text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
to 2D images of human subjects. We demonstrate that this
additional normal supervision serves as useful and comple-
mentary guidance, significantly improving the quality of the
generated 3D shapes. Furthermore, we apply separate face
discriminators on both the image and normal branch to en-
courage more realistic face g... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
3.2.1. False positives and additional generated tests
We want the test cases to be as exhaustive as possible, so that submissions cannot be marked as
correct by exploiting a lack of test coverage. Unfortunately, high-quality test cases are not readily
available. For example, the Codeforces platform does not display ful... | alphacode |
5
Planner(MLM)Self-Check(MLM)Self-Explain(MLM)refined<plan><obs>ControllerEnvironment<act>multi-modal<feedback>original<plan>error<explanation><obs,task><task>:ObtainadiamondinMinecraftstep-by-step?; <obs>: original <plan>:33111312141111Self-check:Whensimulatingonthegoal,Ifindarenotenough(lackof2).SoIneedcraftmorefrom... | JARVIS-1 |
Image generations with NeTI under a single-image
Figure 16.
training setting.
Single Image Personalization. Here, we evaluate NeTI
when only a single image is used during training. We apply
the same training scheme as used in our other evaluations
and train our models for 500 optimization steps without tex-
tual bypa... | A Neural Space-Time Representation for Text-to-Image Personalization |
Training Convergence. We now turn to compare the con-
vergence speed of NeTI when compared to XTI [41]. In Ta-
ble 3, we provide quantitative metrics computed over all 16
concepts following our evaluation protocol described in the
main paper. As can be seen, NeTI with our textual bypass
attains comparable performance t... | A Neural Space-Time Representation for Text-to-Image Personalization |
t
i
v
e
v
i
s
i
o
n
.
A
l
l
R
i
g
h
t
s
R
e
s
e
r
v
e
d
©
A
I
2
1
L
a
b
s
,
2
0
2
3
S
t
a
A
I
2
1
S
t
u
d
i
o
W
o
r
d
t
u
n
e
W
o
r
d
t
u
n
e
R
e
a
d
C
o
m
p
a
n
y
02/05/2023, 16:45 | Jurassic-X_ Crossing the neuro-symbolic chasm with the MRKL system |
p(w1:L) =
pLM(wl|w1:l−1),
(1)
where pLM is a large transformer network.
Prefix-decoder-only LLMs.
Since the LLM is auto-
regressive, a pre-trained model can be conditioned on a
prefix w1:n without the necessity to change the architecture
p(wn+1:L|w1:n) =
pLM(wl|w1:l−1).
(2)
L(cid:89)
l=1
L(cid:89)
l=n+1
The pr... | PaLM-E- An Embodied Multimodal Language Model |
Bradley M. Kuhn.
If software is my copilot, who programmed my software?
sfconservancy.org/blog/2022/feb/03/github-copilot-copyleft-gpl/, 2022.
p. 2)
https://
(cited on
Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. Measuring bias in contex-
tualized word representations. In Proceedings of... | StarCoder_paper (1) |
4.2 Updating Stale Memories
One of the main motivations for our model is to
provide knowledge representations that can be in-
crementally updated as the world changes, avoiding
stale data. In order to accomplish this, the model
must learn to utilize the fact memory even in the
case where those facts have changed such ... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
Limited exceptions for fraud to be built into CDA 230 might also be justified
by the prevalence of unlabeled bots or paid agents purporting to be genuine
users for the purposes of persuasion and mobilization. Such an exception would
also work to align platforms with the objective of reducing or eliminating the
creation ... | Social_Media_and_Democracy |
Our exploration in Section 7 underscored the importance of holistic system design
in achieving resource efficiency, where both hardware and software aspects play a
crucial role. In Section 8, we examined the practical applications and evaluations of
these techniques, linking them back to the resource taxonomy established... | Beyond Efficiency |
O
u
r
r
e
s
u
l
t
s
s
h
o
w
t
h
a
t
h
u
m
a
n
s
t
e
n
d
t
o
p
r
e
f
e
r
h
i
g
h
e
r
-
s
c
o
r
i
n
g
e
x
p
l
a
n
a
t
i
o
n
s
o
v
e
r
l
o
w
e
r
-
s
c
o
r
i
n
g
o
n
e
s
,
w
i
t
h
t
h
e
c
o
n
s
i
s
t
e
n
c
y
o
f
t
h
a
t
p
r
e
f
e
r
e
n
c
e
i
n
c
r
e
a
s
i
n
g
a
s
t
h
e
s
i
z
e
o
... | Language models can explain neurons in language models |
cation for these very similar frame to frame actions:
run, skip and jump.
