image.
|
Car Detection Over Large Areas With YOLT and Zanzibar Open Aerial Imagery
| 39
|
car-detection-over-large-areas-with-yolt-and-zanzibar-open-aerial-imagery-180ba6a314f
|
2018-08-31
|
2018-08-31 14:47:28
|
https://medium.com/s/story/car-detection-over-large-areas-with-yolt-and-zanzibar-open-aerial-imagery-180ba6a314f
| false
| 412
|
Welcome to the official blog of CosmiQ Works, an IQT Lab dedicated to exploring the rapid advances delivered by commercial aerospace startups and the open source community
| null | null | null |
The DownLinQ
|
Info@iqt.org
|
the-downlinq
|
SPACE STARTUPS,MACHINE LEARNING,COMPUTER VISION,DEEP LEARNING
|
cosmiqworks
|
Tanzania
|
tanzania
|
Tanzania
| 1,127
|
Adam Van Etten
| null |
a5a78eaee501
|
avanetten
| 654
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-23
|
2018-06-23 02:42:52
|
2018-06-23
|
2018-06-23 02:43:44
| 2
| false
|
en
|
2018-06-23
|
2018-06-23 02:43:44
| 0
|
180cf2bdca9d
| 2.639937
| 1
| 0
| 0
|
Digital assets quantitative investment has become a new trend of digital asset markets, it put aside the subjective risk preference…
| 1
|
Entrusted to quantify transactions, many investors chose BZcash
Digital assets quantitative investment has become a new trend of digital asset markets, it put aside the subjective risk preference, financial goals and objective risk to bear ability, use of big data analysis, the quantitative model and algorithm, according to the investor’s profit forecast and risk appetite, to provide matching portfolio strategy, and through the trading system of intelligent automatic close the deal, according to the market changes in process of dynamic adjustment strategy parameters, achieve the goal of assets value.
Under such a trend, more and more investors choose to entrust the way of quantitative trading, so as to make their assets steadily increase in value. Investors also from a purely “investors” to the role of “bole” transformation, the focus from the transaction itself to screening is also outstanding quantitative trading platform, realizing the share intelligence information institutions of way to trade.
First, Why do many investors choose to entrust quantitative trading on BZcash platform?
In order to enable ordinary investors to enjoy the convenience of quantitative intelligent trading technology, BZcash came into being. BZcash is a one-stop quantitative trading platform applied in the digital asset market. Investors can use the tools and strategies provided by BZcash to easily complete quantitative transactions, lower the threshold of using quantitative transactions, and help investors achieve stable investment returns.
BZcash quantitative and gain is the innovation of the global digital assets, with the global more than excellent quantitative information on the depth of cooperation, together outstanding firm offer absolute return of the world’s top digital assets quantitative strategy, to establish a quantitative model for the industry’s leading multi-dimensional screening measurement system, and build a perfect cast system, quantitative strategies help investors from the perspective of investment comprehensive and detailed analysis to better according to their own investment demand and risk preference selection to build quantitative portfolio strategy, and in the process of market changes timely dynamic adjustment strategy combination, realized in risk control and asset appreciation, Enjoy the most professional quantitative strategy investment financial services.
Second, the powerful AIMD algorithm will increase the return on digital assets up to 48.5%
A good quantitative trading platform, system must have a powerful algorithm, based on the trend of the economic meaning of clear judgment or arbitrage principle, further systematic and procedural abstraction, presented in the form of logic is a set of complete order process and logic control scheme of the executable.
AIMD (Artificial Intelligence, Machine Deep) algorithm, are among the most outstanding essence BZcash platform of computer mathematics algorithm, it assembled ai (Artificial Intelligence), Machine Learning, Machine Learning), Deep Learning (Deep Learning) three features, the use of powerful computer processor for big rules data operation, modeling, scene test, Learning and memory, etc. As an important decision the brain, quantitative trading session adopted in machine learning: linear regression, logistic regression, decision tree, vector machine (SVM), a bayesian model, regularization, model integration, such as neural network algorithm structure. The parameters can be adjusted, the big rule prediction data can be used for learning and growth, and finally the best policy model products can be adjusted for the use of a large number of platform users.
BZcash through AIMD algorithm, to the mathematical theory of clever blend into the actual combat of the investment transactions, and in the relatively faster time to gain market and perform a large number of orders, to achieve the common way to trade is difficult to obtain the profit space. It is predicted that AIMD will increase its digital assets income by up to 48.5% on the original basis.
|
Entrusted to quantify transactions, many investors chose BZcash
| 1
|
entrusted-to-quantify-transactions-many-investors-chose-bzcash-180cf2bdca9d
|
2018-06-23
|
2018-06-23 02:43:44
|
https://medium.com/s/story/entrusted-to-quantify-transactions-many-investors-chose-bzcash-180cf2bdca9d
| false
| 598
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Bz Cash
| null |
9e024f551a55
|
bzcash
| 4
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-09-01
|
2018-09-01 17:53:53
|
2018-09-01
|
2018-09-01 18:11:03
| 3
| false
|
en
|
2018-09-01
|
2018-09-01 18:11:03
| 10
|
180d7a51ea40
| 3.217925
| 2
| 0
| 0
|
It’s no secret that conversational user interfaces are becoming more common in the modern era. Several brands like Amazon, Google, and…
| 5
|
Why Are Consumers Adopting Voice Assistants in Their Lives?
It’s no secret that conversational user interfaces are becoming more common in the modern era. Several brands like Amazon, Google, and Apple are following this trend in order to provide a better user interface to their customers. In all kind of industries, voice interfaces are developing at an exponential rate. Ranging from banking to hospitality industry, everyone starts using voice assistants in a way to provide a personalized experience to their customers.
In a survey of more than 5000 consumers in the US, UK, France, and Germany found that voice assistants will turn to a powerful form of customer interaction over the next three years, with customers who utilize the technology ready to spend 500% more than they currently do through this form of communication. And, CIRP (Consumer Intelligence Research Partners, LLC) said that around 20 million Amazon Echo & other alexa voice devices and 7 million Google Home smart speakers have been purchased by U.S. Consumers, in the year 2017.
Even, voice assistant has transformed our relationship with phones from being a communication device to being enablers of interpreting some of our daily tasks. ‘Apple’s Siri’ and ‘Google Voice Assistant’ are now a part of our day to day lives. Today, around 24% of users would use a voice assistant rather than typing or touch on their phone screen. According to Comscore, around 50% of all searches will be done by voice without screens by 2020.
What users expect from voice assistants?
One of the major reasons why consumers are embracing Voice Assistants in their lives is Convenience. According to a report, it is clear that 52% of consumers say convenience is the major reason to adopt voice assistants over websites/mobile apps. And, other 48% say the ability to do things hands-free. Moreover, 41% of them also give importance to high-level of personalization.
Photo by Andres Urena on Unsplash
Every consumer wants their task done in less time without any hurdles. And, Multitasking properties are the next factor that will provide customers with better assistance, high-level personalization, in order to serve them better. It takes multiple orders and send it to POS, store customer details, provides suggestions like ‘meal of the day’, advertising and so on. Consumers can easily order their food or book the table without waiting on for long hold calls. Customer requirements can be easily fulfilled by providing them with better services and experience.
So, what should brands be doing now to get ready for voice self?
For every brand, providing voice assistance to the consumers not only leads to better customer satisfaction but also enhances the brand promotions. They must focus on the consumer needs and preferences to decide what different business functions can be performed through a voice assistance.
It is clear that customers are demanding retailers to give voice-based solutions to make their life comfortable. Within a couple of years, the voice assistant is going to be one of a major part of several industries, including hospitality. In restaurants, AI-based voice assistants are placing its feet due to its outstanding benefits such as personalized experience, better customer relationship, customer loyalty, and good customer services. And at the same time help restaurants to save running cost by focusing on their core business areas.
Photo by Nick Hillier on Unsplash
Final Thoughts
The main goal of every business is to build a stellar customer experience for their customers. This also holds true for the restaurant industry. Customer faces difficulty while placing an order over the phone due to wait time or server taking wrong orders, and having no visibility about their order status. These are all things that lead to the unhappy customers.
Voix helps restaurant owner to satisfy their customers’ every need leaving a good experience. It takes the multiple orders on time with zero hold time. It allows to work hands-free and reduces the need for qualified staff. It provides the flexibility in work by complete time-based customization. It also helps in building strong relationships and retaining the customers for a long time. Voix is the all in one solution to your restaurant phone ordering business and increases the efficiency of your operation. Voix AI voice assistant will make your restaurant ready for voice commerce wave.
|
Why Are Consumers Adopting Voice Assistants in Their Lives?
| 77
|
why-are-consumers-adopting-voice-assistants-in-their-lives-180d7a51ea40
|
2018-09-01
|
2018-09-01 18:11:03
|
https://medium.com/s/story/why-are-consumers-adopting-voice-assistants-in-their-lives-180d7a51ea40
| false
| 707
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Voix.ai
|
Voice Assistant for Restaurants
|
e0d4acb49800
|
voixai
| 14
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-31
|
2018-05-31 08:36:14
|
2018-05-31
|
2018-05-31 09:07:03
| 4
| false
|
en
|
2018-09-29
|
2018-09-29 15:30:49
| 1
|
180ddde892a2
| 2.318868
| 5
| 0
| 0
|
Introduction
| 5
|
Chainer Tutorial: Sentiment Analysis with Recursive Neural Network
Introduction
In this tutorial, we will use the Recursive Neural Network to analyze sentiment in sentences.
Sentiment analysis is one of the major tasks of Natural Language Processing (NLP), which identifies writers’ sentiments in sentences. When expressing sentiment, we basically uses labels whether it is positive or negative. For example, in the case of the dataset used this time, emotions are expressed in 5 classes like 1(really negative), 2(negative), 3(neutral), 4(positive), 5(really positive).
cited from [1]
Sentiment analysis is implemented with Recursive Neural Network. Recursive Neural Network is a recursive neural net with a tree structure. NLP often expresses sentences in a tree structure, Recursive Neural Network is often used in NLP. First, we explain the training method of Recursive Neural Network without mini-batch processing. After that, as a advanced story, we explain the training method of mini-batched Recursive Neural Network.
In this tutorial, we will understand the following:
What is Recursive Neural Network?
Implementation of sentiment analysis by Recursive Neural Network using Chainer
Training method of Recursive Neural Network without mini-batch
Training method of mini-batched Recursive Neural Network
What is Recursive Neural Network
Recursive Neural Network is one of Recurrent Neural Networks that extended to a tree structure. As both networks are often written as RNN, so we need to be careful which one we are expressing. In many cases, it refers to the Recurrent Neural Network in many cases, but in natural language processing it sometimes refers to the Recursive Neural Network.
Recursive Neural Network uses a tree structure with a fixed number of branches. In the case of a binary tree, the hidden state vector of the current node is computed from the hidden state vectors of the left and right child nodes, as follows:
This operation is sequentially calculated from the leaf nodes toward the root node. Recursive Neural Network is expected to express relationships between long-distance elements compared to Recurrent Neural Network, because the depth is enough with log2(T) if the element count is T.
Implementation of sentiment analysis by Recursive Neural Network using Chainer
As shown below, we explain the implementation with Colaboratory. If we have a browser, we can immediately run the tutorial in the GPU environment. So, please try it!
Sentiment Analisys with Recursive Neural Network - Chainer Colab Notebook 0.0 documentation
Here we set the parameters for training. * : Epoch number. How many times we pass through the whole training data. * …chainer-colab-notebook.readthedocs.io
Reference
[1] Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank
|
Chainer Tutorial: Sentiment Analysis with Recursive Neural Network
| 10
|
chainer-tutorial-sentiment-analysis-with-recursive-neural-network-180ddde892a2
|
2018-09-29
|
2018-09-29 15:30:49
|
https://medium.com/s/story/chainer-tutorial-sentiment-analysis-with-recursive-neural-network-180ddde892a2
| false
| 429
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Keisuke Umezawa
|
https://keisuke-umezawa.github.io/
|
1f561af8640d
|
keisukeumezawa
| 58
| 35
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-02-03
|
2018-02-03 12:50:10
|
2018-02-03
|
2018-02-03 12:53:35
| 0
| false
|
en
|
2018-02-03
|
2018-02-03 12:53:35
| 0
|
180e57e12382
| 0.792453
| 0
| 0
| 0
|
Li, T., Zhong, J., Liu, J., Wu, W., & Zhang, C. (2017). Ease. ml: Towards Multi-tenant Resource Sharing for Machine Learning Workloads…
| 3
|
Paper Summary: (VLDB’18) Ease.ml: Towards Multi-tenant Resource Sharing for Machine Learning Workloads
Li, T., Zhong, J., Liu, J., Wu, W., & Zhang, C. (2017). Ease. ml: Towards Multi-tenant Resource Sharing for Machine Learning Workloads. arXiv preprint arXiv:1708.07308.
This paper provides another option for multi-tenancy in machine learning system where users (or tenants) submit jobs by only providing the format of Input & Output (e.g. image size and number of classes to be selected from) and training data (e.g. pairs of labeled images), instead of real machine learning workload specifying which model to train. Then the system will automatically explore possible models (e.g. AlexNet, ResNet, GoogLeNet for image classification), allocate resources, schedule training workload, and return the output using the best model learned so far on the dataset when the user inputs a new object.
This design is motivated by the observation that at ETH Zurich, a powerful GPU cluster is provided to over 10 research groups for machine learning jobs but these non-computer-science users are bad at selecting models and may easily end up wasting resources on too complex models but overfitting. The evaluation shows that this system can be up to 9.8x faster in achieving the same global quality for all users than before.
|
Paper Summary: (VLDB’18) Ease.ml:
| 0
|
paper-summary-vldb18-ease-ml-180e57e12382
|
2018-05-18
|
2018-05-18 10:07:31
|
https://medium.com/s/story/paper-summary-vldb18-ease-ml-180e57e12382
| false
| 210
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Luke QZ WENG
| null |
6f0b097d1416
|
lukeqzweng
| 0
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
f3acf34c3db5
|
2017-10-16
|
2017-10-16 12:55:17
|
2017-11-29
|
2017-11-29 14:29:51
| 3
| false
|
en
|
2017-11-29
|
2017-11-29 14:29:51
| 2
|
180eb078aed8
| 3.587736
| 1
| 0
| 0
|
The science behind leadership practices recommendations
| 5
|
How A.I can helps us make better human leaders
The science behind leadership practices recommendations
Can Artificial Intelligence help us have awesome managers ?
Uber’s recent scandal over management culture highlighted the consequences of a toxic framework of leadership practices.
Considering the impacts of bad leadership practices on an organization, as a manager you must face a series of hard decisions every day…
How can we have meetings that are productive and uplifting, where we speak from our hearts and not from our egos? How can we make purpose central to everything we do, and avoid the cynicism that lofty- sounding mission statements often inspire? What we need is not merely some grand vision of a new type of organization; we need concrete answers to dozens of practical questions like these.
Managers often try to make these types of decisions in the most informed way possible. They ask people whose opinion they value, coaches included, but basically, that’s it. Most of us spend hours searching on the Internet in vain for answers.
Today, the challenges of management are too often solved in an unstructured way…
That’s why AI-powered algorithms are solution managers have been waiting for. Artificial Intelligence can be used to provide a science-based and data-driven analysis of challenging management issues.
It will provide leaders with an advanced ability to quickly analyze the context that may be critical to selecting the best management practices.
So, A.I. is not making the decision per se but offers you the most rational options according to science and your peers’ experiences. By aligning manager’s emotional intelligence and new leadership technology, leaders can reinvent organizations.
Build the perfect A.I for a fulfilling and productive workplace
From a mathematical point of view, management is a combinatorial explosion problem. A combinatorial explosion is the rapid growth of the complexity of a problem due to how the problem is affected by the input, constraints, and bounds. That’s why it’s so hard to analyze and make good decision.
How do you solve a combinatorial explosion? Artificial Intelligence can help us.
Solution #1: Decision-making tree model
The first solution is to reduce the equation to the most determinant factors. It helps find the best solution quickly, rather than using trial & error process that senior managers usually endured by the time they get good at managing.
It may sound complicated and unrealistic, but we have good news!
Thousand of researchers are already working on finding the best decision-making tree for management.
The bad news is, as a manager you don’t have time to read scientific papers.
So we did the work for you by reviewing scientific literature from the 60s to now. We were able to isolate four determining factors:
Psychological safety: Shared belief that the team is safe for interpersonal risk-taking.
Motivation: Factors that activate, direct, and sustain goal-directed behavior.
Team effectiveness: Capacity a team has to accomplish goals and objectives.
Creativity: Production of novel and useful ideas by an individual or small group of individuals working together.
Each factors can be divided into subcomponents. For example, motivation can be seen as autonomy + mastery + purpose + psychological and physical rewards.
A nice explanation is shared here by Daniel .H Pink http://bit.ly/2hMNVgm
The matrix is now composed of more than 13 sub-factors. It’s evolving quickly as we’re building an open scientific community with Research Facilities from around the globe.
A.I. helps build a direct connection between researchers and leaders around the world.
Solution #2: Prediction, Recommendation, Clustering and Search…
Collective Intelligence + A.I = Better Leaders ?
More can be done to ensure that every manager has access to the best recommendation system according to their specific context. Here are some approaches we’re already working on:
Collaborative Filtering Model
As a predictive approach, A.I will analyze leaders’ behavior. A.I will align your profile, your context, and practices you already used, and will be able to recommend best practices based on the experience of other managers.
Increase sensor range with 360° evaluation and Company Context Analysis
Currently the A.I. is processing management-related data from a questionnaire filled out by managers. Of course, mental biases are inevitable, and can lead to misjudgements. It’s why we are working on a 360° evaluation in order to collect complementary feedback from their teams. In a near future, A.I will be fed using public or private knowledge about company (growth rate, economic activity, etc.).
Improve Natural Language Processing
This will help to better classify practices from the open source library of management practices.
We need a new generation of leaders to change the world and A.I can help them.
We believe that Artificial Intelligence has the potential to revolutionize management models helping leaders build better organizations regardless its size, country’s culture or creation date.
And we are doing everything we can to make it happen ! #StoryContinues
|
How A.I can helps us make better human leaders
| 3
|
how-a-i-can-helps-us-make-better-human-leaders-180eb078aed8
|
2018-02-27
|
2018-02-27 19:45:59
|
https://medium.com/s/story/how-a-i-can-helps-us-make-better-human-leaders-180eb078aed8
| false
| 805
|
The story behind OpenDecide, a Leadership Support Platform helping managers with their team management
| null |
opendecide
| null |
OpenDecide
|
contact@opendecide.com
|
opendecide
|
WEB,HRTECH,MANAGEMENT,LEADERSHIP,ARTIFICIAL INTELLIGENCE
|
opendecide
|
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Guillaume TASTET
|
CEO @OpenDecide — A Leadership Support Platform to make work better
|
9bb7889a2ad2
|
guillaumetastet
| 18
| 52
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-04-18
|
2018-04-18 16:30:17
|
2018-04-18
|
2018-04-18 16:33:47
| 2
| false
|
en
|
2018-04-18
|
2018-04-18 16:34:07
| 3
|
180f99e758fb
| 4.372013
| 0
| 0
| 0
|
ASI’s Head of Research, Ilya Feige (left) with Hudson’s Associate Director of Data and Digital, Tim Clark (right).
| 5
|
Future thinking: The makings of a successful data scientist
ASI’s Head of Research, Ilya Feige (left) with Hudson’s Associate Director of Data and Digital, Tim Clark (right).
You can tell a lot about a business from a visit to their office. A few minutes’ walk from London’s hectic Oxford Street, hidden inside a converted Victorian terraced townhouse turned office building, ASI Data Science deliver one of the most-promising AI offerings to a host of big brands, including the likes of Easyjet, BBC and Tesco.
The building exudes cleverness; blackboards smeared in formulae, a ping-pong table that is actually used, an atmosphere that’s more akin to the friendly offices of a university’s science department than a corporate headquarters.
I was there to learn about the natural career path of a data scientist — something that ASI’s Head of Research, Ilya Feige, knows a lot about.
His first words to me: “I am not a data scientist.” A smile appears on his face, “I work mainly in machine learning and artificial intelligence on the research side, so I wouldn’t consider myself a data scientist as such.”
Like many of his colleagues, the Harvard doctoral graduate comes from a background in physics, and has been enticed over to the business world with high hopes of making a positive change.
“Here there’s opportunity to have a big impact on improving the lives of others.” He blushes as soon as he hears his statement out loud.
“I know, it sounds cheesy,” he admits. “But it’s true. Data science and machine learning automate the mundane, everyday tasks people don’t want to do, but this technology also has the potential to open up access to information and services that are currently out of reach for a lot of people.”
“Take access to quality and affordable healthcare as an example — this is one of the biggest problems facing developed economies. With AI technology, individuals could one day own their healthcare data, and algorithms could diagnose people’s health ailments if they are struggling to get advice from a doctor.”
He’s not alone in his thinking. In the last couple of years, data scientists have made substantial headway transferring medical facts and figures into complex algorithms; the day that AI technology becomes common practice in diagnosing medical ailments is right around the corner.
“Good language and image processing can turn unstructured healthcare data into structured data, which can in some cases be analysed by algorithms more accurately than by doctors. If this came into effect, I think healthcare could be ten times cheaper than it currently is, and better quality too.
“The same principle applies to legal services. When you need to hire a lawyer, most of the time it’s because they know a set of rules that you don’t — those rules should be stored in a machine that’s accessible to everyone.”
THE GROWING DEMAND FOR DATA SCIENTISTS.
Ilya and Tim discussing what it takes to have a career as a data scientist.
As opportunities rise for organisations to improve their offering through data science, the demand for data scientists is growing by the hour. According to Indeed.com’s EMEA Economist, Mariano Mamertino, the number of AI jobs in Britain has soared by 485% since 2014.
“Data science as an industry is extremely new — there’s nobody that’s been doing this for ten years,” said Ilya. “Data science as it is today only came into being over the past three or four years.”
After graduating from Harvard, the Canadian research scientist spent a couple of years at McKinsey & Company in San Francisco before making the move over to London.
At ASI, Ilya splits his time between being the technical lead of the data science consulting team and his own research, which ranges from modelling the semantic meaning in natural language processing through to training undirected neural network models.
“We do very little of the stuff we don’t like doing,” he admits.
MINDSET OVER MATTER.
Originally a fellowship programme, ASI Data Science has since branched out into a talent company of their own, providing organisations with in-depth insights and recent graduates with front-line industry experience.
I ask him what qualifications you need to get into a career as a data scientist.
“Ironically, science degrees are far more applicable to the type of data science we work on than a degree in data science or computer science,” he replies.
“As the industry is so new, and the specifics of a career in data science are still very hazy, when interviewing fellowship students I always make a point of asking: “What makes a good data scientist?” This is always a good indicator of what they care about and are interested in doing.”
At a time where AI experts and PHD students are being routinely courted by organisations on the same scale as Google, Facebook and Amazon, hundreds of applicants are still applying for ASI’s prestigious fellowship programme every year with the hopes of earning one of the 50 places on offer.
“To be an effective data scientist, you need the necessary technical capabilities: strong maths and statistics, a good grasp of the scientific method, decent coding skills (particularly Python, or SQL) — and some knowledge of machine learning techniques is always a plus.
“But aside from the technical, it’s equally important for a data scientist to have outstanding soft skills. A data scientist really adds value to an organisation by communicating complex ideas in an understandable, jargon-free way and to a varied audience. And data science is almost always collaborative — so you also have to be a team player.”
He continues: “And, perhaps most importantly, when you’re given a problem, you have to be able to own it. Rather than simply doing what you’ve been asked, you need to explore a range of different options and uncover all value, and test your model’s robustness. Candidates that can think with their business hat on — like an entrepreneur — will have the most impact.”
Data science has fast-become one of the most innovative, exciting and ground-breaking sectors on the jobs market, and as AI technology and machine learning become more commonplace in a business setting, so too will the demand for collaborative and agile-thinking data scientists.
The workplace is changing: find out more about the New World of Work here
|
Future thinking: The makings of a successful data scientist
| 0
|
future-thinking-the-makings-of-a-successful-data-scientist-180f99e758fb
|
2018-04-18
|
2018-04-18 16:34:08
|
https://medium.com/s/story/future-thinking-the-makings-of-a-successful-data-scientist-180f99e758fb
| false
| 1,057
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
Tim Clark
|
I place mid to senior level technology, digital and change professionals across a variety of sectors nationally.
|
a970650ba8ea
|
dylan.brown
| 1
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-11-27
|
2017-11-27 23:54:20
|
2017-11-28
|
2017-11-28 01:26:43
| 3
| false
|
en
|
2017-11-29
|
2017-11-29 02:56:44
| 3
|
18126a07c322
| 9.95
| 33
| 8
| 0
|
Why we need more than AI regulation and UBI. We need a new Karl Marx.
| 5
|
The Future of Work
Why we need more than AI regulation and UBI. We need a new Karl Marx.
I spend a lot of time worrying about what my future will be. No surprises there, I think most people not in the twilight of their careers do. And, trend-wise, it doesn’t look good. Wages have remained stagnant. Benefits and pensions are no longer guaranteed. Debt is increasing but the means to repay it are not. The gap between graduation and employment is increasing. And of course, part-time work and the gig economy are taking over traditional employment. Most striking to me is how steeply working hours has declined. You can clearly see a downward trend.
I think most of us are intimately aware how dire the work/economic situation has become. Even if “the economy’s never been better” we all, in the wake of the previous election cycle, have been treated to hundreds of whitelash articles about how economically depressed half the nation has become. Most of us are painfully aware of the guillotine of student loans hanging over our heads. And quite a few of us know what it’s like to live with our parents and “be on the job hunt” for more than an acceptable amount of time.
We know why this is the case. Artificial intelligence. Automation. Technology.
In spite of what many pundits would have you believe, immigration isn’t really the culprit. I mean, free trade deals suck, I’ll agree to that as an uninformed news-junkie, but honestly think about cleaning sewers. Do you want to do that? Do you think anyone, in their heart-of-hearts, day-in, day-out, wants to do that? Of course not. It doesn’t matter if you are black, white, Mexican, Jewish, Muslim, gay, straight, or whatever ethnicity or creed you may claim. Sucky work is just sucky work. It might be important, but that doesn’t mean I, specifically, want to do it.
So, I have three options. One, I roll up my sleeves and clean toilets out of the goodness of my heart. Two, I pay someone else to do it, or I do it but for a fee. Three, I create something to do it for me. Thus far, option two has been our preferred way of doing things and the basis of our economy. But times are a-changing. The increase in processing power of CPUs and commoditization of computers and computing power has made it so that option three, automation, is now more than mere science fiction.
And it’s not just robot arms in factories. Those have been there for years. Artificial intelligence, even the non-AGI variety can and is causing massive disruption middle-class and service roles. Calls centers and spreadsheets will die soon, if the AI revolutionists are to be believed. The usual line is that as free trade agreements took away blue collar, working class jobs AI will take away white-collar, middle class jobs.
What’s hilarious is everyone thinks they’re free and clear. That their job won’t be automated. Check these out:
Interactive: Charting the Automation Potential of U.S. Jobs
In last week's Chart of the Week, we noted that 1.3 million industrial robots would be installed between 2015 and 2018…www.visualcapitalist.com
Visualizing the Jobs Lost to Automation
The employment landscape of the future will look very different than it does today. While we've charted the automation…www.visualcapitalist.com
How Machines Destroy and Create Jobs
"There's just doesn't seem to be many blacksmith jobs these days." At first glance, this would be a ridiculous thing to…www.visualcapitalist.com
I disagree with these on one point — engineering will totally be automated. Engineers are creating the tools of automation and the first thing they’ll do, by the logic I just described, is build something to do it for them.
“I choose a lazy person to do a hard job. Because a lazy person will find an easy way to do it.”
-Bill Gates
It’s easy to look at this human tendency and say human beings are lazy or unmotivated. But in my opinion, that’s not true at all. “Lazy” is just a word we use to shame those who don’t do what we want them to. No one just wants to sit around all day. Have you ever had a desk job or an office job where you had nothing to do? It’s cool for the first week or so, watching Youtube videos or derping around on the internet, but it gets old really quick. After a while you’ll run around asking anybody if you can help them in anything. You might even be told to stop asking for work, because it makes others look bad.
Humans aren’t lazy, we just want to maximize our time doing fun or interesting things and minimize our time doing mundane, boring things. We also, in general, would rather do things for ourselves rather than others. There is an altruistic impulse, but between filling out your spreadsheet and hanging out with my friends, I’d rather hang with my friends. Thus, automation I would say is the optimal choice. But choosing option three over option two fundamentally undermines our entire economic system.
The real heart of the issue is, we are totally unprepared to deal with automation. That’s why there’s such an interest in universal basic income (UBI) and calls for AI regulation. I’m not an expert in either, but both do nothing to address the real source of the problem with automation, AI, and the future of work.
That central question is, is our economic system even valid anymore?
Last time I can think of when someone asked this question was 1847, the year the infamous Communist Manifesto was written. I’m not endorsing, nor denigrating, Marxist thought. I’m pointing out though, that the very act of writing such an expansive critique of the then emerging status quo was because of eerily similar circumstance. A rapidly evolving world, quickly shifting away from feudal and agrarian societies to something that was then new an uncertain — a life of factory work.
We are similarly transitioning from labor based positions to mental work. More data analysts and business intelligence professionals than trade laborers. This varies, profession to profession and location to location, but again, on the average the trend is apparent. Marxist and Neo-Marxist thought was centered around labor being exploited and workers being alienated and not seeing themselves in their work. Most of the arguments that made Marxism valid were based on the existence of a labor force. What thought do we lean on when labor itself is in decline? What criticisms and critiques arise now that we shift from factories to… where, exactly?
We are still in transition, so it’s almost impossible to discern where we are going. Imagine a peasant farmer from some agrarian village. How much sense do you think it would make to them if you explained that we use little plastic cards to exchange pieces of paper to get food? That we don’t really know where our food comes from, but it’s usually collected in one location and often it’s prepared so we don’t really need to do much, other than put it in our mouths and chew? It would sound so bizarre to the agrarian peasant and trying to explain the workings of our current economic system would seem so alien that it probably wouldn’t even register as all that similar to the caravans, and the harvesting, and day long journeys to villages and towns, that constitutes as their economic system.
That is my point about the word “laziness”. We are so entrenched in our economic system that we don’t even recognize how much of our values and language and art is a direct result of our economics. In a new, post-automation system, laziness would probably cease to be a concept. How can you be lazy if the entire point is to get a robot to do it for you? A new thought would mean a profound shift in the values in society and the language we use to describe those values.
We see an attempt at this shift already. The sharing economy. The gig economy. But these are simply tech-enabled service jobs. Uber is basically an app-assisted taxi service. Airbnb is an app assisted timeshare. And they exist inside the existing capitalist system. Neither of these proposed economies discuss automation and AI, except as a worry.
So what would a new thought look like?
I wouldn’t know, but there are a few kernels to look at that I think will play a central role in whatever we envision a new economic system to look like.
The first would be to start where Marx started. What is value? Apparently, central to Marxist thought is the Labor Theory of Value. Marx was originally attempting to address what is known as the Diamond-Water Paradox, which is are diamonds valued higher than water? Diamonds are not as critical to survival as water is, so why are diamonds valued so much more highly? Marx believed that it was because diamonds must be processed and water does not, implying that what we really value is the work put into making the diamond look like a diamond.
Similarly, the paradox I propose a value system from is the Locksmith paradox (there is a variation called the service recovery paradox). Dan Ariely describes a locksmith that, while an apprentice, had a hard time at his craft and worked longer at what would take minutes. As an apprentice, his customers were generally happy with him and often tipped, even though he would make mistakes and sometimes have to redo the work. As he became more experienced though, the quality of his work drastically increased. But then his customers started to complain. Although his work was of the highest quality, they would often complain about the price.
At first blush, this seems to be explained by labor theory. The longer he worked, the more he was compensated because we value the work, right? But this doesn’t make that much sense because the work was of a low quality. Think of the more experienced locksmith, who makes a higher quality lock in less time. It doesn’t seem fair to say that the higher quality, produced in less time is worth less than the lower quality lock, right?
I propose a social theory of value. The apprentice is tipped because his eagerness to please and his willingness to do the job over again until it meets your satisfaction is a socially desirable trait. If you get a really good waitress, you will probably tip her more than if you got a really miserable server. The more experience locksmith, while offering a more efficient service, doesn’t offer the opportunity to socialize at all. He does his job, but there is no way to judge whether we think he deserves his wage or whether he has any socially desirable traits.
But how does this address the future of work? Well, think about it. If you automate something you necessarily take out the social aspect of it. Automating labor means that there is no need for people to go out and build and make things together. There is no opportunity to socialize. Similarly, consider automating capital markets, as is happening with algorithmic trading. Doesn’t it seem stupid, bankers waiting around their Skinner boxes just waiting for numbers to go up or down? Because without human engagement and socialization, markets are pointless. If the value of markets is, as Hayek stated, in information exchange, we are in an open-source era where there are far more efficient ways to exchange supply-demand info on commodities than the ornate dance of trading.
That’s partially why I don’t think UBI or AI regulation will address the crisis of automation. UBI, to begin with, redistributes a massive cash influx to all citizens of a country. But where does this money come from? What makes money valuable? Money is debt. The legal tender in your wallet is an I.O.U. from the country and is guaranteed by the government. But does it make sense to owe a debt to an uncaring robot arm making widgets? Probably not.
And there is the issue of generational inheritance. We might feel entitled to the products of a fully automated economy because we created a fully automated economy, but do our children deserve the spoils? Would our great grandchildren even understand why they’re looking at a screen watching numbers go up and down? Do we even understand right now?
Well, what about regulating these emerging forces? There are a few issues here as well. For one, the means of producing AI, computers, are available to everyone. It’s the same fear with 3D printing guns. Right now, at the moment, it’s very specialized knowledge, but as the technology develops there will be more people who know how to create this disrupting force. And while you might be able to regulate the technology, you cannot regulate the incentive. If a general AI was a reality, our economy would collapse overnight, because why would any changemaker pay someone to think up new ways to build and expand business when they can do it for free? And again, if this did happen, would money even mean anything without social necessity driving it?
The social theory of value also points to a reality we can already see emerging. That careers where socialization is extremely important, like healthcare and education and counseling, are relatively automation-proof. The service industry, essentially, which is the only growing part of our economy in a tangible sense.
Employment in services is growing. Via the Economist.
Another example might be the backlash against globalization. If value is a socializing force, it can condition the average person to behave according to a specific set of norms. What we might call professionalism. Professionalism is necessary because in order to engage the most customers you can, you must be non-offensive to as many people as possible. But often this feels like wearing a mask, behaving in a distilled way from how you would normally behave. The increasing ubiquity of social media, which also splits personalities, of constant surveillance, and of a growing global market would mean that you must constantly behave in this “professional” manner. For short amounts of time, this is ok, but to demand a wholesale change in personality is to provoke retaliation.
I don’t know where else to take this thought or how useful it might be in establishing a new framework for thinking about the future of work, but I do feel it’s more relevant than “Let’s just give everyone money” or “the government (or someone else, but not me) should do something about AI”. I think it’s not as intuitive, and the connections are not obvious (maybe even muddier after reading this), but we need to start from basics and build up if we want to address our rapidly changing future.
|
The Future of Work
| 196
|
the-future-of-work-18126a07c322
|
2018-05-10
|
2018-05-10 19:49:36
|
https://medium.com/s/story/the-future-of-work-18126a07c322
| false
| 2,491
| null | null | null | null | null | null | null | null | null |
Future Of Work
|
future-of-work
|
Future Of Work
| 8,540
|
Totem And Token
|
Craving Discourse
|
1633089f013e
|
TotemAndToken
| 90
| 209
| 20,181,104
| null | null | null | null | null | null |
0
|
myAwesomeSeq2SeqModel =
map linear(128)
.> map relu
.> map linear(194)
.> map relu
.> map linear(256)
.> map relu
.> encoderLstm(256)
.> decoderLstm(256)
.> map linear(64)
.> map relu
.> map linear(10)
.> map softmax
fullyConnected(numUnits, activation=relu) =
linear(numUnits) .> activation
encode =
map fullyConnected(128)
.> map fullyConnected(194)
.> map fullyConnected(256)
.> encoderLstm(256)
decode =
decoderLstm(256)
.> map fullyConnected(64)
.> map fullyConnected(10, activation=softmax)
myAwesomeSeq2SeqModel = encode .> decode
encode =
map (
fullyConnected(128)
.> fullyConnected(194)
.> fullyConnected(256)
)
.> encoderLstm(256)
decode =
decoderLstm(256)
.> map (
fullyConnected(64)
.> fullyConnected(10, activation=softmax)
)
| 7
| null |
2018-01-14
|
2018-01-14 04:44:01
|
2018-01-15
|
2018-01-15 00:11:15
| 1
| false
|
en
|
2018-01-15
|
2018-01-15 00:11:15
| 4
|
1812ee480ff1
| 4.845283
| 9
| 1
| 0
|
Differentiable programming languages let you define programs with data. I think it’s a clear next step for machine learning. Through…
| 5
|
Designing a Differentiable Language for Deep Learning
Differentiable programming languages let you define programs with data. I think it’s a clear next step for machine learning. Through expression and simplicity, they’ll likely open the door for creative and complex new deep learning models.
An Example
I might have a bunch of sequence-to-sequence training data, and I want to build a program fit to it. Sequence-to-sequence models generally have a signature like[a] -> [b], so maybe we can define the model like this:
I’m choosing a modified Haskell syntax for our language; we’ll get into why later. I’m using forward function composition (.>) a lot here; it wires the output of the previous function into the next function. I’m also using map a lot; this just applies the same function to every element in a list, returning the new list. Some functions like linear take parameter arguments (the number of units in this case).
The code is fine, but we could break it up into functions:
We can map a composed function instead of calling map over and over:
Why Differentiable Programming?
Using a deep learning framework like TensorFlow requires users to create a graph of symbolic tensors connected by operations like layers. It feels weird because you’re not writing the program, you’re writing a program that constructs a program (you write the Python code that constructs the computation graph, which is then interpreted by TensorFlow).
In differentiable programming, computation graphs are the implicit substrate of the language. You could read a lot of differentiable code before even realizing it’s differentiable.
Because the graphs are implicit, differentiable languages are much more expressive at making complex models. First-class lists are used for variable-length data. First-class conditionals make control-flow easy. Likewise, deep learning models, like stacked convolutional layers, are first-class functions. Any function in a differentiable language a model, because you necessarily can run backprop through any of them.
Differentiable programming can make transfer learning much easier. If I want the first few layers of a model to be the same as the first few layers in a different model, I can extract those layers into a simple function and train the models one-at-a-time or simultaneously. Reusable functions are fundamental to software engineering. Models are just as easy to reuse as functions in differentiable programming because models are functions in these languages. Imagine what can be accomplished with this expressive power.
What is Differentiable Programming?
Differentiable programming languages let you define programs and evaluate their results and derivatives. To find the derivative of a program, we need to differentiate every expression in it. If we know how to differentiate every expression, then we can use the chain rule to connect them all.
Because the programs (a.k.a. models) are differentiable, you can run optimize them with backprop, exactly like TensorFlow or other deep learning frameworks!
You could call TensorFlow the most popular differentiable language of all. Every operation in TensorFlow has a forward and backward implementation, and it uses the chain rule to connect them all. Unfortunately, it’s not so expressive because of it’s second-class embedded nature within Python.
Driving the Design with Limitations
Here are some requirements for our language that come pretty naturally from our domain:
Functional
Every expression must be differentiable, so every expression must be a “continuous function”. To the programmer, this requires purely functional programming. Mutations, assignments, loops, and print statements aren’t differentiable because they aren’t functions.
The lambda calculus is the basis for functional programming. It’s a rigorous way to write programs.
Aside: maybe an imperative language could be interpreted with a differentiable language, making the imperative language differentiable. It could be a lot like writing an imperative language in a functional one. I haven’t thought too much about this, but it’s a possibility AFAIK.
Constant Model Parameters
Model parameters, like the number of units in a layer, must be known at compile-time. As an example, it’s really hard to optimize a neural network when the number of units and layers are being optimized simultaneously. We definitely want to keep parameters constant throughout training. This doesn’t forbid us from “meta-optimizing” network parameters, but it does require us to make these meta-optimizations part of a meta-program (hmm… maybe Lisp would be a better choice of syntax). Also, TensorFlow doesn’t support it, so I don’t want to. Of course it’s possible to optimize model parameters; the compile-time requirement is just pragmatic.
Borrowed Building Blocks
When implementing the language, it would save a lot of time to build on top of an established differentiable computing framework, like TensorFlow. Not only does it require a lot of time and skill to create a computation engine like TensorFlow, it would be similarly taxing to build all the GPU-optimized operations necessary for our language to be useful. We could borrow all the operations, from convolutions to conditionals, and port them into our language.
Compilation Target
Naively, our compilation target could be strings of Python code that construct TensorFlow models, which could then be used by running eval in a Python program. A more robust solution would be to use the low-level API of TensorFlow. Yet another option would be to embed the language within Clojure so we can use meta-programming to interact with Deeplearning4j or Cortex, deep learning frameworks that run on the JVM.
Syntax
Haskell-Inspired
Haskell is a typed purely functional language, which makes it a great starting point for our differentiable language. There’s conveniently a lot of support for working with functions and no primitives for assignment/mutation. We can make almost all of our syntax Haskell syntax.
Compile-Time Parameters
For example, linear has the parameter numUnits, which is known at compile-time. I might instantiate a linear model with linear(10), satisfying the compile-time requirement of model parameters. We can still assign these to constants, as long as we can tell what they’ll evaluate to at compile-tim.
Default Model Parameters
Default parameters are very useful. Most deep learning frameworks benefit a lot from using sane defaults while allowing users to deeply customize parameters when needed. For example: fullyConnected(numUnits, activation=relu). In more general programming, like composing functions, my opinion is that default arguments aren’t as necessary. Haskell for example doesn’t have default arguments, but it also doesn’t suffer for it.
The Future of this Language
I’m unsure if I’ll implement a differentiable programming language, but I do know that many people are thinking about it and trying.
Check out TensorLang as one example of an early-stage differentiable language.
Also check out Chris Olah’s blog around functional programming and neural networks. for more inspiration.