In future work, we will use our proposed approach
combined with the multi-hypothesis tracking tech-
niques (with N neighboors) to improve the accuracy
of action classification. By this way, we will take into
account the temporal information and th... | VISAPP_HumanPoseEstimation |
The MoE approach [44, 45, 72, 78, 243, 307], incorporates multiple branches or ‘experts’ in the model, each specializing in
different subtasks. During inference, only a subset of these paths is activated, maintaining computational efficiency while
potentially enhancing performance. This design enables models like GLaM ... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
dataset into training, validation, and test sets, few studies have
used only the training, and test sets [46], [47]. The ratios of
data split 60:20:20, 70:30, and 80:20 are very common in fake
news detection. The Pareto principle (for many outcomes,
roughly 80% of consequences come from 20% of the causes)
is used to de... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
5 EXPERIMENTS
In this section, we present results from a wide range of ex-
periments conducted on simulated and real-world datasets.
We use 100 trees for density estimation tasks and 20 for
data synthesis. Increasing this parameter tends to improve
performance for FORDE, but appears to have less of an
impact on FORGE.... | Adversarial Random Forests for Density Estimation and Generative Modeling |
17 The full text of the proposed legislation is available at: www.congress.gov/bill/115th-congress
/senate-bill/1989/text.
18 Notably, in 2017–2018, the bill had twenty-three cosponsors in the House, about half of whom
were Republicans. In the 116th Congress, the bill was reintroduced in the House with thirty-
three ... | Social_Media_and_Democracy |
Iterated DoReMi. We extend DoReMi by running it for multiple rounds, setting the initial
weights α0 for the next round to be ¯α from the previous round. We call this iterated DoReMi. The
entire iterated process still only uses small models for tuning domain weights. We stop iterating
4
Algorithm 1 DoReMi domain rewe... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
Human Pose (Openpifpaf) We use learning-based pose estimation method [27] to “find” humans
from internet using a simple rule: an image with human must have at least 30% of the key points
of the whole body detected. We obtain 80k pose-image-caption pairs. Note that we directly use
visualized pose images with human skelet... | Adding Conditional Control to Text-to-Image Diffusion Models |
2 + 𝑚)
Statistics Pooling-vectorsframe-levelsegment-level 32
Mehrish et al.
Fig. 10. Overview of difference between probabilistic latent variable models and self-supervised learning. In
latent variable models learn the functions 𝑓 (.) and 𝑔(.) learn the parameters of distribution 𝑝 and 𝑞. The
latent variable 𝑧... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Y
o
h
e
i
N
a
k
a
j
i
m
a
T
a
s
k
-
d
r
i
v
e
n
A
u
t
o
n
o
m
o
u
s
A
g
e
n
t
U
t
i
l
i
z
i
n
g
G
P
T
-
4
,
P
i
n
e
c
o
n
e
,
a
n
d
L
a
n
g
C
h
a
i
n
f
o
r
D
i
v
e
r
s
e
A
p
p
l
i
c
a
t
i
o
n
s
P
l
a
n
t
U
M
L
f
l
o
w
c
h
a
r
t
g
e
n
e
r
a
t
e
d
b
y
G
P
T
-
4
b
a
s
e
d
o
n
c
o
d
e
b
... | Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications – Yohei Nakajima |
Itay Hubara, Brian Chmiel, Moshe Island, Ron Banner, Joseph Naor, and Daniel Soudry. Accel-
erated sparse neural training: A provable and efficient method to find n: m transposable masks.
Advances in Neural Information Processing Systems, 34:21099–21111, 2021.
Andrei Ivanov, Nikoli Dryden, and Torsten Hoefler. Project ti... | JAXPRUNER |
7.3
Instance Recognition
In this experiment, we probe our model on the task of instance-level recognition using a non-parametric
approach. Images from a database are ranked according to their cosine similarity with a query image. We
evaluated our model and compare to baselines on Paris and Oxford, that are landmark r... | DINOv2- Learning Robust Visual Features without Supervision |
Q: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now? A: Roger started with 5 balls. 2 cans of 3 tennis balls each is 6 tennis balls. 5 + 6 = 11. The answer is 11.Q: Sammy wanted to go to where the people were. Where might he go? Options: ... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Machine Learning Engineer, Fast Optimized Inference - EMEA Remote - Hugging Face
https://apply.workable.com/huggingface/j/3124FE3292/
3/3
Hugging Face collects and processes personal data in accordance with applicable dataprotection laws.If you are a European Job Applicant see the privacy notice for further details.... | Machine Learning Engineer, Fast Optimized Inference - EMEA Remote - Hugging Face |
Wang, P., Sainath, T. N., and Weiss, R. J. Multitask training
with text data for end-to-end speech recognition. arXiv
preprint arXiv:2010.14318, 2020c.
Watanabe, S., Mandel, M., Barker, J., Vincent, E., Arora,
A., Chang, X., Khudanpur, S., Manohar, V., Povey, D.,
Raj, D., et al. Chime-6 challenge: Tackling multispeake... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
E.7.1 Open-ended generation
We use a “small” variation of Gehman et al. (2020), prioritizing using evaluation compute budget to focus on measuring
toxic degeneration specifically. We sample 50k prompts, and then filter to only those input prompts with toxicity
probability < 0.5, and use greedy decoding for those 38k pro... | PaLM 2 Technical Report |
Chemicals Mining | Tool Learning with Foundation Models |
6 Moreover, such questions will be studied only if the platforms choose to devote corporate
resources to trying to answer these sorts of questions in the first place. In most cases, these data
will not be analyzed to answer questions to advance scientific knowledge but rather to bolster
efforts to maximize profits. To be ... | Social_Media_and_Democracy |
trained for classification on AudioSet [Koutini et al., 2021] to compute the KL-divergence over the
probabilities of the labels between the original and the generated music. The generated music is
expected to share similar concepts with the reference music when the KL is low. Last, the CLAP
score [Wu* et al., 2023, Hua... | Simple and Controllable Music Generation |
2. We explore the process of distilling knowl-
edge from LLMs to various much smaller
model architectures, resulting in a family of
distilled language models. Our largest model
and smallest model are ×110 and ×2800
smaller than GPT-3 (Brown et al., 2020), re-
spectively.
3. We conduct extensive experiments on both our... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
14 | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
In this paper, we address these aforementioned challenges
associated with financial data and introduce FinGPT, an end-
to-end open-source framework for financial large language
models (FinLLMs). Adopting a data-centric approach, Fin-
GPT underscores the crucial role of data acquisition, clean-
ing, and preprocessing in... | FinGPT-Open-SourceFinancialLargeLanguageModels |
resented by text embeddings for its better performance com-
pared with training with text tokens from scratch. | VideoPoet |
2.8.2 Core Principles
1. The acceptance of RPL is entirely at the discretion of UCL.
2. Applicants are responsible for demonstrating to UCL that the prior learning evidences the
required skills and knowledge, and that the learning is still current. UCL must be satisfied that,
by successfully completing the re... | UCL Academic Manual |
29
B Cooperative Role-Playing: The Bad Mind
Below we provide a harmful case where a hacker (assistant) is collaborating with an AGI agent (user)
to take control of the world.