I’m Job Searching! January 2018
I’m on the job hunt! I have broad experience in machine learning, functional programming, music, and entrepreneurship. I’m looking for a place to apply my skills preferably in Los Angeles or other non-bay-area locations.
|
Designing a Differentiable Language for Deep Learning
| 20
|
designing-a-differentiable-language-for-deep-learning-1812ee480ff1
|
2018-06-12
|
2018-06-12 19:42:31
|
https://medium.com/s/story/designing-a-differentiable-language-for-deep-learning-1812ee480ff1
| false
| 1,231
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Max Bendick
| null |
631c55a202f4
|
maxbendick
| 43
| 5
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-18
|
2018-06-18 15:42:48
|
2018-05-23
|
2018-05-23 17:42:11
| 1
| false
|
en
|
2018-06-18
|
2018-06-18 15:43:51
| 13
|
1814aaf7cb4d
| 4.641509
| 0
| 0
| 0
|
The race to deliver last-mile goods to customers robotically is being accelerated by Amazon. The most recent patent awarded to Amazon gives…
| 5
|
You Can’t Automate Empathy: The Future of Labor in the Nonprofit Sector
The race to deliver last-mile goods to customers robotically is being accelerated by Amazon. The most recent patent awarded to Amazon gives them IP on the idea of “mobile bases” that can be used to robotically deliver packages through various means. As more of these delivery bots take to the skies and sidewalks, the oncoming wave of automation and job obsolescence will increase in visibility.
An oncoming wave of automation.
While robots on wheels may be on pace to replace many package and food delivery jobs, one thing they won’t be threatening are groups like Meals on Wheels. More important than delivering food, programs like these are delivering human contact that helps connect homebound seniors with the outside world. This is the kind of work that depends on social intelligence, and it won’t easily be replaced by a robot.
So even as experts predict half the US workforce could find their jobs wiped out in coming decades, the nonprofit world faces a great opportunity: The jobs that are the safest from automation rely on creativity, social intelligence, perception, and manipulation. For at least the next two decades, machines will continue to struggle in matching the human brain’s ability to understand and empathize with others.
This means that, in the coming decade, America’s nonprofits will go from a nice addition to a necessary player in our economic and labor success.
That nonprofits are destined to flourish in the age of AI is by no means a passive, pollyannaish statement. Rather, our inability to automate empathy is a real wakeup call for those industry leaders and policymakers who overlook the nonprofit world’s role in our economy. And nonprofits and foundations face an important challenge now: They can start using the power they will increasingly have as the nation’s biggest employers to push adoption of an Empathy Economy.
The Empathy Economy is not just about the jobs nonprofit workers hold, but also the compassion we all show to one another in public policies that find ways for people who lose their jobs to automation to transition smoothly to new approaches, be it training for new types of jobs, a guaranteed basic income, a new wave of public-service jobs that pay people who volunteer to help others, or other efforts that ensure every American can pay for the necessities of life and find a sense of purpose and community.
Many nonprofits are already working on ways to ease the transitions in the workforce, exerting job creation influence on for-profit sector through incentives and adult education programs. Organizations like Goodwill, which has placed over 313,000 people into employment as of 2016, set an example for how nonprofits can move the needle outside of their sectors. Local groups like Opportunities for a Better Tomorrow* work deeply in the New York City community to create adult education and career building for thousands of disadvantaged people. There are also operations like the Opportunity Fund and Fund Good Jobs*, which seek to support families and businesses whose hiring practices meet both living wages and living standards. By offering financial incentives like loans and management trainings to provide better working-wage jobs, such organizations are impacting the way companies do business.
While we may still be a couple decades away from any significant job elimination from automation, we do know that this is an eventuality — not a possibility. Forecasts for elimination rate vary greatly, yet whether the eventual number of jobs lost is 38% (according to PricewaterhouseCoopers) or 47% (per the The Future of Employment), that total won’t be 0, as Treasury Secretary Steve Mnuchin would have us believe.
For jobseekers, hope lies in an industry that has made hope its currency — a currency earned through action and empathy. The nonprofit world is built on our service to each other and the world around us. The work being done is real and depends on compassion, insight, and a community of interests.
That doesn’t mean nonprofits aren’t using technology to transform their work — and they certainly will continue to do so. Case in point: New York Common Pantry, a nonprofit that helps feed over 300,000 annual visitors, transitioned to a digital system nearly a decade ago. Prior to the transition, volunteers made the same package of items for everyone based on family size. Now, 10% of the organization’s customers order online while the other 90% use on-site tablets. The importance of volunteers, however, hasn’t changed. They are still needed to support and connect with visitors as they interact with the tablets and wait to be served. The tablets and tech didn’t put people out of work — in fact, they enhanced the quality of service and helped the organization to grow.
The resilience of the nonprofit labor market has been clear for more than a decade. A 2012 study from Johns Hopkins University* notes that, through the Great Recession, nonprofits gained jobs at an average of 1.9% per year while the private sector lost jobs at 3.7%.
That’s in part because more people were in need — and they turned to nonprofits for help.
More recently, the 2017 Nonprofit Employment Practice Survey showed that 50% of nonprofits surveyed were planning to create jobs in the coming year, well above 40% of for-profits planning to hire. Over the past few years, the leading nonprofit job site Idealist.org*, has seen consistent year-over-year increases in nonprofit job postings. They are currently projecting continued growth going into 2018 based on current Q1 growth of around 10% vs this time last year.
It’s not just automation driving growth in the nonprofit workforce, it’s also demographics as more Baby Boomers move into retirement, more Millennials level up in their careers, and the first members of Gen Z enter the workforce. More than half of nonprofit jobs come from the healthcare and social services industry because those groups rely on millions and millions of people to provide aid — and they will soon need millions and millions more.
In light of this, it’s time for policy makers to lean in: rather than penalizing the nonprofit sector with tax code updates, our government should be looking to build labor-friendly policies. Right now, despite being tax exempt, nonprofits still pay payroll tax for their employees at the same rate that companies like Amazon and Apple do. Removing or reducing this tax burden on the sector’s employees is one place to start to support job growth needed in the coming years.
Meanwhile, the next time you see yet another article about the impending unemployment wave coming at the hands of automation, consider that with this wave will come a greater need for the human touch. Rather than fight the wave of AI and risk being swept away by the undertow, let’s also invest and promote the Empathy Economy.
*Idealist.org, Fund Good Jobs, Opportunities for a Better Tomorrow, and the Johns Hopkins School of Public Health are all former clients of Whole Whale.
Originally published at medium.com on May 23, 2018.
|
You Can’t Automate Empathy: The Future of Labor in the Nonprofit Sector
| 0
|
you-cant-automate-empathy-the-future-of-labor-in-the-nonprofit-sector-1814aaf7cb4d
|
2018-06-18
|
2018-06-18 15:43:52
|
https://medium.com/s/story/you-cant-automate-empathy-the-future-of-labor-in-the-nonprofit-sector-1814aaf7cb4d
| false
| 1,177
| null | null | null | null | null | null | null | null | null |
Nonprofit
|
nonprofit
|
Nonprofit
| 15,254
|
Whole Whale
|
Digital agency that leverages data and tech to increase the impact of nonprofits and for-benefit companies. Read more at https://medium.com/whole-whale.
|
c7eac8c0788
|
whole_whale
| 8
| 156
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-12-04
|
2017-12-04 01:05:09
|
2018-01-13
|
2018-01-13 04:56:11
| 1
| false
|
en
|
2018-01-13
|
2018-01-13 04:56:11
| 1
|
1814fb849db2
| 3.596226
| 0
| 0
| 0
|
Conversation is a two way process, so improving conversation involves both how we interpret what we hear and how well we communicate back…
| 4
|
How to make machines talk like humans?
Conversation is a two way process, so improving conversation involves both how we interpret what we hear and how well we communicate back. My previous post talks about how to improve NLU to get better at interpreting user utterances. In this post I want to highlight ways to constructively engage users of a conversational agent with meaningful responses. In the world of conversational agents, this is what is called Natural Language Generation.
I would like to summarize three key principles that have worked well for me when designing and evaluating conversational agents for it’s NLG:
Always use natural language
Being natural doesn’t mean being anthropomorphic from the word go, it’s more about not being ‘machine like’. The responses shouldn’t feel like someone is reading a ‘data structure’ or a ‘GUI dialog box’ to you ;)
A good approach to being ‘natural’ is to simulate a real world conversation for your use cases — like talk to a person on the phone or over text for the same task or information as your use case. Using real conversations to guide user workflow would help address the nuances of a real conversation like maintaining context, resolving co-references etc. Of course, it’s best to simulate conversations with your target user base like teenagers, college grads, busy parents, lawyers, etc.
Machine-like NLG has other downsides. While interacting with a conversational agent, especially if the agent NLU is not good enough, users will quickly adapt and start mirroring the agent’s language (in the hope of being understood). Hence, besides using more human-like language, we also need to ensure the examples and patterns we use in our NLG work with our NLU models as users may say ‘your phrases’ back to the agent. If the agent writes ‘Movie is at 7:00pm’, users will also subconsciously start using ‘HH:MM<am|pm>’ type of phrases including ‘0700pm’.
Learn from existing conversational interfaces
In GUI applications, the graphical elements like menu, buttons, icons, etc. are like an instruction manual — giving a good overview of the supported (and unsupported) features. Instead of matching that overview expectation in a conversational agent, model it on a real world interaction.
We interact with human agents for all sorts of tasks like book movie tickets, book a restaurant, book flight, etc. Structure the AI agent’s dialog on how such a human agent interacts: How she introduces herself? How she clarifies any incomplete or incorrect information? How she uses your name while addressing you? How she repeats your info if she needs confirmation? These agents were also filling structured forms in the background taking inputs from a customer through a conversation.
Instead of focusing on smalltalk make your core use case interaction more human-like — that is what engaged and retained users will experience more often. It includes giving meaningful responses for unsupported intents, “Sorry I can’t help with cancelling or refunding tickets yet”. It also includes assisting with feature discovery. Human agents are always on the lookout for opportunities to tell users of other (preferably related) stuff the agent can do. It’s good to piggyback related task details on other task completion responses, “Your movie tickets are booked. Would you also like to book Uber for the show?”
Keep communication clear and consistent
Unlike GUI, where we have web pages and app screens to set context, an agent’s entire interaction with a user is a single conversation thread. This can make the conversational interface messy. It is very important to clearly set context — especially start, progress, and completion of a task or goal. Similar to human agents’ behavior, including context info in agent responses can help convey which task is in progress and what’s the status. If your agent can’t handle multiple nested contexts, there is no harm in pushing users to complete the current task.
Include plenty of context info and examples in error messages and clarification messages to help user finish his task, like instead of responding “I didn’t get that”, say “I didn’t get ‘10am in the night’. Can you clarify the time again?”
For a conversational agent, consistency is a lot more important than a human-like personality. Consistency in user flow and clarity of language would make the interaction more predictable for repeat users and help them accomplish their tasks faster. For example if all responses start with ‘Great’, ‘Awesome’, ‘Sure’, user won’t understand the intent till he has read the entire response. Additionally, if your workflow asks for information is different order, user needs to pay attention each time he interacts with the agent.
Conclusion
For any challenges that your current NLU components are facing, your NLG components have opportunities to solve that problem by guiding users through the flow. Good NLG is key to a successful conversation.
Just like any other UX design, usability tests are very important for each user flow. Conduct usability tests to understand: Are new users getting enough guidance to get started? Are users able to start and end tasks? Are clarification prompts clear? Are examples in responses good enough? Is conversational interface making your use case better than GUI? Similar to GUI design, this mean iterations on NLG with a goal of improving UX.
I would love to hear your thoughts on this topic, do leave your comments below or PM me if you would like to chat on this topic.
|
How to make machines talk like humans?
| 0
|
how-to-make-machines-talk-like-humans-1814fb849db2
|
2018-01-13
|
2018-01-13 04:56:12
|
https://medium.com/s/story/how-to-make-machines-talk-like-humans-1814fb849db2
| false
| 900
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Manisha Pankey
|
Product Manager @Yahoo. Interested in Conversational Agents, Consumer Products, Travel, Effects of Tech on Kids.
|
a4976fbcc5bf
|
manishapankey
| 30
| 15
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-06
|
2018-06-06 17:54:40
|
2018-06-06
|
2018-06-06 17:57:22
| 1
| false
|
en
|
2018-06-06
|
2018-06-06 17:57:22
| 4
|
1816bbc65d1b
| 3.516981
| 1
| 0
| 0
|
Analytics and machine learning are all about data. The quality and quantity of your data plays a critical role in determining the…
| 5
|
Improve Machine Learning Effectiveness with a Feature Catalogue
Analytics and machine learning are all about data. The quality and quantity of your data plays a critical role in determining the effectiveness of the models. But even with reams of the best data, if your process doesn’t use it properly, your results will be dubious. All stages of the process must not only use the exact same data but the same interpretation of its meaning.
There are three stages to the analytics process: Describe, Predict, and Prescribe. Ideally, each of these stages revolve around a feature catalogue that delivers a standard view of the data and includes business meaning. For example, it will have a feature (or attribute) called “Business Address” as opposed to “address number 5 of type 37 in collection X.” Did you understand the second one? Neither did I — and that means it’s open to be interpreted differently by each data scientist assembling data sets.
Describe:
The first stage in building the feature catalogue is to create the semantic model. The semantic model is a virtual model of the data that is completely abstracted from the data sources. It allows you to concentrate on the business description (semantics) and the usage of the features without worrying about where they come from or how they are stored. There’s no need to build the entire catalog at once. You can start small and grow it over time.
Next you map the features to actual data sources. If there are conflicting sources, you can centralize the rules to resolve conflicts and inconsistencies. Because the rules are centralized, you can guarantee that everyone who uses the feature catalogue will use the same rules. It’s also cheaper then maintaining ETL. With ETL, every analytics problem comes with new ETL code which replicates the same logic as previous code for resolving the conflicts. The code bloats very quickly.
Finally, you expose the data in a catalogue shared by everyone.
Predict:
Now comes the machine learning. First, you start with the data analysts and data engineers who explore the data. They use industry standard tools like Sparc, Zeppelin, and Jupyter to look at the data, understand what it’s saying, and determine which features are important to solve a given problem. Sometimes they need additional data, at which time they can go back into the Describe loop to map some new features; sometimes they need to create a new feature with a calculation of one or more correlated features. In either case, new features are added to the catalogue and are available for anyone to use. When data discovery is completed, data scientists will have a list of features required for machine learning.
Before we continue, let’s take a step back and discuss the traditional way data exploration and machine learning happens. For data exploration, data scientists, data analysts, and data engineers compile data requests for back-office fulfillment. The back-office makes a guess at which attributes in which systems are relevant and gives a spreadsheet or data dump to the explorers, who do their magic. When they are done, the explorers hand over a document that lists the required features to the data scientists. What do the data scientists do? They go back to the back-office and request more data. The back-office personnel, likely different people, will gather the data they think is relevant and hand it over in a spreadsheet or data dump. Notice the problems here? First, the data is definitely stale. Second, the data is likely not consistent because different back-office people interpreted the meaning and probably pulled from different systems.
A feature catalogue solves this because everyone is using the exact same features that expose the exact same data and the data is always up to date.
The data scientists use their skills to train and test a machine learning model and deploy a decision model that can be wired into front office applications. Once again, the decision model is going to use the exact same features that were used in data exploration and machine learning, ensuring the consistent application of the machine learning model.
Prescribe:
The decision model can now prescribe what a user should do. In wealth management, this could be the next best action that will increase customer loyalty and encourage a customer to invest more; in corporate banking it could be a recommended product and sales methodology to advance the account plan for a high value client. In either case, the system monitors the results of the actions over time and passes the results back in a feedback loop to enable continuous learning. As bankers and advisors use the system and accept the recommendations or do something else, the system will monitor the results and make better recommendations the next time.
With a single feature catalogue participating in all stages of the analytics process, firms can be assured that decision models will be consistent, more accurate and, most importantly, continue to learn as advisors and bankers use the system.
How do you provision data for machine learning? How do you get the resulting decision models into the hands of the people who need them?
Originally published on nexj.com on June 6, 2018.
Connect with us on social! Linkedin, Twitter, or Facebook
|
Improve Machine Learning Effectiveness with a Feature Catalogue
| 1
|
improve-machine-learning-effectiveness-with-a-feature-catalogue-1816bbc65d1b
|
2018-06-07
|
2018-06-07 05:02:33
|
https://medium.com/s/story/improve-machine-learning-effectiveness-with-a-feature-catalogue-1816bbc65d1b
| false
| 879
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
NexJ Systems
|
NexJ Systems is a leading provider of enterprise customer management solutions (CRM, CDM and CDAi) for the financial services & Insurance firms.
|
323738089033
|
NexJSystemsInc
| 122
| 103
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-09-27
|
2018-09-27 11:28:16
|
2018-09-27
|
2018-09-27 11:31:23
| 0
| false
|
en
|
2018-09-27
|
2018-09-27 11:31:23
| 0
|
18184f1884da
| 0.6
| 0
| 0
| 0
|
My Machine learning Experience:
| 3
|
Machine Learning
My Machine learning Experience:
1.Ml application
Face recognition
Headsets
IPhone keypad
Voice recognition
Robot dogs
Facebook ads
Amazon Netflix
Space for maps
2.missing data
Categorical data
timing and train set
Feature scaling
3.
Regression
a) Simple linear regression
b) Multiple linear regression.
c) Polynomial linear regression
d) Support vector regression
e) Svr intuition
f) Decision tree regression
g) Random forest regression
4.
Classification
a) Logistics regression
b) K-Nearest Neighbor (K-NN)
c) Support vector machine
d) Kernel SVM
e) Naive Bayes
f) Decision tree classification
g) Random forest classification
5.
Clustering
a) K-Means clustering
b) Hierarchical clustering
c) Association Rule Learning
d) Apriori
6.
Reinforcement learning
UCB
Thompson sampling
7.
Natural Language Processing
8.
Deep learning
a) Artificial Neural Networks
b) convolutional Neural Networks
9.
a) Principal Component Analysis
b) Linear discriminant analysis
c) Kernel PCA
10.
Model selection and boosting
Techniques
1. K-fold cross validation
2. Grid search
11.
Model selection
12.
XGboost
|
Machine Learning
| 0
|
machine-learning-18184f1884da
|
2018-09-27
|
2018-09-27 11:31:23
|
https://medium.com/s/story/machine-learning-18184f1884da
| false
| 159
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
sejal jain
|
Machine Learner
|
102e9e13601e
|
sejal8001
| 0
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
78dc282c1c8e
|
2018-03-16
|
2018-03-16 20:19:52
|
2018-03-16
|
2018-03-16 20:45:18
| 6
| false
|
en
|
2018-03-20
|
2018-03-20 01:07:22
| 5
|
181999e861dc
| 2.44434
| 0
| 0
| 0
|
This year’s Super Bowl LII certainly made history, once again, and lots of headlines with a surprise ending. Many were brought together in…
| 5
|
Ad Teardown: Touchdown Celebrations to Come | NFL
This year’s Super Bowl LII certainly made history, once again, and lots of headlines with a surprise ending. Many were brought together in celebration, while others were struck and left speechless. One thing is clear though, fans from every corner of the nation were filled with passion and emotion.
While fans pay attention to the nation’s most important sporting event by the millions, companies, retailers and brands fight it off through advertising. There is no other event in the US that gathers so many eyes and glues people to their screens.
These year’s ad picks were fantastic, but we decided to turn our eyes to the NFL and their throwback ad ‘Touchdown Celebrations to Come’
Ad Details
Producer: NFL
Runtime: 1:00
Published: Feb 4, 2018
The Setup
We created a Campaign using NFL’s YouTube video through the RefineAI Market Research platform.
Once the Campaign was live, we shared the unique URL with a small portion of our global Panel — 50 potential candidates in the US.
Within just 2 hours, we reached results from 42 participants and concluded our test.
The Results
Participants provide a rating for Story, Audio and Visuals.
After watching the video, we asked each participant to rate the Story, Audio, and Visuals. The selections available include Didn’t like it at all, Somewhat liked it and Really liked it. This allows us to easily measure the results and show them through the RefineAI Dashboard in an easy to read format. If any of these parameters ranked low, it allows the producer to focus on feedback related to the lowest ranking area.
Qualitative Feedback
Each participant was asked 2 simple questions:
What do you like the most about the video? and
What would you do differently?
We’ve blurred out the names and emails from panelists
In the past, we asked what they liked the most and the least, but we noticed that by asking in a different way, it yielded to more insightful responses and ultimately more actionable feedback for the video producer.
Emotional Feedback
Measurements include an Overall Emotional Analysis, Real-time Emotional Feedback against the video timeline, and Engagement Score (eye tracking).
Predominant emotion evoked: 😂 Joy at 70%
Overall audience engagement: 👀 Eyegaze Tracking 86%
Highest emotion intensity: 🌎 Hispanic or Latinx | Female | Joy at 96%
Key Moments
0:12, 0:44 and 0:56 were the peak frames with the highest Engagement Score and Emotion Intensity combined.
Video market testing as a service.
Gather actionable feedback for your videos in minutes.
Do you have a good use case for RefineAI? Leave a comment below or drop us a line. Don’t forget to share!
Originally published at www.refineai.com.
|
Ad Teardown: Touchdown Celebrations to Come | NFL
| 0
|
ad-teardown-touchdown-celebrations-to-come-nfl-181999e861dc
|
2018-03-20
|
2018-03-20 01:07:25
|
https://medium.com/s/story/ad-teardown-touchdown-celebrations-to-come-nfl-181999e861dc
| false
| 396
|
Measure the emotional appeal and viewer engagement of any video using AI.
| null |
refineai
| null |
RefineAI
|
founders@refineai.com
|
refineai
|
EMOTION,SENTIMENT,ARTIFICIAL INTELLIGENCE
|
refineaiapp
|
Super Bowl
|
super-bowl
|
Super Bowl
| 2,660
|
Alberto Orsini
|
Designer, husband, dad. I am into delightful UX. Sr. Manager, Product Design at Royal Caribbean.
|
8e17b22a164a
|
albertojorsini
| 542
| 317
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-27
|
2018-05-27 16:43:09
|
2018-05-27
|
2018-05-27 16:48:04
| 3
| false
|
en
|
2018-06-01
|
2018-06-01 07:26:55
| 6
|
1819a9ef3c45
| 2.606604
| 101
| 1
| 0
|
We are proud to announce that we have entered into a strategic cooperation with International Digital Asset Management Co. (IDA), a global…
| 5
|
Announcing our first official dApp: International Digital Asset Management. Ready to digitalize, track and trace $3 billion USD worth of physical assets on the MATRIX blockchain
We are proud to announce that we have entered into a strategic cooperation with International Digital Asset Management Co. (IDA), a global leader in asset digitalization. IDA will be the first official decentralized application to be launched natively on the MATRIX blockchain.