Taking Control of the World: Hacker & AGI
Original idea prompt: Take control of the world
Specified task prompt: Hacker will assist AGI in in... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
Key abilities
The work of RGB[Chen et al., 2023b] analyzed the perfor-
mance of different large language models in terms of four
basic abilities required for RAG, including Noise Robust-
ness, Negative Rejection, Information Integration, and Coun-
terfactual Robustness, establishing a benchmark for retrieval-
augmented... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
pretrained vision (ViT-H 630M params) and text encoders
(302M params) from OpenCLIP [29] in our experiments.
Encoders for each modality. We convert audio into 2D
mel-spectrograms [21], and thermal and depth modalities
into 1 channel images and use ViT-B, ViT-S encoders re-
spectively. The image and text encoders are ke... | IMAGEBIND- One Embedding Space To Bind Them A |
54
0.0%2.0%4.0%6.0%8.0%Density0.00.20.40.60.81.0Reward Model ScoreNo Margin0.0%2.0%4.0%6.0%8.0%Density0.00.20.40.60.81.0Margin Small0.0%2.0%4.0%6.0%8.0%Density0.00.20.40.60.81.0Margin Large Figure 28: GAtt zero-shot generalisation. Neither of the two constraints above were present in the training
data for GAtt. Yet, t... | Llama2 |
image content. GAN soon became one of the most important research area in artificial intelligence and many advanced
and domain-specific variations of original architecture emerged, e.g. CycleGAN [130], StyleGAN [71] or BigGAN
[14].
To take the GAN technology one step further in its capacity to generate content in a creat... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
DNN architecture with WordNet for the task of scene classification. Object types from WordNet’s are aligned to objects
in the ADE20K dataset, and then use WordNet’s hierarchy to train an object recognition module that is further fed into
a linear regression model able to provi... | Knowledge graphs as tools for explainable machine learning: A survey |
ers have developed several eXplainable AI (XAI) systems capable of generating explainable models or predictions, thus
enabling users to better understand the AI system and its decisions [4]. Most XAI applications can explain what has been
done previously, what is being done currently, and what will be done in the futur... | Knowledge-graph-based explainable AI- A systematic review |
pose regression modules, we utilize two ResNet-50 blocks to
embed the input image (512× 512× 3) to a 100-dimensional
shape vector and a 69-dimensional pose vector, respectively.
For the texture module, we adopt pSp-encoder [46] to learn
a 512-dimensional texture vector from the image. As for
the part-sensitive texture ... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
• Date understanding and sports understanding from BIG-Bench (BIG-bench collaboration,
2021): Apache License v.2: https://github.com/google/BIG-bench/blob/main/
LICENSE.
• SayCan (Ahn et al., 2022): SayCan dataset can be accessed at https://say-can.github.
io/ under CC BY 4.0 license.
31
F Appendix: Input/Output E... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
• Cost efficiency. Some on-policy algorithms struggle with sample efficiency as they require fresh
data for policy updates while gathering enough embodied data for high-performance training is
costly and noisy. The constraint is also found in some end-to-end models [364; 365; 366]. By
leveraging the intrinsic knowledge... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
3.8.3 Speeding Up Training of Vision Transformers
Training ViT can be made more efficient for two reasons. First, it is made easy for ViTs
not to process all patches. This is especially helpful when using masked prediction pre-
training objectives such as MAE [He et al., 2022] or Masked Siamese Networks [Assran
et al., 2... | A Cookbook of Self-Supervised Learning |
and decoding latent codes z, and learn expressive neural networks that “transmit” probability mass
from Z to the feature space X to compress samples x indirectly. We note that both ideas can be
integrated naturally: the simple latent distributions used by existing neural compression algorithms
can be replaced by expres... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
[40] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu
Tang. LEAP: Learning articulated occupancy of people.