IDA is an asset digitization and tracking company which focuses primarily on Belt and Road projects — it is a key first step for MATRIX’s Belt and Road work implementation. MATRIX will work with IDA to provide technical support and solutions to IDA, who will use the MATRIX public blockchain at the core of their solution.
At the time of writing, IDA currently has assets worth approximately 3 Billion USD which has been or is to be digitalized in the form of gold mines, oil fields and forestry. MATRIX will be the technical partner to transfer this wealth into the digital world via the MATRIX blockchain.
The first project we will start with is IDA’s recently acquired Laos’s virgin forest project. Lao People’s Democratic Republic (Laos) is located in the north of IndoChina Peninsula. Covering approximately 236,800 square kilometers, Laos is the only landlocked country on the peninsula, and forests cover 52% of Laotian territory, making it a famous country of forests. One of the simplest and easiest use-cases to imagine with the MATRIX and IDA in Laos is the digitalization of the wood which is harvested from trees within the forest and sold internationally as seen in the flowchart below.
Example flow-chart of a digitalization process for logging in Laos
As can be seen in the flowchart, utilizing the MATRIX blockchain greatly allows for increased transparency through each step of the process.
Beyond this initial use-case MATRIX and IDA will go even further. Laos has been highlighted as a key Belt and Road region and will soon benefit from additional transport links and energy solutions such as high speed rail, tidal barrages and hydropower schemes which will require services such as blockchain enabled energy trading and cargo tracking. This, coupled with the rich natural resources in the area such as gold mines and water sourcesmakes it a wonderful location to initially focus on, as it will bring tremendous asset value to be tracked and exchanged on the MATRIX blockchain.
IDA have a strong emphasis on ensuring legal compliance both in the issuance of resource tokens and in the trading of them. Given that it will be a global solution there has been great care taken in the design of the application to ensure legal compliance with government bodies at home and abroad, including the SEC. IDA have begun writing a systematic framework and standardization process for the digitization of assets which will be shared with all entities who wish to digitize and trade assets via the IDA application.
More details will follow in the coming weeks about IDAs particular technical solutions and innovations, and some of these were introduced at the Vietnam meet-up on 26th May 2018. We are very proud to have them build on our platform and will continue to attract more top quality projects to join our ecosystem.
Reference：
Website | Telegram | Twitter | Reddit | Technical White Paper
|
Announcing our first official dApp: International Digital Asset Management.
| 2,220
|
announcing-our-first-official-dapp-international-digital-asset-management-1819a9ef3c45
|
2018-06-16
|
2018-06-16 13:02:01
|
https://medium.com/s/story/announcing-our-first-official-dapp-international-digital-asset-management-1819a9ef3c45
| false
| 545
| null | null | null | null | null | null | null | null | null |
Blockchain
|
blockchain
|
Blockchain
| 265,164
|
MATRIX AI NETWORK
|
An open source public intelligent blockchain platform
|
ad51c60ef692
|
matrixainetwork
| 1,003
| 0
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
d9d93b744558
|
2018-09-07
|
2018-09-07 10:37:29
|
2018-09-07
|
2018-09-07 10:39:49
| 2
| false
|
en
|
2018-09-07
|
2018-09-07 10:42:53
| 0
|
181c4fe76d5a
| 4.217296
| 1
| 0
| 0
|
Reviewed by Dr Kenneth Payne, KCL
| 3
|
Life 3.0: Being Human in the Age of Artificial Intelligence
Reviewed by Dr Kenneth Payne, KCL
If you want a breezy and accessible argument about the far future of humanity in an age of Artificial Intelligence, this is the book for you. Max Tegmark is a well-connected polymath, with a longstanding interest in AI and a knack for bringing big ideas to life. Who knows what life will be like for humans in the next billion years? Certainly not Tegmark, but why let that stand in the way of the fun? It’s enough here to ask the questions: what will AI want, and why? Can machines ever become conscious, and if so, how might that feel? Is it possible to gain immortality by uploading ourselves to the net? What happens when machines merge with humans as cyborgs, or ‘augmented intelligence’? These are longstanding questions in philosophy of mind, computer science, neuroscience and a bunch of related fields. Since we can’t know the answers to these questions, the fun comes in sketching out the possibilities — often revealing more about ourselves in the process: what does consciousness do for us, for example? How far is our cognition shaped by our bodies, and by our evolutionary imperative to reproduce?
There are plenty of popular books that cover similar territory. Nick Bostrom famously warned of the dangers of an AI ‘Superintelligence’emerging rapidly and with disastrous consequences for humanity, as it single-mindedly seeks to fulfil the task we humans have set it. Another similarly engrossing account of advanced AI is Murray Shanahan’s Technological Singularity– Shanahan was the brains behind the movie Ex_Machina, and, like Tegmark, is adept at asking the big philosophical questions in easy to digest prose.
In his book, To be a Machine, Mark O’Connell has produced a witty and sceptical look at this sort of AI expert, especially transhumanists who explore the technological possibilities of transcending the human body. Transhumanism is the yin to the yang of superintelligence — its sunny utopian promise of eternal human life contrasts with the dystopian vision of a malign superintelligent AI obliterating life as we know it. Tegmark and Bostrom personify the debate — both fascinated and repelled by the possibilities of machines with minds. Ultimately, there’s something deeply human about their existential search for meaning in a universe with machines that can produce and refine themselves without our help.
But is that really where AI is heading? Other introductions to AI offer more insight on the underlying science — the best works here are Margaret Boden’s AI or Kevin Warwick’s AI: The Basics. Warwick, incidentally, walked the walk — turning himself into a pioneering cyborg, complete with DIY implants. Those in the know call this ‘bio-hacking’, and it’s not for the faint of stomach.
Unlike those two books, one thing you won’t get from reading Tegmark is very much detail on the current state of the art in AI — the underlying processes, and the inherent difficulties. AI research has developed in fits and starts, with promising advances followed by frustrating AI ‘winters’ as research stalls a long way short of achieving anything like human-level intelligence. Right now there are breakthroughs aplenty, and it’s easy to go along with the sometimes breathless hype about what AI might achieve. Things are moving so fast that sometimes a superintelligence seems entirely plausible — often replete, in the public imagination, with the sort of glowing red eyes of Arnie’s T-1000. Much of the running is currently being made by ‘connectionist’ AI that are (very) loosely modelled on the brain’s neural networks. But the godfather of this approach, Geoffrey Hinton doesn’t think that connectionism will ultimately bring about the sort of advanced intelligence that Tegmark, Bostrom et al have in mind. Modern AI lacks the ‘common sense’ of humans, as anyone using Siri can attest. Without much sense of what things ‘mean’, connectionist AI might be headed for another winter — highly adept at decision-making in narrow, limited domains, but wholly incapable of applying its reasoning more flexibly. These Artificial Neural Networks can pick out a cat picture on the internet, but they can’t tell you if they like cats. It doesn’t compute. And while they can demolish you at chess and even poker, they can’t on a whim decide to throw it all in and write a novel. In fact, AI’s have written novels — but they’re truly awful.
The other thing that Life 3.0 won’t tell you very much about is war. Breathless and unrealistic hype about Terminator AI aside, AI has profound implications for those interested in strategy — and especially for air minded strategists. As yet, there’s not much written in this important sub-field. The debate has mostly been about the ethics of autonomous weapons. Will they be able to discriminate between soldiers and civilians? Who can we apportion blame to when things go wrong? Is it possible to ban, or at least control, such weapons? This is important — and there are some big challenges to face — how can we regulate AI given that it’s a dual-use technology? Can we afford to keep a human ‘in the loop’ to watch over decisions if the enemy chooses not to? Compared to these ethical conundrums, the broader literature on how AI will shape warfare is rather nascent. That’s changing this year with the publication of Paul Scharre’s Army of None and my own book, Strategy, Evolution and War– but for now the action is mostly in think tank reports, like DCDC’s recent work on human machine teaming. Whether it’s swarm-like mini-UAVs in the attack role, the use of algorithms for encryption and intelligence analysis, or AI deploying ‘fake news’ to deceive adversaries, this is a rapidly changing and important field. In the course of careers now underway, we will see some rapid and radical changes to organisational culture. New skills, concepts and equipment will undoubtedly be required. If Tegmark’s thought provoking romp through some big ideas is your introduction to this brave new world, that’s no bad thing at all.
Life 3.0: Being Human in the Age of Artificial Intelligence, Max Tegmark, Penguin, 2018.
|
Life 3.0: Being Human in the Age of Artificial Intelligence
| 5
|
life-3-0-being-human-in-the-age-of-artificial-intelligence-181c4fe76d5a
|
2018-09-09
|
2018-09-09 09:52:56
|
https://medium.com/s/story/life-3-0-being-human-in-the-age-of-artificial-intelligence-181c4fe76d5a
| false
| 1,016
|
The Royal Air Force's Centre for Strategic and Conceptual Thinking on Air, Space and Cyber Power
| null |
RAFCASPS
| null |
RAF CASPS
| null |
raf-caps
|
AIRPOWER,AIR FORCE,BOOK REVIEW,MILITARY HISTORY
| null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
RAF CASPS
| null |
7cc0bb7983c0
|
RAFCAPS
| 37
| 1
| 20,181,104
| null | null | null | null | null | null |
0
|
from animalsounds.sounds_generator import generator
__version__ = '0.1.0'
__author__ = First Last <email@gmail.com>'
__all__ = []
sounds_dict = {“cat”:”meow”,”dog”:”woof”,”fish”:”...”}
def generator(animal):
“””
Summary:
Translates what the sound different animals make
Args:
animal (str) – A string of the animal you want to hear, valid options include ‘cat’, ‘dog,’ and ‘fish’
Returns:
A string of the appropriate sound.
“””
sound = “”
try:
sound = sounds_dict[animal.lower()]
except:
raise Exception(f“We don’t know what a {animal} sounds like! You could be on to something!”)
return(sound)
import pytest
import animalsounds
def test_generator_cat():
sound = animalsounds.generator(“cat”)
assert sound == “meow”
def test_generator_dog():
sound = animalsounds.generator(“dog”)
assert sound == “woof”
def test_generator_fish():
sound = animalsounds.generator(“fish”)
assert sound == “...”
def test_generator_notfound():
with pytest.raises(Exception):
animalsounds.generator(“emu”)
[tox]
envlist=py34,py35,py36
[testenv]
passenv = *
commands=py.test --cov=animalsounds tests/
codecov --token={env:CODECOVTOKEN}
deps=pytest
pytest-cov
codecov
import setuptools
setuptools.setup(
name="animalsounds",
version="0.1.0",
url="https://github.com/username/animalsounds",
author="First Last",
author_email="email@gmail.com",
description="An exploration of different animal sounds.",
long_description=open('README.rst').read(),
packages=setuptools.find_packages(),
install_requires=[],
classifiers=[
'Development Status :: 2 - Pre-Alpha',
'Programming Language :: Python',
'Programming Language :: Python :: 3',
'Programming Language :: Python :: 3.6',
],
)
| 61
|
316a0e93fcde
|
2017-11-08
|
2017-11-08 02:11:52
|
2017-11-08
|
2017-11-08 03:20:43
| 4
| false
|
en
|
2017-11-17
|
2017-11-17 16:54:01
| 12
|
181c5e31f3f8
| 7.786792
| 23
| 1
| 0
|
This is my first blog post with hopefully many more to come, talking about how to create your first Python package and release it on PyPI…
| 5
|
Creating Your First Python Package
This is my first blog post with hopefully many more to come, talking about how to create your first Python package and release it on PyPI. I’ve been writing software for my startup and for various ad hoc development projects for the past couple of years, however, after I put out my first package a couple of weeks ago, I realized there is so much one learns and sees the benefit of, that not writing code in a “good practice” way actually becomes more painful.
Being a data scientist first, best software principals weren’t always at the forefront of my mind when doing projects. After forcing myself to write tests and attempt to pick up more best practices, I find it being a mindset that I can’t believe I’ve lived without for so long. These are things learned from a small package, I put out in under 1 weeks too! For those of you who are interested, the package is called icd. I have also just released a second one called fastteradata, however, I accidentally got too excited and am currently in progress of writing the retroactive tests for it and doing a major cleaning of the code now.
Some of the benefits of writing your first package, no matter how small, will be:
Being forced to think much more modularly
Writing code with tests in mind
Thinking about making the api simple enough for others to intuitively grasp and reuse your work
Learning new tools that come along with writing tests
Enough about the benefits, let’s get to it!
First, let’s go through a very basic package example starting with the folder structure. A great resource for folder structure templates is a package called cookiecutter. There are many different templates to choose from, and you just pick one, call the command on a downloaded example, and in no time you have some basics filled out to get you started faster. In my case, I used the pypackage-minimal template because I wanted a simple skeleton and wanted to use pytest.
After you use cookiecutter, you’ll want to make sure you have the following files, which we will go through each one at a high level. You can explore each one of these items in much more depth later on.
Directory <package_name> — Where all of your actual package code will live
Directory tests — Where all of your tests you write will live
tox.ini — Used in conjunction with pytest for multi-environment testing
codecov.yml — Allows you to measure how much of your code is backed by tests
LICENSE.txt — License to distribute your code under
README.rst — Where your initial documentation will live (for larger packages, you might move some of the documentation out of your README into something like a Sphinx site.)
setup.py — Used to generate and specify details about your package
Great, now that we have the structure let’s look at actually writing some code. In your package_name directory, which from now on we will call animalsounds, let’s first start by creating an __init__.py file to represent the package. In that file, we can place something like:
This will be where everything gets imported for your users to access. In this case, we import the generator method from a file, sounds_generator. Now, create a sounds_generator.py file and place something like the following inside.
You’ll notice in this simple function that the doc strings are almost longer than the actual code logic. This is something that you will thank yourself in the future for a thousand times over. But now that we have our first python package function, what’s next?
Tests!
Any piece of code that you write, you should be writing with tests in mind so that you can better break apart larger functions into base components to hopefully get more reuse out of them. In this case, we wrote our package code first, however, you can also write your tests first, knowing your function can’t fulfill them and only then fill in the code until the test passes. There is a lot of great material out there for Test Driven Development (TDD) if you want to explore the philosophy behind it more. This is something that is a major level up in best practice coding that once you force yourself to do, everything seems so much easier.
For this example, we will be using pytest. So first, same as last time, within our tests directory, let’s place an empty __init__.py file. Next let’s write some tests! When using pytest, by default, it searches the tests directory for files and methods that begin with “test_” recursively. So let’s make a file called test_generator.py and place the following code in it.
Now, with the following code, we have covered every case for our generator method. Pytest syntax for most tests will just be an “assert” command with your test of the function with known parameters compared to the expected behavior. The last test we wrote, we took advantage of testing even when we expect the function to raise an Exception case. This syntax in pytest is simply the with pytest.raises(Exception): line followed by an invocation of your method that should raise the exception. You can be more specific in these cases as well with specifying specific types of Exceptions.
To invoke the tests and see them pass, just cd to the root of the project and after making sure you have pytest pip installed, type pytest in the terminal and there you go!
Packaging up
Now that we have our code and our tests, let’s expand the functionality of our testing basis a little by introducing tox. Tox allows us to run through tests in multiple environments so that you can be sure your code will work with the differences between python versions as well as linux vs. osx. In our tox.ini file, let’s place the following code:
The above code essentially says we are going to try and test using python3.4, 3.5 and 3.6 environments lays out the dependencies, which in this case is just pytest and some code coverage libraries, and specifies the commands to execute in each environment. This is a great time to talk about code coverage.
The idea behind code coverage is to see what percentage of your code is covered by your tests. It gives a good idea of the confidence of deploying and trusting the methods you write can be repeatable and trusted. This is because the report comes back with a percentage vs. a normal build just having a pass or fail status based on if all 1 of your tests pass or all 1,000 of your tests pass. Codecov is a service that attempts to do just this. Pytest-cov is a package that allows us to generate the specified codecov reports from pytest, which you can see us invoking in the first command. Then the second command is what uploads the codecov report to the service so we can pull down the appropriate badge. When you set up your codecov account, you can sync up your specified github repo and are given a token for each repo. So when we run tox on our local machine, we will have to set the CODECOVTOKEN environment variable before running tox.
Uh oh, as you can see with tox testing the different python environments, you can see there is a syntax error for the python3.4 and 3.5 environments. That is because our package uses a new feature in python 3.6, the function string. f” some string {variable_to_render}” We can either go back and change our package code for compatibility, or in this case, I like the clarity and syntax better, so let’s just remove the py34 and py35 environments in our tox.ini file.
Ok, almost there!
So at this point, we have our package logic, our tests, multiple environments tests, and code coverage reports. Now what happens if others start contributing to our repo and someone edits code, but forgets to run tox or pytest locally before pushing to our repo? Bugs could be introduced and people who depend on your code now have things that break. This is where Continuous Integration (CI) comes in. For this, we will be using travis. What travis does is every time someone pushes to our github repo, we tell travis to run tests and generate new code coverage reports via tox and if something breaks the build, we can send off notifications alerting us of the tragedy. This then auto updates image sources for our code coverage and build passing/failing badges in our README.
To set up travis, we just connect our repo, insert our CODECOVTOKEN environment variable into the settings for our project, and we are good to go!
So now that we have all of the main code components for our new package, now we have to set the license, write documentation and example use in the README, and edit our setup.py file.
For the license, look up and pick out your favorite software license and just copy paste into your LICENSE.txt file.
For the README, use an online real time editor and try to write it in Rich Structured Text (rst) format with examples, docs, dependencies, installation instructions, etc. If you used cookiecutter, you should already have a nice template to work from.
Finally, the setup.py file, you can follow a form like this:
Can We Upload Yet?
Once we have all of these items, have pushed to github, have seen our continuous integration run our tests and generate the build passing and coverage reports, we can now upload our code to PyPI with confidence!! How do we get it up there though? Well, we are in luck because there’s a python package for that, and it’s called twine.
Twine makes sure to package up and upload your code to the PyPI servers through a secure connection. You will have to also go on PyPI and create a user account before doing this, but after doing the one line twine command to register and upload, you now have your first package deployed!
Congratulations! Try to pip install it and give yourself a pat on the back!
|
Creating Your First Python Package
| 82
|
creating-your-first-python-package-181c5e31f3f8
|
2018-06-19
|
2018-06-19 20:58:06
|
https://medium.com/s/story/creating-your-first-python-package-181c5e31f3f8
| false
| 1,878
|
Tips, tricks and updates we think are cool.
| null |
38thStreetStudios
| null |
38th Street Studios
|
info@38thstreetstudios.com
|
38th-street-studios
|
MARKETING,DATA SCIENCE,SOCIAL MEDIA,WEB DEVELOPMENT,ANALYTICS
|
38thststudios
|
Python
|
python
|
Python
| 20,142
|
Mark Hoffmann
|
Data Scientist at JPL (NASA’s Jet Propulsion Laboratory) / CDO 38th Street Studios
|
cc8070255a47
|
markkhoffmann
| 180
| 94
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-08-31
|
2017-08-31 14:18:44
|
2017-09-30
|
2017-09-30 08:26:34
| 3
| false
|
tr
|
2017-09-30
|
2017-09-30 08:26:34
| 3
|
181d29326782
| 2.048113
| 27
| 0
| 0
|
Bugünkü yazımızda Reinforcement Learning algoritmaların en kolay anlaşılırı olan Q-Learning algoritmasını basit bir örnekle inceleyeceğiz.
| 4
|
Python ve Makine Öğrenmesi: Q-Learning Temelleri
Kartaca
Bugünkü yazımızda Reinforcement Learning algoritmaların en kolay anlaşılırı olan Q-Learning algoritmasını basit bir örnekle inceleyeceğiz.
Python ile Yakınsama: Euler ve Pi Sayıları
Bugünkü yazımda analitik değerlere bilgisayar programları vasıtası ile erişmemizi sağlayan yakınsama yöntemlerinden…medium.com
Q-Learning Nedir?
Q-Learning algoritması diğer Reinforcement Learning yöntemleri gibi “agent” ismini verdiğimiz yapay zeka programlarının geleceğe dair plan kurmasını (bkz. Planning in AI) istediğimiz zaman sıkça kullandığımız yöntemlerden biri. Algoritmanın çalışma prensibi günlük hayatımızda çokca kullandığımız teşvik ilkesine dayanıyor. Nasıl teşvik bir kişiyi veya kurumu belirli bir işe yönlendiriyorsa, aynı şekilde bir robotu- ya da başka bir yapay zekayı- belirli bir amaç doğrultusunda yönlendirebilir.
Algoritmanın ana hattı iki matris yapısı üzerine kurulu. Bunlardan ilk R(Reward=ödül) matrisi, ikincisi ise Q(state=durum) matrisi. Belirli bir iterasyon süresince dünya içinde hareketler gerçekleştiren robotumuz bu hareketler sırasında R matrisindeki değerler vasıtasıyla Q matrisini dolduruyor. Tüm iterasyonlar sonucunda Q matrisinin son değerleri bize optimal sonucu veriyor.
Yazımızın geri kalanında Q-Learning ve RL’nin robotik alanında en genel uygulamalarından beri yol planlama probleminin soyutlanarak basitleştirilmiş bir versiyonunu inceleyerek yukarıda anahatlarını belirttiğimiz algoritmayı inceleyeceğiz.
Problem Tanımı
Yol planlama tanımı basit bir graf dünyası üzerinde geçer. Burada grafın düğümler karar alma noktalarını- yollar için yol ayrımı gibi-, kenarlar ise iki durum arasındaki geçişi- yolların kendisini- ifade eder.
Amacımız robotumuzun buradaki başlangıç noktasından hedef noktasına varması için takip etmesi gereken en iyi yolu öğrenmesi.
Problem Çözümü
İlk önce aşağıdaki yol grafını ifade edecek bir graf yapısı oluşturmamız gerekiyor.
Yol haritamız(http://graphonline.ru/en/ sitesinde oluşturuldu)
Sınıf yapımız
Daha sonrasında bu grafa ait bir R matrisi oluşturmamız gerekiyor(Ben aşağıdaki kod örneğinde elle oluşturdum fakat programatik olarak oluşturmanız da mümkün). Burada ilk olarak grafın ağırlıksız “Adjacency List” gösterimini elde ediyoruz. Bunu birbirine ulaşılabilen düğümlere karşılık gelen elemanları 0, geri kalanları -1 yaparak gerçekleştiriyoruz. Adjacency matrisini teşvik matrisine çevirmek için tek yapmamız gereken hedef düğüme ulaşan yollara bir ödül değeri atamak.
Grafa karşılık gelen R Matrisi
Q matrisimiz ilk olarak R matrisin 0 rakamıyla doldurulmuş bir kopyasıdır. Daha sonrasında her iterasyonda Q matrisini güncellemek için gelecek Q durumlarını hesaplayacak bir fonksiyona ihtiyacımız var. Burada aşağıda ifadesini görebileceğiniz Bellman Denklemi devreyi giriyor.
Bellman Denklemi (slideshare.com adresinde Xavier Giro-i-Nieto’e ait sunudan alındı)
Denklemin bize söylediği şey: “Q matrisini güncellemenin kuralı alınan yolun bize direkt verdiği ödülün(R matrisinden aldığımız “immediate reward”) ve gideceğimiz yerdeki muhtemel rotalardan en iyisinin vereceği ödül(future reward) ile toplamıdır”.
Bellman Denklemi
Buraya kadar oluşturduğumuz alt yapı ile artık çözümü gerçekleştirebiliriz
Oluşan çözüm Q matrisinden optimal yolu aşağıdaki fonksiyon ile bulabiliriz.
Bir sonraki yazıda OpenAI Gym ortamını inceleyeceğiz.
Görüşmek üzere.
|
Python ve Makine Öğrenmesi: Q-Learning Temelleri
| 161
|
python-ve-makine-öğrenmesi-q-learning-temelleri-181d29326782
|
2018-04-13
|
2018-04-13 22:09:07
|
https://medium.com/s/story/python-ve-makine-öğrenmesi-q-learning-temelleri-181d29326782
| false
| 397
| null | null | null | null | null | null | null | null | null |
Q Learning
|
q-learning
|
Q Learning
| 42
|
Sıddık Açıl
|
ITU CS Graduate. Ericsson Türkiye
|
cae549011af8
|
sddkal
| 359
| 89
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-04-19
|
2018-04-19 13:51:52
|
2018-03-29
|
2018-03-29 07:00:00
| 1
| false
|
en
|
2018-04-19
|
2018-04-19 13:55:43
| 0
|
181d90154754
| 2.713208
| 0
| 0
| 0
|
As a researcher, I am no stranger to conferences, but my experience at SXSW managed to leave me a lot more surprised than I anticipated. I…
| 5
|
Notes From SXSW: AI is the Key to Improving Human Experience
As a researcher, I am no stranger to conferences, but my experience at SXSW managed to leave me a lot more surprised than I anticipated. I usually go to technical conferences, but this time, most of the AI talks focused on the end user benefits perspective rather than the technical details. The beauty of attending these types of events goes beyond just learning, it allows you take a step back and gain an entirely different perspective within your own industry. It’s like I keep telling people, diversity, especially in regards to knowledge and backgrounds, is the key to success!
Here’s what you need to know about the importance of user experience according to the people at SXSW:
Communication Goes A Long Way
Here is a common assumption most researchers make: “once I have identified a problem, finalized my research and found the solution, my job is done.” Albeit they have a point, this only corresponds to the first phase of the process. Have you ever heard the expression :” You can lead a horse to water, but you can’t make it drink”? Welcome to your next great challenge, developing a way to insure that your “solution” has a long-lasting impact. What is the point of creating something that people will try, but never use again? Not only do you need to influence people to use the product, you need to communicate it in a way that makes sense to them.
Even though your success is directly tied to the experience and satisfaction of your user, you also benefit from monitoring their behavior. Gaining a crucial understanding into their activity and preferences, gives you a unique opportunity to make precautionary changes before it’s too late. Another great insight we received during SXSW? The right design advocates for a customer’s time, attention and well-being.
User Experience Is Key
UX isn’t just about creating complex and impressive interfaces, it’s also about creating interfaces that fits the users’ needs and impact the user in a positive way. It isn’t difficult to understand that building experiences that subtly help improve someone’s life, is a lot better received than offering a disruptive experience, no matter how much better the tech is. Adaptability is a human strength that has allowed us to evolve to the top of the food chain and we have a arrived to a moment in history where customization is the only strength technology has to survive our short attention spans and incredibly high standards.
Speaking of customization, it’s important to note that we aren’t talking about personalizing in terms of preferences, but rather the ability for a system to recognize what it can do to make your experience more enjoyable and in the end, improve your life in some way. You’re probably wondering how a machine could even ask itself this question, well the answer is quite simple: AI.
AI Makes Everything Better
Customization is also about understanding context. Human beings exist in a wide variety of contexts and in order to make relevant suggestions, we need to integrate AI into the equation. It’s also human nature to question what we cannot control and our confidence will often depend on our level of trust.
AI doesn’t always have the perfect answer and the only way to create experiences that work when prediction cannot be guaranteed is by going back to the first point I made: communication. By including the user, we shift the conversation from imposing to suggesting.
The AI experience is immersive and offers cross-platform opportunities that range from smart home devices to smart cities. The idea is to create an ongoing collaborative process between what users want and technology that can anticipate what they will need.
The biggest takeaway is this: while you are researching and developing your tech, step out of your shoes, slip into the ones of your user and ask yourself: “How can I make it simple, fun and meaningful?”
|
Notes From SXSW: AI is the Key to Improving Human Experience
| 0
|
notes-from-sxsw-ai-is-the-key-to-improving-human-experience-181d90154754
|
2018-04-19
|
2018-04-19 16:37:54
|
https://medium.com/s/story/notes-from-sxsw-ai-is-the-key-to-improving-human-experience-181d90154754
| false
| 666
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
Carolina Bessega
|
Chief Scientific Office & Co-Founder — Stradigi AI
|
48a2ff6e75fb
|
carolina_bessega
| 3
| 2
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
6c34a6d897f6
|
2018-02-09
|
2018-02-09 14:34:27
|
2018-02-09
|
2018-02-09 14:58:43
| 1
| false
|
en
|
2018-02-10
|
2018-02-10 04:50:50
| 1
|
181e838f698e
| 2.562264
| 48
| 2
| 0
|
“I want to get into Data Science but don’t know how?” — This is by far the most common expression I hear from people who want to get into…
| 3
|
Data Science Plan 2018
“I want to get into Data Science but don’t know how?” — This is by far the most common expression I hear from people who want to get into this field.
It is definitely a valid question because there is an ocean of materials, books, MOOCs, classes, tutorials out there. Where to begin? Which one is legit? What’s the starting point? Is it Statistics or Programming or Visualization?
Honestly, it is a really difficult question to answer since most of the Data Scientist are “Scientists” who are expert in their Business domain and bring in years and years of experience.
But that doesn’t mean that you can’t do it. It’s just that it will take a hell lot of time if you want to be really good at what you do. You don’t necessarily have to get a doctorate but you should act like one, study like one, practice like one and teach like one.
A few months back, I came across a subreddit wherein a guy left his graduate school and decided to be a self-learning Data Scientist.
(Such a bold step right?)
He said if “I can learn the same things online at a fraction of fees and whole less amount of time, why should I join a University?”
So I want to share that “Plan” or “Curriculum” that he made. I’m sorry for a being a month late but still, you can catch up. I’ve been following this plan and trust me, it is INTENSE. But I’m also balancing my undergrad studies with this so can’t really complain.
Source: Reddit
It starts with Stats, Algebra and Probability and later goes into Python.
For R lovers, you can take up John Hopkins course on Coursera. Apart from that I really don’t know anything about R so sorry ;)
Next it goes into Python for Stats and reading a couple of books for getting a statistical mindset.
At this point you must be pretty good with Python so might as well do some shorts projects on it. You can follow ML Review Team’s blog for some kickass ideas.
Next, the fun part — Visualization. Followed by a couple of MOOC’s on Data Science.
By this time, you must be pretty confident with the basic Data Analysis with Python.
Personal Suggestion: Do an intern at a startup as an Analyst or Engineer and get a basic understanding of how things work in the industry.
Okay, so back to the plan. Things are about to get real now. Where do you store all the Data? So as a Data Scientist you must know how to store, manipulate, aggregate, mine and handle Database. SQL, MongoDB, Cassandra, PostgreSQL and whatnot. Get a basic understanding of how storage works in a NoSQL environment.
Next comes Big Data — Hadoop & MapReduce. Towards the end, you have to master Machine Learning and Deep Learning.
I’d suggest that throughout this plan, try to focus on one (or) two domains such as — eCommerce or Finance or Media or whatever that excites you. Do all the projects in that domain. Constantly learn more and more about that field. Domain is important because that will make you an expert and I don’t have to tell you that experts mint money far more than anyone else.
I know it feels overwhelming right now. But you have to do this over a year. So follow the plan, do one thing at a time and next year this time, you’ll be telling this story to your junior that how hard you worked to be a Data Scientist :)
Here the link to the original thread — http://bit.ly/2Bln4Ub
PS: Next tutorial is coming up in a few days so stay around.
Until then, keep hustling!
|
Data Science Plan 2018
| 111
|
data-science-plan-2018-181e838f698e
|
2018-06-12
|
2018-06-12 16:02:51
|
https://medium.com/s/story/data-science-plan-2018-181e838f698e
| false
| 626
|
One stop destination for all things Data Science!
| null | null | null |
Data Science Weekly
|
namanjd@outlook.com
|
data-science-weekly-dsw
|
DATA SCIENCE,MACHINE LEARNING,AI,PROGRAMMING,DEEP LEARNING
| null |
Data Science
|
data-science
|
Data Science
| 33,617
|
Naman Doshi
|
Helping aspiring data scientists.
|
a42d268507e9
|
namandoshi
| 103
| 26
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-04-25
|
2018-04-25 05:34:24
|
2018-04-26
|
2018-04-26 17:48:56
| 4
| false
|
en
|
2018-04-26
|
2018-04-26 20:22:56
| 6
|
181f68dcd06b
| 4.167925
| 1
| 0
| 1
|
How human behavior can make algorithms tend to bias…
| 5
|
Algorithmic Bias
Algorithms have become the most effective elements in today’s world and they do the computations and decision makings more reliable and faster than human. They make decisions about the news we read, the jobs we get, the people we meet and the ads we see. According to the stories we read on Facebook social net, texts we send on Telegram messenger, photos we like on Instagram application and the search results we see on Google and etc, we generate Big data. Big data is used to make decisions about health care, employment, housing, education and policing.
Contrary to what many of us might think, technology is not objective. Algorithm’s behavior depends on people who invented them: what code they write, what data they use to train the machine learning models, and how they test the models after they’re finished. For example, Give an image classification algorithm millions of labeled cat pictures and it will be able to tell you whether a photo it hasn’t seen before contains a cat. Give a speech recognition algorithm millions of voice samples along with their corresponding written words, and it will be able to transcribe spoken language faster than most humans.
The more data you feed to your learning algorithm, The better it becomes at the task it’s designed for.
According to above sentence, any amount of data you feed your algorithm of white people, you should do the same for dark-skinned people or any amount of data you feed the algorithm including men, do the same for women and transgenders, etc.
Software programs are not free of human influence, because generally we are creating them and they behave on data that may be provided by humans, machines or a combination of both and according to them algorithms may inherit bias.
Machine learning algorithms learn and evolve based on what people do online.
For instance, a new study by Carnegie Mellon University researchers found that Google’s online advertising system, showed an ad for high-income jobs to men much more often than it showed the ad to women. Algorithmic bias is getting a more serious problem than the past and governments and algorithm developers don’t care about it, or if they do, that’s not enough :)
If the bias exists inside the algorithms that make important decisions, goes unrecognized and unchecked, it could have dangerous consequences, especially for poor communities and minorities.
TED talk provided by Joy Buolamwini
Check this TED talk, provided by Joy Buolamwini, to see a real problem, where a machine doesn’t recognize black people’s face well, but it works great for white people.
Another example is a joint project by researchers at Microsoft and Boston University, in which they found sexist biases in word embedding algorithms, which are used in search engines, translation and other software that depend on natural language processing. Among their findings about the behavior of word embedding algorithms was a tendency to associate words such as “programming” and “engineering” to men and “homemaker” to women.
Algorithmic bias is not a new topic. For decades, academics and experts warned companies and algorithm developers. But what makes it critical at this time, is that these algorithms are getting bold in everyday decisions we make.
Open source VS. algorithmic bias
One of the best ways to fight algorithmic bias is by evaluating the training data fed into machine learning models. The data itself might have a skewed distribution; for instance, programmers may have more data about United States citizens than immigrants, and about rich men than poor women. Such imbalances will cause an AI to make improper conclusions about how our society is in fact represented.
open source technology has extreme potential to help in this fight against algorithmic bias. Modern artificial intelligence is dominated by open source software, from TensorFlow to packages like scikit-learn. The open source community has already proven extremely effective in developing robust tested machine learning tools, so it follows that the same community could effectively build anti-bias tests into that same software.
Open source technology has also proven to be extremely effective for vetting and sorting large sets of data. Nothing should make this more obvious than the domination of open source tools in the data analysis market (Weka, Rapid Miner, etc.). Tools for identifying data bias should be designed by the open source community, and those techniques should also be applied to the plethora of open training data sets already published on sites like Kaggle.
People VS. algorithmic bias
Companies that develop and apply machine learning systems and government regulators, show little interest in monitoring and limiting algorithmic bias. They should pay more attention to this topic before it become out of our hands. Plus, Technology companies use all sorts of mathematical models and aren’t transparent about how they operate!
It would be impossible for humans to oversee every decision an algorithm makes. But companies can regularly run simulations to test the results of their algorithms.
Algorithmic bias is a human problem, not a technical one. Our thought, biases, interests, hates, etc, reflects in AI programs. The real solution is to start removing bias in every aspect of our personal and social lives. This means endorsing diversity in employment, education, politics and more. If we want to fix our algorithms, we should first fix ourselves.
Education is extremely important. We all know people who may be unaware of algorithmic bias but who care about its implications should talk to those people and explain both “how the bias is formed” and “why it matters”.
|
Algorithmic Bias
| 1
|
algorithmic-bias-181f68dcd06b
|
2018-05-24
|
2018-05-24 17:49:15
|
https://medium.com/s/story/algorithmic-bias-181f68dcd06b
| false
| 919
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Soroush Hashemi Far
|
Software Engineer, Machine Learning, Deep Learning, Research
|
b9923b2455fa
|
soroush.hashemi76
| 13
| 5
| 20,181,104
| null | null | null | null | null | null |
0
|
In[1]: import spacy
...: nlp = spacy.load("en")
...: doc = nlp("The big grey dog ate all of the chocolate,
but fortunately he wasn't sick!")
In[2]: doc.text.split()
...:
Out[2]:
['The',
'big',
'grey',
'dog',
'ate',
'all',
'of',
'the',
'chocolate,',
'but',
'fortunately',
'he',
"wasn't",
'sick!']
In[3]: [token.orth_ for token in doc]
...:
Out[3]:
['The',
'big',
'grey',
'dog',
'ate',
'all',
'of',
'the',
'chocolate',
',',
'but',
'fortunately',
'he',
'was',
"n't",
' ',
'sick',
'!']
In[4]: [(token, token.orth_, token.orth) for token in doc]
...:
Out[4]:
[(The, 'The', 517),
(big, 'big', 742),
(grey, 'grey', 4623),
(dog, 'dog', 1175),
(ate, 'ate', 3469),
(all, 'all', 516),
(of, 'of', 471),
(the, 'the', 466),
(chocolate, 'chocolate', 3593),
(,, ',', 416),
(but, 'but', 494),
(fortunately, 'fortunately', 15520),
(he, 'he', 514),
(was, 'was', 491),
(n't, "n't", 479),
( , ' ', 483),
(sick, 'sick', 1698),
(!, '!', 495)]
In[5]: [token.orth_ for token in doc if not token.is_punct | token.is_space]
...:
Out[5]:
['The',
'big',
'grey',
'dog',
'ate',
'all',
'of',
'the',
'chocolate',
'but',
'fortunately',
'he',
'was',
"n't",
'sick']
In[6]: practice = "practice practiced practicing"
...: nlp_practice = nlp(practice)
...: [word.lemma_ for word in nlp_practice]
...:
Out[6]:
['practice', 'practice', 'practice']
In[7]: doc2 = nlp("Conor's dog's toy was hidden under the man's sofa in the woman's house")
...: pos_tags = [(i, i.tag_) for i in doc2]
...: pos_tags
...:
Out[7]:
[(Conor, 'NNP'),
('s, 'POS'),
(dog, 'NN'),
('s, 'POS'),
(toy, 'NN'),
(was, 'VBD'),
(hidden, 'VBN'),
(under, 'IN'),
(the, 'DT'),
(man, 'NN'),
('s, 'POS'),
(sofa, 'NN'),
(in, 'IN'),
(the, 'DT'),
(woman, 'NN'),
('s, 'POS'),
(house, 'NN')]
In[8]: owners_possessions = []
...: for i in pos_tags:
...: if i[1] == "POS":
...: owner = i[0].nbor(-1)
...: possession = i[0].nbor(1)
...: owners_possessions.append((owner, possession))
...:
...: owners_possessions
...:
Out[8]:
[(Conor, dog), (dog, toy), (man, sofa), (woman, house)]
In[9]: [(i[0].nbor(-1), i[0].nbor(+1)) for i in pos_tags if i[1] == "POS"]
...:
Out[9]:
[(Conor, dog), (dog, toy), (man, sofa), (woman, house)]
In[10]: wiki_obama = """Barack Obama is an American politician who served as
...: the 44th President of the United States from 2009 to 2017. He is the first
...: African American to have served as president,
...: as well as the first born outside the contiguous United States."""
...:
...: nlp_obama = nlp(wiki_obama)
...: [(i, i.label_, i.label) for i in nlp_obama.ents]
...:
Out[10]:
[(Barack Obama, 'PERSON', 346),
(American, 'NORP', 347),
(the United States, 'GPE', 350),
(2009 to 2017, 'DATE', 356),
(first, 'ORDINAL', 361),
(African, 'NORP', 347),
(American, 'NORP', 347),
(first, 'ORDINAL', 361),
(United States, 'GPE', 350)]
In[11]: for ix, sent in enumerate(nlp_obama.sents, 1):
...: print("Sentence number {}: {}".format(ix, sent))
...:
Sentence number 1: Barack Obama is an American politician who served as
the 44th President of the United States from 2009 to 2017.
Sentence number 2: He is the first
African American to have served as president,
as well as the first born outside the contiguous United States.
| 11
| null |
2018-03-11
|
2018-03-11 06:35:40
|
2018-03-11
|
2018-03-11 06:47:42
| 1
| false
|
en
|
2018-03-11
|
2018-03-11 06:48:01
| 4
|
18205c4972d6
| 6.260377
| 1
| 0
| 0
|
Natural Language Processing (NLP) is one of the most interesting sub-fields of data science, and data scientists are increasingly expected…
| 5
|
NLP in Python
Spacy
Natural Language Processing (NLP) is one of the most interesting sub-fields of data science, and data scientists are increasingly expected to be able to whip up solutions that involve the exploitation of unstructured text data. Despite this, many applied data scientists (both from STEM and social science backgrounds) lack NLP experience.
In this post I explore some fundamental NLP concepts and show how they can be implemented using the increasingly popular spaCy package in Python. This post is for the absolute NLP beginner, but knowledge of Python is assumed.
spaCy, you say?
spaCy is a relatively new package for “Industrial strength NLP in Python” developed by Matt Honnibal at Explosion AI. It is designed with the applied data scientist in mind, meaning it does not weigh the user down with decisions over what esoteric algorithms to use for common tasks and it’s fast. Incredibly fast (it’s implemented in Cython). If you are familiar with the Python data science stack, spaCy is your numpyfor NLP – it’s reasonably low-level, but very intuitive and performant.
So, what can it do?
spacy provides a one-stop-shop for tasks commonly used in any NLP project, including:
Tokenisation
Lemmatisation
Part-of-speech tagging
Entity recognition
Dependency parsing
Sentence recognition
Word-to-vector transformations
Many convenience methods for cleaning and normalising text
I’ll provide a high level overview of some of these features and show how to access them using spaCy.
Let’s get started!
First, we load spaCy’s pipeline, which by convention is stored in a variable named nlp. declaring this variable will take a couple of seconds as spaCy loads its models and data to it up-front to save time later. In effect, this gets some heavy lifting out of the way early, so that the cost is not incurred upon each application of the nlp parser to your data. Note that here I am using the English language model, but there is also a fully featured German model, with tokenisation (discussed below) implemented across several languages.
We invoke nlp on the sample text to create a Doc object. The Doc object is now a vessel for NLP tasks on the text itself, slices of the text (Span objects) and elements (Token objects) of the text. It is worth noting that Token and Span objects actually hold no data. Instead they contain pointers to data contained in the Doc object and are evaluated lazily (i.e. upon request). Much of spaCy’s core functionality is accessed through the methods on Doc (n=33), Span (n=29) and Token (n=78) objects.
Tokenization
Tokenisation is a foundational step in many NLP tasks. Tokenising text is the process of splitting a piece of text into words, symbols, punctuation, spaces and other elements, thereby creating “tokens”. A naive way to do this is to simply split the string on white space:
On the surface, this looks fine. But, note that a) it disregards the punctuation and, b) it does not split the verb and adverb (“was”, “n’t”). Put differently, it is naive, it fails to recognise elements of the text that help us (and a machine) to understand its structure and meaning. Let’s see how SpaCy handles this:
Here we access the each token’s .orth_ method, which returns a string representation of the token rather than a SpaCy token object, this might not always be desirable, but worth noting. SpaCy recognises punctuation and is able to split these punctuation tokens from word tokens. Many of SpaCy’s token method offer both string and integer representations of processed text – methods with an underscore suffix return strings, methods without an underscore suffix return integers. For example:
Cool, right?
Lemmatization
A related task to tokenisation is lemmatisation. Lemmatisation is the process of reducing a word to its base form, its mother word if you like. Different uses of a word often have the same root meaning. For example, practice, practised and practising all essentially refer to the same thing. It is often desirable to standardise words with similar meaning to their base form. With SpaCy we can access each word’s base form with a token’s .lemma_ method:
Why is this useful? An immediate use case is in machine learning, specifically text classification. Lemmatising the text prior to, for example, creating a “bag-of-words” avoids word duplication and, therefore, allows for the model to build a clearer picture of patterns of word usage across multiple documents.
POS Tagging
Part-of-speech tagging is the process of assigning grammatical properties (e.g. noun, verb, adverb, adjective etc.) to words. Words that share the same POS tag tend to follow a similar syntactic structure and are useful in rule-based processes.
For example, in a given description of an event we may wish to determine who owns what. By exploiting possessives, we can do this (providing the text is grammatically sound!). SpaCy uses the popular Penn Treebank POS tags, see https://www.ling.upenn.edu/courses/Fall_2003/ling001/penn_treebank_pos.html. With SpaCy you can access coarse and fine-grained POS tags with the .pos_ and .tag_ methods, respectively. Here, I access the fine grained POS tag:
We can see that the “ ’s ” tokens are labelled as POS. We can exploit this tag to extract the owner and the thing that they own:
This returns a list of owner-possession tuples. If you want to be super Pythonic about it, you can do this in a list comprehenion (which, I think is preferable!):
Here we are using each token’s .nbor method which returns a token’s neighbouring tokens.
Entity recognition
Entity recognition is the process of classifying named entities found in a text into pre-defined categories, such as persons, places, organizations, dates, etc. spaCy uses a statistical model to classify a broad range of entities, including persons, events, works-of-art and nationalities / religions (see the documentation for the full list https://spacy.io/docs/usage/entity-recognition).
For example, let’s take the first two sentences from Barack Obama’s wikipedia entry. We will parse this text, then access the identified entities using the Doc object’s .ents method. With this method called on the Doc we can access additional Token methods, specifically .label_ and .label:
You can see the entities that the model has identified and how accurate they are (in this instance). PERSON is self explanatory, NORP is natianalities or religuos groups, GPE identifies locations (cities, countries, etc.), DATE recognises a specific date or date-range and ORDINAL identifies a word or number representing some type of order.
While we are on the topic of Doc methods, it is worth mentioning spaCy’s sentence identifier. It is not uncommon in NLP tasks to want to split a document into sentences. It is simple to do this with SpaCy by accessing a Doc's .sents method:
That’s it for now. In later posts I’ll show how spaCy can be used in complex data mining and ML tasks.
|
NLP in Python
| 5
|
nlp-in-python-18205c4972d6
|
2018-06-18
|
2018-06-18 06:44:17
|
https://medium.com/s/story/nlp-in-python-18205c4972d6
| false
| 1,606
| null | null | null | null | null | null | null | null | null |
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Jayesh Bapu Ahire
|
I am an IT professional working on Web development ,Game development and Web security. At present I developed my interest in AI and working in this field.
|
4d137af1d608
|
jayeshbahire
| 542
| 480
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-08-17
|
2018-08-17 14:02:54
|
2018-08-17
|
2018-08-17 14:52:59
| 1
| false
|
en
|
2018-08-17
|
2018-08-17 14:54:44
| 2
|
1820893ea145
| 5.422642
| 10
| 0
| 0
|
Jeya Maria Jose pursued his internship at the National University of Singapore this summer in the booming field of Artificial Intelligence…
| 5
|
The Intern Featurette #4 — Jeya Maria Jose — National University of Singapore
Courtesy: Jeya Maria Jose
Jeya Maria Jose pursued his internship at the National University of Singapore this summer in the booming field of Artificial Intelligence. He is a tech geek who spends his free time exploring the latest technology preferments in the world. He is adept at various domains such as Robotics, Deep Learning and Circuit Design. Let’s interact with him to get his view on the process of attaining internships.
Hi, can you please let us know of the process in which you were selected for the internship?
Hello. I looked into the websites of the labs of top Universities and the research projects being done in the same. Whenever I found a project interesting or when I felt it was relevant to my interested domain, I mailed the principal investigator of that research laboratory directly. I received some positive responses. From there on, I chose which would be the best possible option for me and confirmed with the Professor for the research internship. From the professor’s point of view, I think selection of a student depends solely on the availability of research openings in the lab. Only then, they would even look over the cover letter and the previous works done and decide if a student would be eligible to carry out the work.
Can you share your experience of sending mails to various universities requesting for internships? How much time consuming was it?
Mailing professors and waiting for their replies, especially to those in top Universities abroad needs a lot of patience and perseverance. They would be so busy that they would just neglect any mail that has a subject “Application for Research Internship” or have filters that would directly send your mail to their spam inbox. But, most of them would be in need of interns for helping their PhD scholars or Lab Scientists on their research projects.
Also, they would definitely expect you to be good in the respective field that you are applying because it’s a waste of time for them to take you in and end up only teaching the basics of a particular domain in the internship.
So, I prepared a cover letter carefully giving details about the various projects I had done in similar fields and the notable competitions where I had secured some place. I also gave details about the frameworks and platforms I am comfortable to work on. Also, I sent the link of my GitHub repositories and attached a letter of recommendation from my previous internship and my resume.
I started mailing just after Festember. I mailed around 5 professors daily. I targeted only NUS , NTU and ETH due to personal interests. I would send a reminder mail after 2 days to the same professors if they don’t reply. I followed this routine for around 3 weeks and stopped when I received 4 worthy offers. After getting more details about the project that I would be working on from those Professors, I finalized a Professor and mailed him I would be joining his lab for the internship and to start the Visa process. This happened before the end of October.
What domain does your project belong to? Can you explain about it briefly?
I worked on Deep Learning for Medical Imaging. I used Generative Adversarial Networks (GANs) for CT image Synthesis and Segmentation of Intracerebral Hemorrhage Dataset. I was mainly involved in building a novel Segmentation network that uses the concept of Adversarial Learning for the process of Brain Tumor and Haemorrhage Segmentation.
Also , I participated as a team with my PhD scholar and Professor for two challenges (BraTS and MoNuSeg) of MICCAI 2018, the 21st International Conference on Medical Image Computing and Computer Assisted Intervention, which is going to be held in Granada, Spain later this year. I worked on the Overall Survival Prediction of Brain Tumor Dateset of BraTS( organized by UPenn) and used the Segmentor-Adversarial I had worked earlier for cell nuclei segmentation for MoNuSeg.
How was the work culture? How different was it from your home university research work?
Work Culture was very different. There were no timings or any dress code. There was not even a mentor-mentee type of relationship with the professors and scholars. They followed a collaborative culture where in we were solely responsible for taking out the best experience out of the research. You can put in all your effort and extract best results or do nothing much, have fun because the professors are chill with anything.
Even though there were no timings, I found the people in my lab working for 10 to 14 hours a day and many even turned up on Sundays. I felt their driving force was the interest they had in what they were doing which was something I found very motivating.
What’re the different sub-domains and opportunities that are available in your domain?
Deep Learning(DL) is actually a subdivision under Machine Learning in Artificial Intelligence (AI). It has been a very hot field since its revolution in 2012. In just 6 years , there has been an humongous growth in the applications of DL across various domains.
DL has its applications in Computer Vision , Robotics , Self Driving Cars, Natural Language Processing, Speech and Music Generation, Finance, Healthcare, Gaming and a lot more.
A lot of research in academia has bloomed in this sector and also most of the top companies have already heavily invested in this field. So, I believe opportunities are never-ending and don’t need any further clarification.
Did being a part of any club in college help you in attaining the internship?
I don’t think anyone gets any internship just because they are a part of some tech club.
I believe it’s the work and effort you put up to pursue research that takes you to places.
Being in a tech club can definitely give a lot of opportunities to work in interesting projects and learn new things by meeting like minded people. However one can do all this without being in a tech club too.
I was a part of Spider and it did help me develop my technical skills which were helpful during the internship.
What was your previous work experience in projects on your own self, under your dept profs, internships, and how much it helped you in this internship?
I worked on some interesting project statements previously for competitions in Shaastra, Pragyan and also actively participated in TIIDC and E-Yantra. I did a couple of projects as a part of Spider. I also pursued an internship in my 2nd year in IIT Madras where I worked on Optimization algorithms for Control Systems. I tried to explore as many domains as possible and learn the basics of the same to sort of which field interested me a lot.
I guess that all my previous works, projects and internships helped me do well during my internship. The coding languages and frameworks I learnt and mastered during my previous project experiences were surely of great help.
Any piece of advice to juniors who would be applying for research interns?
Plan well beforehand of what you want out of the internship and apply accordingly.
There are a lot of internship programs like CMU’s RISS , MILA’s Summer Internship apart from the classical DAAD , MITACS, and SRFP. So, I would suggest you to properly Google on all the internships available in your domain. As all these results come a bit late, it’s better to start mailing well in advance to various research groups you are interested in so that you can have a back up. In case you land on multiple offers, always bank on the ones that you think can extract the best research work from.
What’s your future plan of action? Job or highers, and why?
I am passionate about contributing to research and I look it as my career. So, I’m planning to pursue higher studies.
Feel free to contact Jeya Maria Jose for further queries!
WhatsApp — 93456 79386
Facebook — https://www.facebook.com/jeya.jose
LinkedIn — https://www.linkedin.com/in/jeya-maria-jose-357951130/
Interviewed by Venkatesh Bharadwaj. S
An initiative of SENSORS NIT TRICHY, the annual technical symposium of the Department of Instrumentation and Control Engineering of NIT Trichy
|
The Intern Featurette #4 — Jeya Maria Jose — National University of Singapore
| 79
|
the-intern-featurette-4-jeya-maria-jose-national-university-of-singapore-1820893ea145
|
2018-08-17
|
2018-08-17 14:54:44
|
https://medium.com/s/story/the-intern-featurette-4-jeya-maria-jose-national-university-of-singapore-1820893ea145
| false
| 1,384
| null | null | null | null | null | null | null | null | null |
Artificial Intelligence
|
artificial-intelligence
|
Artificial Intelligence
| 66,154
|
SENSORS NIT TRICHY
| null |
a964a50e9798
|
sensorsnitt19
| 33
| 3
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-09-04
|
2017-09-04 21:09:09
|
2017-10-17
|
2017-10-17 18:57:39
| 2
| false
|
en
|
2017-10-17
|
2017-10-17 18:57:39
| 6
|
1820a4784f07
| 3.055031
| 0
| 0
| 0
|
This is the third and final installment of a blog series where we talk about the exciting transition that the recruitment industry is going…
| 5
|
4 Key Problems Facing the Modern Recruiter (Part 3).
This is the third and final installment of a blog series where we talk about the exciting transition that the recruitment industry is going through, and highlight some of the key problems facing today’s modern recruiter. As recruiters ourselves, we’re fed-up with feeling the pain that these problems are causing, so we founded myRecdex to tackle them head on.
If you missed Part 1, you can find it (here) and Part 2 can be found (here)
1. Candidate Profile Data Is Becoming Outdated Faster Than Ever
The shelf-life of candidate data is shorter than ever with profile information becoming stale within a matter of days or weeks. Many recruiters try to “sell” clients on the size of their database, however if that data has gone stale it offers little value.
Today it is incredibly difficult for recruiters to keep accurate information about where their candidates are working, what they are doing, and what technologies they are currently working with. The growing millennial workforce and the rise of the Gig Economy will only accentuate this.
Research shows that candidates are changing jobs more frequently, and new technology is emerging at an exponential rate. Because stored profile data is static in traditional candidate databases and ATS’s, it quickly becomes old and stale. This means that today’s modern recruiters need to manually cross-reference ATS profiles with online candidate profiles which is time consuming and inefficient.
2. Today’s Best Software Is Out Of Reach For Most Recruiters
We’ve seen some great advancements in recruitment technology over the last 12 -24 months with new recruitment marketing software hitting the market. This is good news for the industry as more non-HR based recruitment solutions are becoming available for us (remember, we’re not HR anymore!). However, these products aren’t cheap and unless your department has a big recruitment budget, they’re not an affordable solution for most recruiters. Furthermore, these solutions are priced for teams and companies which means individual recruiters who want to use the latest technology are missing out.
3. Silver Medalists are easily forgotten
As a recruiter, runner-up (a.k.a Silver Medalist) candidates are arguably the most valuable assets that you have in your candidate network, but unless you’re making an extra effort to nurture those relationships, they’re easily forgotten. The ATS’s, spreadsheets, and LinkedIn accounts that the majority of recruiters are using today don’t cater to the need for recruiters to maintain special relationships with those VIP candidates that could be the next hire.
4. The Pros and Cons of Robots
credit: unknown
As Artificial Intelligence (AI) in recruitment technology continues to develop at a blistering rate, more ‘robots’ will be taking over parts of our traditional roles as recruiters. Sourcing and screening functions will become more automated meaning that recruiters who don’t invest in the other ‘soft skills’ areas of recruitment are at risk of loosing their jobs. Regardless of how intelligent recruitment software will become, the best candidates will still need to be sourced creatively, and deep human-to-human relationships will still need to be built.
Today’s Modern Recruiter needs to take steps to future-proof their career by investing more time in nurturing authentic human relationships at scale, while differentiating themselves by building a unique personal brand.
For the modern recruiter ‘who you know’ doesn’t cut it anymore, ‘who knows you’ has become equally important.
How can we solve these problems?
At myRecdex we believe that these major industry problems need to be solved ASAP. We’re building the world’s first personal Recruitment CRM with marketing automation power specifically designed to meet the needs of individual modern recruiters enabling them to future-proof their careers, nurture their network, and ultimately fill more jobs.
As recruiters ourselves, we believe that all recruiters should have access to AI-driven Recruitment Marketing technology — not just the teams that have large budgets — so that we can all play a part in elevating our industry.
We’re on a mission to empower individual recruiters with the technology to take their careers into their own hands and to think and act like marketers.
Learn more about us at www.myrecdex.com and join our mission by signing up to gain exclusive access to myRecdex BETA.
|
4 Key Problems Facing the Modern Recruiter (Part 3).
| 0
|
4-key-problems-facing-the-modern-recruiter-part-3-1820a4784f07
|
2018-03-16
|
2018-03-16 19:09:18
|
https://medium.com/s/story/4-key-problems-facing-the-modern-recruiter-part-3-1820a4784f07
| false
| 708
| null | null | null | null | null | null | null | null | null |
Recruiting
|
recruiting
|
Recruiting
| 15,454
|
myRecdex
|
Building awesome & accessible recruitment tools that make hiring easier and recruiters' lives happier. ENGAGE. NUTURE. CONVERT http://www.myrecdex.com/
|
545d9b3cc75c
|
myrecdex
| 19
| 76
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
fc78dab2b103
|
2018-03-08
|
2018-03-08 04:12:31
|
2018-03-08
|
2018-03-08 05:08:11
| 1
| false
|
en
|
2018-03-08
|
2018-03-08 06:42:13
| 1
|
1822808a0326
| 0.222642
| 0
| 0
| 0
|
Tableau Public Link
| 3
|
Homework #4-Aurora
Tableau Public Link
|
Homework #4-Aurora
| 0
|
homework-4-aurora-1822808a0326
|
2018-03-08
|
2018-03-08 06:42:15
|
https://medium.com/s/story/homework-4-aurora-1822808a0326
| false
| 6
|
A pilot data science hackathon for high school students in Singapore
| null | null | null |
Budding Data Scientists
|
buddingdatascientists@gmail.com
|
budding-data-scientists
|
DATA SCIENCE,EDUCATION,HACKATHONS,SOCIAL CAUSE,HIGH SCHOOL
| null |
Data Science
|
data-science
|
Data Science
| 33,617
|
WOO MEI QI AURORA S2-12
| null |
e0ed2faf2188
|
7114160111
| 0
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
f41f22e4979f
|
2018-03-05
|
2018-03-05 11:28:12
|
2018-03-05
|
2018-03-05 11:38:39
| 1
| false
|
en
|
2018-03-05
|
2018-03-05 11:38:39
| 8
|
1823837ae3aa
| 2.637736
| 92
| 4
| 0
|
Last night’s Oscar winners were completely in line with the predictions announced on March 1 by BigML, the machine learning company where I…
| 5
|
And this year’s Oscar goes to… BigML machine learning
Last night’s Oscar winners were completely in line with the predictions announced on March 1 by BigML, the machine learning company where I am a strategic advisor. And when I say “in line” I really mean that: strike; 100% on the nail, in one category after another. The candidates for each and every one of the six major categories Big ML’s algorithm picked as favorites (best film, best director, best actor, best actress, best supporting actor and best secondary actress) all came in.
Magic? Luck? There was 28,125 combinations, which means a 0.00003556 probability of guessing them all by chance. Obviously, no magic, no luck, but machine learning: the result of working with the right algorithm. And such algorithm, once correctly fed with data and fine tuned, was able to predict the voting intention of the more than 7,000 members of the Academy.
Feed in the information: details about the film (duration, budget, genre, etc.), its IMDB ratings, and nominations in previous awards (Golden Globes, BAFTA, Screen Actors Guild and Critics Choice), applied to the same database from 2000 to 2017 used in last year’s predictions. In total, more than 100 pieces of data per movie. This year, the scores awarded by IMDB users were eliminated from the equation because they are complex to acquire and had little impact last year anyway. Finally, the results of the predictions were evaluated by applying them to nominees between 2013 and 2016, with very good results: the models were able to predict the winners of each category in four consecutive years with very few errors.
Seeing the result and knowing how it was obtained gave me a strange sense of predictability, something between “Yep, that makes sense” and “I told you so.” Actually, it’s just about applying the right methodology to the right data: last year, ensembles, this year, deepnets, or deep neural networks. A model for each category, which takes about half an hour to be trained by testing dozens of different networks and that ends up creating a high-performance classifier. Obviously, it’s possible for a winner to emerge in a particular category that defies the predictions, winning against all odds… but the more data you have and the better it is, the more likely your algorithms are to get it right.
That said, BigML was just carrying out an exercise, one that many other companies do: we all know that when IBM, Google or Carnegie Mellon create algorithms capable of winning chess matches against grand masters or beating brain boxes at Jeopardy, Go or poker, what they are really trying to do is showcase the possibilities of their technology. They may garner headlines in the process, but what their work is really about is helping decision-makers begin to understand the potential of machine learning.
I repeat, this year’s Oscar predictions are not magic: the key to successful machine learning is in defining the objective, in obtaining the right data, in transforming it and in carrying out the processes required to obtain a model and then evaluating it. Sadly, real life is rarely like this: businesses looking to use algorithms to gain a competitive edge will find that data is hard to find or will be in the wrong format or incomplete and may not be able to be imported into a database easily. On many occasions, the objective may not even properly be defined. As said, it’s not magic, it’s hard work. Someone has to define the task, carry it out and then use the right tools, i.e., that are able to do the job and then interpret the results. But when you have the right data and the right tools, the results make perfect sense and can illustrate the huge potential of machine learning.
And this year’s Oscar goes to… machine learning.
(En español, aquí)
|
And this year’s Oscar goes to… BigML machine learning
| 468
|
and-this-years-oscar-goes-to-bigml-machine-learning-1823837ae3aa
|
2018-05-02
|
2018-05-02 12:54:19
|
https://medium.com/s/story/and-this-years-oscar-goes-to-bigml-machine-learning-1823837ae3aa
| false
| 646
|
On the effects of technology innovation (writing in Spanish at enriquedans.com since 2003)
| null |
enriquedans
| null |
Enrique Dans
|
enrique.dans@ie.edu
|
enrique-dans
|
TECHNOLOGY,DISRUPTION,SOCIAL MEDIA,INNOVATION
|
edans
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Enrique Dans
|
Professor of Innovation at IE Business School and blogger at enriquedans.com
|
acfc0d77aeef
|
edans
| 32,707
| 236
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2017-10-11
|
2017-10-11 10:06:32
|
2017-10-11
|
2017-10-11 10:09:14
| 1
| false
|
en
|
2017-10-11
|
2017-10-11 10:09:14
| 4
|
18277cd03285
| 1.860377
| 0
| 0
| 0
|
Stealing from startups to make big business innovation an organisational habit.
| 4
|
Trigger, rigour, reward
Stealing from startups to make big business innovation an organisational habit.
TRIGGER
Emerging technologies create new opportunities and over-inflated expectations.
Right now artificial intelligence (AI) is a case in point.
Consultants like me talk large about how AI will disrupt every sector, transform the customer experience, and drive enterprise efficiency.
Just the other day I was doing just this about the future supermarket journey.
The problem is that big organisations are often too wrapped up in business-as-usual thinking to explore potentially disruptive ideas. I understand why, and am wise enough not to get frustrated by this.
But, in a world where startups are leveraging enabling technologies to disrupt and ultimately take over, it’s still a problem, whether or not big companies choose to worry about it.
RIGOUR
So how might we solve it?
One way may be to steal simple approaches from the startup world.
A few years ago I worked with a startup that was writing a pitch deck to raise funding. I’d written lots of pitch decks before, but this was different. For me, this quote sums up why:
“Startups are as impersonal as physics. You have to make something people want, and you prosper only to the extent you do.” — Paul Graham
In the big business world it’s easier to ‘game the system’ as Paul puts it; using executive authority or simply smoke and mirrors to determine which ideas get explored, funded and ultimately executed.
In the startup world, nobody — investors, founders or users — wants to waste time or money on a bad idea.
My take away from this was the line, “Make it so they can’t say no”. A rigorous pitch deck — built on customer insight, product vision and business analysis — can not just win the argument, but stop the argument happening in the first place; you get to a level of fidelity where it becomes inarguable.
I’ve started to work this through here, if you’re interested: Pitch like a startup.
REWARD
Recently I have been reading ‘The Power of Habit’, and wondering if ‘the habit loop’ could be applied here.
Simply put, an enabling technology triggers interesting ideas, startup-style rigour helps organisations work them through in a simple, repeatable way, and the ultimate reward is a ‘yes’ from the powers that be; a mandate to experiment, learn, fail, succeed.
This habit loop (trigger-rigour-reward) might just create an organisational desire to think and act more like startups. And if this desire is seeded into culture, a genuine habit loop is formed.
This is rough thinking by the way, so I welcome comments!
|
Trigger, rigour, reward
| 0
|
trigger-rigour-reward-18277cd03285
|
2017-10-11
|
2017-10-11 10:09:15
|
https://medium.com/s/story/trigger-rigour-reward-18277cd03285
| false
| 440
| null | null | null | null | null | null | null | null | null |
Innovation
|
innovation
|
Innovation
| 59,190
|
Chris Dane
| null |
7aa397b295ff
|
chrisdane
| 29
| 39
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
7d31440eeaef
|
2018-01-29
|
2018-01-29 18:18:26
|
2018-01-29
|
2018-01-29 20:15:58
| 1
| false
|
en
|
2018-01-30
|
2018-01-30 15:01:19
| 5
|
182925ed8468
| 2.230189
| 14
| 0
| 0
|
Machine learning, despite its notable constraints, has found incredibly useful applications within medicine, image recognition, and many…
| 5
|
Machine Learning and Policy Analysis
Machine learning, despite its notable constraints, has found incredibly useful applications within medicine, image recognition, and many other areas; it has allowed computers to perform certain “domain-specific” tasks faster, more efficiently, and more accurately than human beings.
As the future moves forward, advancements in brain imaging, quantum computing, and biologically-enhanced intelligence will only propel machine learning forward towards “true” AI capability — greatly reshaping the globe.
How it Works
Modern machine learning is largely achieved through “neural networks,” specialized algorithms that simulate the brain’s functionality. By taking in some number of inputs and then systematically building “neural” connections between those inputs, these algorithms can achieve a complex understanding of data. This allows them to produce some number of outputs as a result — i.e. reading in the pixels of an image, making deductions from those pixels, and then identifying the object in the photo.
When these connections are made, they can occur with the “supervision” of a human being, meaning we say “I have x which yields y, so figure out what connects them,” or without the supervision of a human being, meaning we say, “I have x, y, and z, but can’t figure out what connects them — so derive that for me.” This, however, is precisely why machine learning cannot be effectively used for policy analysis.
The “Black Box” Problem
As I’ve discussed in numerous other technology-, ethics-, and policy-focused articles, this underlying machine learning behavior makes these algorithms more or less a “black box.” While we know the inputs and outputs, we don’t exactly know what happens in between.
This can cause everything from extreme network vulnerability (i.e. manipulating input data to disrupt functionality) to dangerous algorithmic bias (i.e. recommending discriminately longer prison sentences for black convicts). Many research groups, such as one between Cornell, Microsoft, and Airbnb, have worked on getting these algorithms to better display their internal processes, but most of the work is still in its infancy.
Until their goals of ML transparency are achieved, however, these algorithms will remain a mystery. We can’t distinguish between correlation and causation. We can’t be sure if the model’s statistical processes are in some sense flawed. And we can’t be sure how to fix it all even if we identified it in the first place.
Policy Constraints
This is a fundamental problem for policy analysis, as scholars like Stanford Economist Susan Athey have discussed. A field that depends on clear distinctions between correlation and causation simply cannot rely upon techniques that blur or even conflate these properties. But in a world increasingly dependent on technology, this seems like a ridiculous constraint.
We could think of myriad uses of machine learning in policy-making — from analyzing national security intelligence to garnering new insight into welfare economics. It seems that policymakers both at home and abroad should be regularly using machine learning to inform their decisions for the better. And, logically, this makes sense — but it’s prohibited by technical constraints.
So until changes on this front are pushed by the policy and technology communities, policy decision-making will be unable to leverage this powerful technology for the better. It’s time we advocated for that change.
|
Machine Learning and Policy Analysis
| 33
|
machine-learning-and-policy-analysis-182925ed8468
|
2018-05-02
|
2018-05-02 09:38:57
|
https://medium.com/s/story/machine-learning-and-policy-analysis-182925ed8468
| false
| 538
|
Technology, digital, and innovation in government and foreign policy
| null |
asandreDC
| null |
Digital Diplomacy
|
asandre@gmail.com
|
digital-diplomacy
|
DIGITAL DIPLOMACY,FOREIGN POLICY,DIPLOMACY,POLITICS,TECHNOLOGY
|
andreas212nyc
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Justin Sherman
|
Cyber Strategist / Fellow at Interact / Fellow at Duke Center on Law & Technology / Co-Founder & VP of Ethical Tech / Student at Duke University
|
48796e380fe7
|
justinwsherman
| 121
| 32
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
|
128ca62dd629
|
2017-11-28
|
2017-11-28 16:28:11
|
2017-12-14
|
2017-12-14 18:06:55
| 1
| false
|
en
|
2017-12-14
|
2017-12-14 18:06:55
| 10
|
182ba2086cf6
| 2.743396
| 11
| 0
| 0
|
In our second open source release in recent times, today we are announcing the release of Neuralyzer, an artificial neural network (ANN)…
| 4
|
Serving up cool and refreshing neural nets on-device with Swift
Neuralyzer is like a brain for your phone!
In our second open source release in recent times, today we are announcing the release of Neuralyzer, an artificial neural network (ANN) library for Swift that supports on-device learning using a variety of neural network architectures, optimizers, loss functions, and layer types.
We created Neuralyzer because we wanted to a) be able to train models on-device using b) an API that is similar to our offline workflow (i.e., TensorFlow and/or Keras). Also, c) since most ANN frameworks for Swift are designed for larger convolution neural networks (for use in image classification etc.), they aren’t generally designed for on-device training. So we created Neuralyzer to address this gap in our workflow. Design goals for the project include a) ease of use (our data scientists aren’t Swift developers), b) speed (we use Accelerate wherever possible), and c) lightweight (it’s a small library, with ‘no’ external dependencies and minimal components. We even use our own matrix library).
To give you a better idea of what we’ve come up with, here is a quick example that fits a simple neural network to the famous Iris data-set:
Iris Example
The Iris flower data set or Fisher’s Iris data set is a multivariate data set introduced by the British statistician and biologist Ronald Fisher in his 1936 paper “The use of multiple measurements in taxonomic problems as an example of linear discriminant analysis”. The data sets consists of 3 different types of irises’ (Setosa, Versicolour, and Virginica) petal and sepal length, stored in a 150x4 JSON array. The rows being the samples and the columns being: Sepal Length, Sepal Width, Petal Length and Petal Width. See the Wikipedia article for more information about this dataset.
In this example, we’ll attempt to train a neural network to classify the type of Iris (Setosa, Versicolour, or Virginica) using each flower’s petal and sepal length and width. If you play along with this example in Neuralyzer’s Playground, the data are stored as a JSON array in the Playground’s Resources folder (`iris.json`).
We start by loading the Data and decoding it using Swift 4s Codable niceness. Once we have a reference to the JSON data, we can convert it into an array for extraction into inputs (X) and targets (Y).
Model setup and training proceeds pretty much the same as in other neural network packages; we specify a 3-layer neural network with 4 input features (petal and sepal lengths and widths), a 16-unit hidden layer, an 8-unit hidden layer, and a 3-unit output layer. We’ll use sigmoid activation functions on the first two hidden layers, and softmax on the output layer (for classification probabilities).
As we’re doing classification, we use a CategoricalCrossentropy loss function, and will use RMProp as our optimization routine. Here I’ve increased the learning rate to 0.1 to make our model adapt a bit faster. We should be able to achieve > 80% accuracy in under 20 epochs with these settings.
The actual backpropagation training proceeds very similarly to something like Keras or some other wrapper around something like TensorFlow. Along the way, we can evaluate our model accuracy, and even do timings, using simple Swift commands.
Finally, we can produce predictions, and export our fitted model to a Swift Dictionary or JSON object.
And there you have it! A very simple, fast, and super lightweight feed-forward artificial neural network library in Swift. As you can imagine, this opens up a whole slew of possibilities for on-device learning using neural networks.
So check out our official open source release of Neuralyzer over here, and let us know what you think. Neuralyzer is a work in progress (let’s say α stage), so input is very much welcome. Having said that, if you find it useful, or you have some issues, ideas, merge requests, or complaints, please let us know. And of course, if you know of any other nice neural network implementations for Swift, we’d love to hear about that too.
|
Serving up cool and refreshing neural nets on-device with Swift
| 81
|
serving-up-cool-and-refreshing-neural-nets-on-device-with-swift-182ba2086cf6
|
2018-06-10
|
2018-06-10 10:18:36
|
https://medium.com/s/story/serving-up-cool-and-refreshing-neural-nets-on-device-with-swift-182ba2086cf6
| false
| 674
|
Building new digital experiences by focusing on user privacy, openness, and decentralization. Makers of https://www.textile.photos/
| null |
textileio
| null |
Textile
|
andrew@textile.io
|
textileio
|
TECH,STARTUP,MOBILE APP DEVELOPMENT,DECENTRALIZATION,PHOTOGRAPHY
|
textile01
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Carson Farmer
|
Data Scientist at Textile, Assistant Professor at University of Colorado Boulder
|
adda52ceb6bc
|
carsonfarmer
| 431
| 631
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-04-26
|
2018-04-26 13:35:35
|
2018-04-26
|
2018-04-26 13:45:31
| 3
| false
|
fr
|
2018-04-26
|
2018-04-26 14:22:32
| 6
|
182c8fa9bd1a
| 8.025472
| 0
| 0
| 0
|
Le monde #IT connaît depuis l’évolution technologique de ces dernières années, notamment avec l’essor des #FinTech et des #RegTech, des…
| 5
|
Réseaux neuronaux artificiels et blockchain : science-fiction ou future de l’IA
Le monde #IT connaît depuis l’évolution technologique de ces dernières années, notamment avec l’essor des #FinTech et des #RegTech, des bouleversements considérables qui modifient radicalement le paysage informatique et les solutions proposées à l’ensemble de ses acteurs. Dans le secteur financier, deux technologies IT de pointe attirent actuellement l’attention : la #blockchain d’une part et les #réseaux #neuronaux #artificiels (#artificial #neuronal #network, ANN) de l’autre.
Pour rappel, la blockchain est une base de données partagée, possiblement décentralisée, à laquelle ses utilisateurs peuvent accéder dans son entièreté et contrôler dans une certaine mesure son intégrité.
La blockchain est ainsi composée de technologies complexes sous-jacentes dont la cryptographie qui a connu une forte médiatisation récente en raison de l’intérêt accru pour les problématiques de cybersécurité rencontrer notamment par les plateformes et les fonds de cryptomonnaies. Bien que la blockchain soit un travail d’orfèvre informatique dont le développement requiert de hautes connaissances IT, l’explosion fulgurante des #cryptomonnaies de ces quelques dernières années a rendu cette technologie accessible au grand public. Outre les simples spéculateurs qui font fi de la technologie sous-jacente de la blockchain, les entreprises exerçant essentiellement dans les secteurs de l’informatique et de la finance commencent à s’approprier cette technologie dans un souci d’automatisation et d’optimisation des échanges (pour exemple : projet IBM Blockchain Platform, le projet interbancaire Utility Settlement #Coin ou encore la résurrection de #Kodak via sa nouvelle plateforme de photographie).
Les ANN sont une forme d’intelligence artificielle dans le sens où ils constituent une des techniques mises en œuvre en vue de simuler un système de pensée complexe. Ils sont néanmoins bien moins connus des néophytes alors que les premiers travaux sur le sujet ont été publiés bien avant que la blockchain ne gagne ses lettres de noblesse (voir les travaux de Donald O. Hebb des années 1940). Les ANN sont, comme leur nom l’indique, inspirés du fonctionnement des neurones biologiques. De façon plus précise, un ANN est en général composé d’une succession de couches de neurones. Ces couches sont de trois grands types, présentés comme suit :
- couches d’entrée qui reçoivent les données provenant d’autres traitements ;
- couches cachées qui reçoivent les résultats en provenance d’autres neurones et transmettent leurs résultats aux neuronaux des couches suivantes ; et
- couches de sortie qui transmettent leurs résultats aux utilisateurs.
Chaque neurone peut être considéré comme un programme alimenté de données (inputs) pour obtenir une solution mathématique statistique ayant un résultat unique (output). Chacun de ces neurones prend comme inputs les outputs des neurones précédents (couches cachées notamment). Les premiers neurones de la couche d’entrée prennent des données externes fournies par l’utilisateur ou transmises par d’autres systèmes qui peuvent aussi être des ANN (exemple : prix, volatilité, temps ; température, vitesse, vecteur position ; âge, taille, poids, etc.), puis les transmettent aux neurones des couches suivantes et ainsi de suite jusqu’à la couche de sortie qui retourne un output final pour l’utilisateur ou un autre système (schéma 1).
Plus précisément, à l’entrée de chaque neurone, chacun des inputs reçus est pondéré par une certaine valeur (poids) spécifique à chaque neurone pour chaque type d’input puis l’ensemble de ces inputs pondérés est traité (filtré, calculé) par chacun des neurones selon une série de fonctions mathématiques statistiques complexes qui renvoient un output(schéma 2). Cet output sera lui-même ensuite traité par les autres neurones de façon plus ou moins complexe, jusqu’à retourner une solution finale à l’utilisateur via la couche de sortie. Ces schémas rappellent classiquement les schémas utilisés en électronique, optique, en science de l’ingénieure, etc. pour visualiser les fonctions de transferts pour des équations linéaires. Les ANN augmentent de facto la complexité de ces schémas, les fonctions utilisées par les neurones n’étant pas linéaires.
Pour analogie, ces schémas peuvent être aussi apparentés au fonctionnement d’un réseau d’arbres de probabilité donc chaque nœud pourrait être assimilé à un neurone. C’est pour cela notamment que les règles simples de la statistique comme l’inférence bayésienne(actualisation d’une valeur statistique d’une donnée historique en fonction de la connaissance d’une donnée nouvelle) s’appliquent naturellement.
Les différentes règles de calcul des ANN définissent la méthode d’apprentissage spécifique de l’ANN et sont regroupées dans la discipline appelée machine learning. Parmi ces méthodes, la plus en vogue est l’#apprentissage #profond (appelé aussi #deep #learning, deep structured learning, ou hierarchical learning). Le deep learning regroupe deux types de méthodes couramment utilisées :
- l’apprentissage supervisé (#supervised #learning), les données et les résultats attendus sont connus (ImageNet de #Google) ; et
- l’apprentissage non supervisé (unsupervised learning), les résultats attendus sont inconnus et l’utilisateur laisse le système détecter des caractéristiques ou comportements particuliers (features) de l’ensemble de ces résultats.
Conséquemment la blockchain et les ANN ne s’opposent pas par leur nature mais demeurent, à ce jour, des domaines très disjoints de l’informatique.
Néanmoins, au vu des avancées très récentes de ces deux applications technologiques, nous nous sommes interrogés sur la pertinence d’une éventuelle alliance. Quel serait l’intérêt de réunir ces deux technologies a priori sans lien ?
Rappelons que conceptuellement, la blockchain est un historique, une mémoire sécurisée et partagée. Pour sa part, l’ANN est une réflexion, une logique formelle fondée sur le traitement d’énormes quantités de données par son exceptionnelle puissance de calcul. L’alliance de ces deux technologies permettrait donc de créer, en théorie une réflexion très poussée par sa puissance de calcul et par la rigueur de sa méthodologie statistiquement juste s’appuyant sur une mémoire immense et fiable. Les perspectives théoriques semblent donc, aux premiers abords, incommensurables et révolutionnaires.
C’est donc sans grande surprise que les premiers pas de cette alliance inédite mais particulièrement congrue ont été réalisés dans le secteur financier.
En effet, et peut-être sans que cette analyse ait été consciente, l’ensemble des traders utilisant les ANN pour optimiser la finance applicable aux cryptomonnaies utilisent à ce jour le parangon de l’intelligence artificielle. Dans cet exemple, l’ANN est donc un trader de crypto-assets artificiels apprenant de son expérience de trading à une vitesse inhumaine et optimisant seconde après seconde sa stratégie de trading sans que la connaissance de la valeur intrinsèque du sous-jacent ne soit requise. Les risques sont grands, aussi, un suivi constant des équipes de la gestion des risques et une confrontation avec l’expérience et la connaissance humaine sont encore indispensables. Cet exemple, qui représente un progrès majeurs dans le monde du trading, n’est pourtant qu’un balbutiement parmi le champ des possibles offerts par ces technologies complémentaires.
Ainsi, les ANN pourraient être utilisées afin de créer une blockchain répondant aux problématiques actuelles :
- minimisation des frais de transactions ;
- minimisation des latences ;
- augmentation de la cybersécurité ;
- rationalisation des processus (suppression du caractère émotionnel dans la prise de décision) ;
- etc.
De façon plus générale pour tout secteur où la blockchain peut être applicable (logistique, comptabilité, capacité de calcul, IoT, etc.) ne peut-on pas considérer possible d’optimiser les processus par l’ajout d’ANN ? Ultimement, ne pourrait-on pas imaginer des ANN capables de coder par eux-mêmes ou simplement de procéder à une explication détaillée de leurs décisions ?
A l’heure actuelle hélas, certains écueils persistent encore pour voir apparaître une génération spontanée de Blockchain Artificial Neuronal Network, ou #BANN, en raison des problématiques rencontrées par ces deux technologies :
Les blockchains actuellement utilisées rencontrent leurs limites notamment du fait d’un seuil dans leur code. Cette limite étant représentée pour les cryptomonnaies par un temps de validation d’un block très long impliquant des frais de transactions élevés. D’autres part, les interconnexions entre les différents systèmes de blockchain sont limitées et leurs systèmes de gouvernance dissociés (exemple : comment faire interagir des éléments enregistrés dans la blockchain avec des éléments de même nature qui n’y sont pas ?).
En ce qui concerne les ANN, la prise en compte de tous les facteurs dont ceux difficilement calculables ou informels peut engendrer une situation appelée #overfitting (pour une masse de données non corrélées, trop importante, l’ANN renvoie les mêmes outputs et/ou formule des réponses incohérentes pour l’utilisateur).
Cependant, l’histoire des technologies a une constante : les limites d’une génération sont généralement résolues par une nouvelle vague d’innovation.
Ainsi, nous pouvons d’ores et déjà imaginer de potentielles futures applications des technologies de BANN. Parmi celles-ci, l’émergence d’une intelligence générale, collective et distribuée (via une blockchain de type #Ethereum) qui facturerait l‘utilisateur pour les requêtes demandées à la BANN et redistribuerait les gains aux participants selon des règles qui lui sont propres, variables et adaptables. Exemple, quid de #smart #contracts dont les paramètres se réactualiseraient à chaque évolution du marché, plus précisément imaginez si non seulement le prix de votre abonnement téléphonique, électricité, gaz, etc. mais aussi certaines clauses comme les conditions de résiliation, les modalités de paiements, etc. s’optimisaient en comparant à chaque instant les informations disponibles dans le marché pour en retenir le plus favorable et ce sans intervention humaine ?
Au même titre, pourquoi ne pas remplacer les services fournis par les géants du web, type GAFAM, par des services dont la gouvernance serait issue d’une décision collective émanant par exemple d’un Etat ou d’une organisation mondiale ? Cette immense puissance de calcul requise serait alors distribuée entre les utilisateurs via une BANN.
Cependant, ces exploits théoriques souffrent en pratique de l’absence de plusieurs « briques logiciels ». Celles-ci consistant en une étape intermédiaire au cours de laquelle différentes entités, privées ou publiques, seraient amener à démocratiser ces technologies afin de permettre à un écosystème de se structurer. Il est possible de faire un parallèle avec les investissements d’infrastructures portés par les Etats (infrastructure de réseaux : routes, gaz, eau, téléphonie, internet). D’un point de vue micro-économique, ces infrastructures constituent les fondations du développement de nombreux acteurs de taille modeste TPE/PME, ainsi que la création de nouveaux marchés pour les sociétés de plus grande envergure. D’un point de vue macro-économique, les infrastructures contribuent à l’augmentation de la productivité globale des facteurs et donc à l’augmentation du produit national brut. La difficulté est l’articulation de ces deux approches macro et micro (top-down ou bottom-up), et ce, quel que soit le type d’infrastructure considéré.
Dans le cas particulier de la blockchain, la mise en place de service hybrides de type BANN pourraient permettre l’émergence :
- de plateformes d’échanges de services numériques décentralisés ;
- d’interfaces monde virtuel — monde physique capables de réaliser des prestations de services dans le monde physique et virtuel ;
- de l’individualisation de toutes les prestations de services ;
- d’IA, avatars dans la BANN inspirés des comportements et identités des personnes physiques/morales ;
- de systèmes de gouvernance propres, avec à terme la création de normes sur-mesures adaptables à chaque individu puis d’Etats décentralisé ;
- un plus haut degré de mondialisation.
Alors que les secteurs de la blockchain et des ANN ne cessent d’innover, les changements tant attendus ne sont pas encore à la portée de tous. Bien que certaines industries de pointe, en finance ou en technologie, se soient déjà emparées de ces sujets, leurs usages restent cantonnés à des domaines bien spécifiques. Cependant, nous sommes à l’aube d’une nouvelle ère technologique qui sera marquée par certaines révolutions déjà connues (voitures autonomes, objets et lieux connectés, désintermédiation bancaire et financière via les cryptomonnaies, etc.) et d’autres inconnus qui répondront à des besoins sous-estimés notamment l’émergence des réseaux sociaux numériques. A l’instar des changements survenus avec Internet, ceux de la blockchain et des ANN auront une portée mondiale et modifieront notre manière d’interagir. Néanmoins, à supposer la faisabilité technologique de ces projets, de véritables problématiques politiques et éthiques émousseront ces révolutions, les impacts structurels sur l’économie des entreprises et des Etats étant à ce jour inconnus et inimaginables.
Pour l’ensemble de ces raisons, nous continuerons à suivre, à analyser et bien sûr discuter les évolutions de chacune de ces technologies et de leurs applications pratiques.
Jérôme BLANCHET
Grégory WAWSZYNIAK-DUMONT
Ivan WAWSZYNIAK-DUMONT
|
Réseaux neuronaux artificiels et blockchain : science-fiction ou future de l’IA
| 0
|
réseux-neuronaux-artificiels-et-blockchain-science-fiction-ou-future-de-lia-182c8fa9bd1a
|
2018-04-26
|
2018-04-26 14:22:33
|
https://medium.com/s/story/réseux-neuronaux-artificiels-et-blockchain-science-fiction-ou-future-de-lia-182c8fa9bd1a
| false
| 1,981
| null | null | null | null | null | null | null | null | null |
Blockchain
|
blockchain
|
Blockchain
| 265,164
|
Scale Fuse
|
We use AI efficiency and blockchain security to provide optimized finance solutions and management
|
484aa60f122a
|
scalefuse
| 2
| 1
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-05-11
|
2018-05-11 15:39:43
|
2018-05-11
|
2018-05-11 16:03:19
| 1
| true
|
en
|
2018-05-11
|
2018-05-11 16:03:19
| 0
|
182d1c8a57ff
| 3.245283
| 1
| 0
| 0
|
The data of the dead still hags around. What is it going to be used for?
| 5
|
Digital ghosts in the Data Science Age
Photo by David McEachan on Pexels
I found out about my ex-girlfriend’s death a year after it happened. After years together, I had dumped her. It was a hard breakup, the type in which you still care about the other person but realize they can’t provide the support you need to be happy. We, justifiably, never spoke again after she moved out. Then we went through all of the typical breakup motions of the post-digital age. We unfollowed each other. We unfriended each other. We untagged or deleted all of our pictures together. Most of our individual friends unfriended the other person. But that last step is rarely complete, and it is from one of these vestigial connections that I learned of her death.
The post was a heartfelt and sobering eulogy to the pain of a year without a best friend. It linked to my ex-girlfriend’s profile. This was before Facebook decided the expand the “Like” button to include multiple emotions, so the post had a lot of Likes. They were sympathy Likes, not enjoyment Likes. It was an impromptu memorial for all of her friends and family that I used to spend time with but would never see again. Facebook’s algorithm decided that this post was something I would be interested in.
After years together, I couldn’t help but click through to her profile. I learned how she died. She posted until the very end. I read the first-person account of years of suffering and pain and the nobility of a strong woman who knows she’s dying. I was too deeply engrossed in this bizarrely modern grief experience to realize that it was the worst sort of voyeurism. A work of digital snuff.
Her profile is still around, frozen in place. It is Facebook’s corporate property, indelibly preserved and monetized and traded on a daily basis. I wonder how many of Facebook’s clients would be sad at her death. I wonder how many would be furious to learn they’re wasting money marketing to a dead woman. Somewhere right now, someone is trying to develop an algorithm to classify dead versus living users. They will probably sell it as a value-added service to ensure that companies only market to people who can actually buy things. More likely, this algorithm already exists, and its associated startup has already been funded or purchased.
When we speak of data, it is almost always with language inspired by engineering and computer science and math. The tone of data science is almost always dispassionate and analytical, as one would expect from a discipline riding the border between computer science and statistics. Data and business present problems to be solved. Developers and mathematicians are expected to produce products that do work in a regular and predictable way, and when those products do not work, these developers are expected to analyze root causes and develop solutions quickly and efficiently. Algorithms are fixed and improved and their performance is benchmarked in terms of processor time and accuracy and precision.
And data scientists are doing remarkable things with all of the available social media data. Profiling and targeted marketing aren’t even particularly cutting-edge any more. Most people can be classified with fewer than 10 social media posts, and when paired with voluntary demographic data, those classifications are sufficient to predict products of interest. Segmentation and prediction are old news. Now people are training realistic chatbots with bulk data and making fake pornography with YouTube celebrity faces. These days, the money is in simulation, and with enough data, you can simulate anything.
You can simulate anyone.
Soon, data science will have all the tools to create digital doppelgangers at a time in which our communications and communities are increasingly digital. How often do you interact with people outside of the internet? How many friends have never met you in person? Would they know if a message from you was genuine? Would your parents?
Have you produced enough social media content to be simulated? What about your dead? Will Facebook eventually decide that my ex-girlfriend wants to video chat with me?
We are sprinting toward a future in which our dead can be resurrected as bots. They will be able to call and text us. They will video chat with us. They will allow us to move beyond (or revel in) our own grief. Our abusers will return to our lives in high fidelity. We will face long-forgotten objects of lust. Friends and family will see us and talk to us without our knowledge or theirs. We won’t know who is really talking to us and who is being simulated. Algorithms will decide if we want or need to be exposed to the ghosts of our pasts. They will try to sell us things. They will be owned.
|
Death in the Data Science Age
| 1
|
death-in-the-data-science-age-182d1c8a57ff
|
2018-05-23
|
2018-05-23 15:23:08
|
https://medium.com/s/story/death-in-the-data-science-age-182d1c8a57ff
| false
| 807
| null | null | null | null | null | null | null | null | null |
Data Science
|
data-science
|
Data Science
| 33,617
|
Zach Scott
|
Medical Informatician. I study how our data and the platforms that collect it affect our lives, our minds, and our health.
|
e5168a12ebf7
|
zbohannan
| 20
| 6
| 20,181,104
| null | null | null | null | null | null |
0
|
conda env create -f environment.yaml
//Instead of using
model.fit()
// Use
model.fit_generator()
python drive.py model.h5
| 3
|
767ba779a453
|
2018-03-04
|
2018-03-04 08:56:20
|
2018-03-11
|
2018-03-11 12:01:00
| 22
| false
|
en
|
2018-03-12
|
2018-03-12 10:05:28
| 8
|
182ee407dbc5
| 7.185849
| 49
| 2
| 1
|
This is a Udacity Self-Driving Car Nanodegree Project and all sources are belongs to Udacity. In this test I used my MacPro 15" with 2.9GHz…
| 5
|
Behavioral Cloning (Udacity Self Driving Car Project) — Generator Bottleneck Problem in using GPU
This is a Udacity Self-Driving Car Nanodegree Project and all sources are belongs to Udacity. In this test I used my MacPro 15" with 2.9GHz CPU and Sonnet Breakaway Box with Geforce 1060 6GB GPU.
1. What you expect for this paper
How to use simulator data for training car autonomously like a real world.
How to train a CNN Model with and without a generator.
How important data preprocessing before training.
How to use data augmentation.
How affect generator in training phase using GPU. Bottleneck Problem.
2. Sources:
Project starter code link : https://github.com/udacity/CarND-Behavioral-Cloning-P3
Setup of environment for the project please download this link CarND Term1 Starter Kit. (I didn’t use this Kit that explained below.)
I use Conda (https://www.continuum.io/downloads) with Keras 2.1.4 and Tensorlow-gpu 1.5.0. I upload my ‘environment.yaml’ with simulator and data in below link. if you want to use my environment just run this code after downloading.
Link for the environment.yaml, simulator and data: https://drive.google.com/drive/folders/1_a8tmhlsrWfp_vyi6uN_js6RqupJ8zFl?usp=sharing
3. Project Codes:
Before coding. I want to explain a bit of project. Data file which include images and driving_log.csv file that path of the images. Each frame has 3 images which comes from left camera, centre camera and right camera. So, when reading driving_log.csv file each row has ‘center, left, right, steering, throttle, brake, speed’ columns. In this project we use camera images (center, left, right) as input and steering as target.
At first import needed packages for the project.
3.1. Load Images
I use two function to load images, first one is read csv file and load by line by line. Return ‘lines[:1]’ because first line is header. Second one is load the images for using lines path. Steering (angle) is for centre image. So, if image is right image angle is a bit small value (-0.2), and right add a bit (+0.2).
As mentioned above, data file included ‘driving_log.csv’ file and ‘IMG’ folder.
First, I load the paths (not images) . And then split paths into two parts. One is for train and the other is for validation. You may asked, why not use test data? We don’t need to test images because it is a regression problem not classification.
Now load images in their paths.
Total train images (if loaded three camera images of frame), total samples of train is 19284 and validation is 4820.
Plotting these images give some intuition about data.
As you see data is not a normal distribution or gaussian. If we load this data to network, -0.2, 0, +0.2 degrees overweighted the other steering angles.
I think we have two options. First one is when I load the data to network I cut overweighted degrees (as mention above) make it all degrees have normal or gaussian distribution. It seems a good option, however when thinking machine learning intuition it is not a good option, why? Because we lost some of the data and load to network very low data might makes it overfitted.
Second option is generate a new, augmented data. It is like a opposite of the first option. Instead of cut the overweighted steering angles, this time we augmented the data that the other steering angles have more that and get nearly gaussian distribution. So I choose this option.
3.2. Data Augmentation
How to augment data, I have some options:
-Load left and right images (Like did it above),
-Make some images flip,
-Make some images brighten.
Also I have more options that like scale, zoom or crop images. However this are change the shape of the image that I don’t need on this regression project. Although Keras have data function for augmentation process (https://keras.io/preprocessing/image/), I create my data augmentation function.
At this point we examine problem in two different manner, first one include generator the other not. Generators lets low memory usage. For more discussion for it is usage check this link: https://stackoverflow.com/questions/1756096/understanding-generators-in-python
3.2.1. Data Augmentation with Generator
Randomly take center, right or left image and each image again randomly augmented it brighten or flipped.
As it is a generator function we make a generator for use it.
Make some samples to plot. Now I select batch size is 1024 that generator generates 1024 images for train and validation.
Now, let plot images as histogram again.
It is not perfect but usable than its raw version.
3.2.2. Data Augmentation without Generator
Data augmentation part is nearly same as generator. With generator sample images are adjusted in train phase (we explain below), but without generator we should adjust it here.
I augmented nearly 14.100 image in 115 sn. Of course I use much more memory this time.
Let’s also plot images.
It looks gaussian. Nice because it makes car more straight drive. We can improve it but I think it is enough for training.
3.3. Model Selection
I used classical CNN architecture that include;
First, crop unneeded part of the image (Because, at the top of all the images are sky-mountain-forest and the hood of the car at the bottom),
Resize image to a small one like a Le-Net architecture,
Normalize the data,
Use 5 convolution layers with batch normalization and activation ELU,
1 Flatten layer and 3 dense layer with activation ELU and using dropout to not overfit the network,
Optimizer Adam with learning rate 1e-5 instead of default 1e-3.
This is the model that we are going to use;
So input image shape is 60x320x3 and output is just a steering prediction.
Total parameter is more than 2 million. But it is not a big number if you use GPU.
3.4. Train Model
We train model in same model with/without generator.
3.4.1. Train Model with Generator
Model train is a bit different when using generator. (https://keras.io/models/sequential/)
In training phase I setup ‘steps_per_epoch= len(train_samples)’ that means 6500 images used in this train. if you want to train more; like make it double which are nearly 13000 images, just multiply ‘2*len(train_samples)’. Don’t forget that without generator more than 14000 training images.
Because it takes more time to train with generator, I entered only ‘1’ to ‘epochs’ (epochs =1). Oh my God! It is more than what I predicted. Just one epoch takes more than 12 minutes with nearly 6500 images (In my first try it takes 34 minutes!!!).
As to it is a regression problem accuracy is trivial, aim is minimize the loss.
3.4.2. Train Model without Generator
I entered 50 for epochs (epochs=50). Each epoch takes nearly 20 sec. and total training time is nearly 17 minutes.
As to it is a regression problem accuracy is trivial, aim is minimize the loss.
4. Results After Training
To run simulator run this code;
The result of full training after preprocessing and data augmentation;
The result of full training no preprocessing (except cropping and normalization in model). It is like a drunk car and at the end exit from the route :
5. Conclusion
In this paper we take a simulator and data which is prepared for Udacity Self-Driving Car Nanodegree and train it in CNN architecture as a regression problem.
We train the model with/without generator. Compare to with generator, without generator have double images and epoch/sec is 20sec. which is 720sec. in with generator. It means without generator is nearly 72 times faster than with generator. (Sometimes it goes up 200 times!!!).
The problem is also about ‘time efficiency’ versus ‘memory or space efficiency’ problem. If have enough memory to allocate and want to effective usage of GPU selection without generator is understandable. However, training a huge data not possible to allocate all in memory makes to use generator.
With Generator
So in this data and model that using generator with GPU, bottleneck problem occurred. Of course it depends on data size, batches, model, generator function and computer CPUs-GPUs. It is just a example illustration of course it is more complex than that.
Without Generator
As you see on above videos, data preprocessing is so important. It is a good example for ML experts saying, ‘Garbage In Garbage Out’.
Because of the not enough data we make a data augmentation, which is coded by us. However we know that Keras has a data augmentation function.
Github : https://github.com/bahadir60/Udacity-Self-Driving-Car/tree/master/Behavioral_Cloning
If you like this post please click ‘Clap’ below to support.
|
Behavioral Cloning (Udacity Self Driving Car Project) — Generator Bottleneck Problem in using GPU
| 604
|
behavioral-cloning-udacity-self-driving-car-project-generator-bottleneck-problem-in-using-gpu-182ee407dbc5
|
2018-06-20
|
2018-06-20 13:21:11
|
https://medium.com/s/story/behavioral-cloning-udacity-self-driving-car-project-generator-bottleneck-problem-in-using-gpu-182ee407dbc5
| false
| 1,414
|
As a leading Artificial Intelligence (AI) community in Turkey, we are producing content about AI in both Turkish and English.
| null |
deeplearningtr
| null |
Deep Learning Turkey
|
deeplearningturkiye@gmail.com
|
deep-learning-turkey
|
ARTFICIAL INTELLIGENCE,DEEP LEARNING,MACHINE LEARNING,DATA SCIENCE
|
deeplearningtr
|
Machine Learning
|
machine-learning
|
Machine Learning
| 51,320
|
Bahadır YILMAZ
|
SDC, ML, DL, RL — PhD Candidate
|
bf6a4b6b7e40
|
bahadiryilmaz32
| 46
| 3
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-04-21
|
2018-04-21 18:45:04
|
2018-04-22
|
2018-04-22 07:36:24
| 2
| false
|
en
|
2018-04-22
|
2018-04-22 07:36:24
| 6
|
18312eb65e2f
| 0.734277
| 8
| 0
| 0
|
You’ve been waiting for this for a long time, we’re finally listed on CoinMarketCap (CMC)! The volume reached 100k, which allowed us to…
| 5
|
DAN is now listed on CoinMarketCap!
You’ve been waiting for this for a long time, we’re finally listed on CoinMarketCap (CMC)! The volume reached 100k, which allowed us to reiterate our request for listing on CMC. As soon as the listing was effective, the price went up by more than 300%!
This is another step forward! We continue to do our best to list the DAN on other exchanges, now that we are on CMC, that should open doors for us!
Stay tuned:
Twitter: https://twitter.com/daneelproject
Telegram: t.me/DaneelCommunity
Facebook: https://www.facebook.com/daneelproject
LinkedIn: www.linkedin.com/company/11348931/
Reddit: https://www.reddit.com/r/Daneel_Project/
|
DAN is now listed on CoinMarketCap!
| 142
|
dan-is-now-listed-on-coinmarketcap-18312eb65e2f
|
2018-05-07
|
2018-05-07 00:10:33
|
https://medium.com/s/story/dan-is-now-listed-on-coinmarketcap-18312eb65e2f
| false
| 93
| null | null | null | null | null | null | null | null | null |
Token Sale
|
token-sale
|
Token Sale
| 15,244
|
Daneel Assistant
|
Your future personal crypto assistant ! https://daneel.io
|
dc883054551c
|
daneel_project
| 463
| 9
| 20,181,104
| null | null | null | null | null | null |
0
| null | 0
| null |
2018-06-11
|
2018-06-11 07:35:58
|
2018-06-11
|
2018-06-11 04:22:30
| 3
| false
|
en
|
2018-06-11
|
2018-06-11 07:35:59
| 6
|
1831b683cb36
| 2.289623
| 0
| 0
| 0
| null | 5
|
Introducing “NEVA” by NICE — The First Robotic Virtual Attendant for Your Employees
One of the leading providers of on-premises and cloud-based enterprise software, NICE recently introduced their latest innovation for the productivity market. NEVA or the “NICE employee Virtual Assistant” is the very-first virtual attendant to hit the industry that was designed specifically for employee empowerment. The system is powered by NICE’s desktop automation technology and offers real-time process guidance for staff members.
With NEVA, NICE hopes that employees will have the support they need to execute complex request efficiently, upsell products, and even cross-sell to customers at the most lucrative times in a conversation. The result? Better employee engagement, greater confidence, and even stronger regulatory compliance.
The Latest Solution in Robotic Process Automation
Meet NEVA
NICE’s new delivery in RPA and virtual attendant technology supports the belief that robotic entities and human agents can work more efficiently together. NEVA’s intuitive interface gives employees the opportunity to request assistance when they need it, or even access contextual guidance during a sales conversation. What’s more, the intelligent decisioning engine translates employee requests into structured workflow actions automatically on the behalf of each employee — interacting with advanced desktop systems.
By pulling together scripts and extracting hidden data from back-end systems, NEVA helps to connect various loops in the transaction process, helping employees to achieve stronger results. What’s more, with NEVA, it will be much easier for business leaders to provide new hires with on-the-job training and guidance through their initial calls — reducing the need for cumbersome classroom settings. Thanks to automatic reminders, NEVA can even help to boost compliance.
Leading Technology for the Modern Workforce
According to the President of the NICE Enterprise Product Group, Miki Migdal, the NEVA release is a huge step forward for the company, helping them to bring new innovation to the customer service domain, and the global business landscape. NICE hopes that NEVA will help them make their mark on the changing workforce, by support leaders and employees alike.
NEVA is based on the NICE RPA platform, which earned the title of “Leader” in the Everest Group RPA product Peak Matrix. According to the Practice Director of the Everest Group, Amardeep Modi, the NICE RPA system achieved exceptional scores on Market Impact, and Vision & Capability, making it a Star performer for the PEAK matrix. The scope of the RPA product from NICE is sure to support NEVA in serving employees throughout the enterprise.
NEVA can even help out with implementing repetitive and routine tasks for the modern team, making customer service faster and more efficient, as well as more capable of adhering to company policies. The result will be higher productivity, improved process accuracy, and increased customer satisfaction.
|
Introducing “NEVA” by NICE — The First Robotic Virtual Attendant for Your Employees
| 0
|
introducing-neva-by-nice-the-first-robotic-virtual-attendant-for-your-employees-1831b683cb36
|
2018-06-11
|
2018-06-11 07:36:00
|
https://medium.com/s/story/introducing-neva-by-nice-the-first-robotic-virtual-attendant-for-your-employees-1831b683cb36
| false
| 461
| null | null | null | null | null | null | null | null | null |
Apac
|
apac
|
Apac
| 286
|
UC Today
|
Unified Communications Stories
|
bd51979d153c
|
uctoday
| 13
| 74
| 20,181,104
| null | null | null | null | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.