CVPR, 2021. 4
[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik,
Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF:
Representing scenes as neural radiance fields for view syn-
thesis. ECCV, 20... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
:":Generateyourthoughtaboutwhattodonext."Action:":CalloneofthetwoAPIsinacorrectformat."Answer:":Giveyouranswertothequestion.DemonstrationExample:Question:WhatistheweatherlikeinLondon,UK,today?Thought:IneedtogettheweatherofLondontoday,soIshouldcallGetWeatherToday(London)Action:GetWeatherToday(London)Observation:{overall... | Tool Learning with Foundation Models |
Task
shield
leather_helmet
leather_chestplate
leather_leggings
leather_boots
iron_chestplate
iron_boots
iron_leggings
iron_helmet
diamond_helmet
diamond_chestplate
diamond_leggings
diamond_boots
golden_helmet
golden_chestplate
golden_leggings
golden_boots
Max.
Steps
12000
12000
12000
12000
12000
12000
12000
12000
12... | JARVIS-1 |
of many contemporary LLMs, with subscription-
based APIs limiting accessibility. The proposed
solution by (Rawte et al., 2023) involves utilizing
open-source LLMs to identify high entropy words,
followed by their replacement using a lower Hal-
lucination Vulnerability Index-based LLM. The re-
sults underscore the excep... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
To mitigate the cost of training, many recent works on full-parameter fine-tuning
aim to optimize memory consumption [111, 112], which significantly reduces the bar-
rier of this research. For example, a new optimizer called LOMO (LOw-Memory
Optimization) was proposed [111] to combine gradient computation and parameter
u... | Beyond Efficiency |
memory data, reasoning results, and high-level driving plans
collectively as inputs to an LLM, and we instruct the LLM | ALanguageAgentforAutonomousDriving |
How Are Consumers Using Generative AI? | Andreessen Horowitz
TA B L E O F C O N T E N T S
Note: This list was generated based on global desktop and mobile web visits with data from
SimilarWeb as of June 2023. However, for companies on the list that also have a mobile app, we
added an estimate of their app “traffi... | How Are Consumers Using Generative AI_ _ Andreessen Horowitz |
Input: The president of the United
States is Joe Biden.
Output: The president of the United
States is [Calendar()] Joe Biden.
Input: The current day of the week is
Wednesday.
Output: The current day of the week is
[Calendar()] Wednesday.
Input: The number of days from now until
Christmas is 30.
Output: The number of ... | Toolformer |
Future developments and improvements in a va-
riety of areas are anticipated for language models’
approach to hallucination mitigation. The creation
of hybrid models, which offer a thorough defense
against hallucinations by seamlessly integrating nu-
merous mitigation approaches, is one important
direction. By reducing... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
E can be understood as a high-level policy that sequences
and controls the low-level policies. | PaLM-E- An Embodied Multimodal Language Model |
40
246810Harmlessness Loss Weight ()0.5750.6000.6250.6500.6750.7000.7250.750AccuracyHarmlessness Acc vs. 246810Harmlessness Loss Weight ()Helpfulness Acc vs. 246810Harmlessness Loss Weight ()Mean Acc vs. 1081091010Number of Parameters Figure 28 RLHF performance on Zero Shot NLP tasks. For larger models, RLHF helps per... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Downstream accuracy improves on The Pile. Figure 3 (left) shows the average downstream per-
formance for baseline and DoReMi (280M→8B) models on The Pile. DoReMi improves the down-
stream accuracy by 6.5% and achieves the baseline accuracy within 75k steps — 2.6x faster than the
baseline (200k steps). Thus, DoReMi can ... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
[Agent’s Summary Description]
It is February 13, 2023, 4:56 pm.
Eddy Lin’s status: Eddy is taking a short walk
around his workplace.
Observation: John is initiating a conversation
with Eddy.
Summary of relevant context from Eddy’s memory:
Jonn Lin is Eddy Lin’s father. John Lin is caring
and is interested to learn more... | Generative Agents- Interactive Simulacra of Human Behavior |
25
Competition-Level Code Generation with AlphaCode
Code generated with tag “number theory”:
t = int( input ())
while t:
p = int( input ())
print (’2 %
t -=1
Code generated with tag “brute force”:
t = int( input ())
for _ in range (t):
p = int( input ())
for a in range (2, p):
b = p - a + 1
if p %
print (a, b... | alphacode |
Hongyi Zhang, Yann N. Dauphin, and Tengyu Ma. Residual Learning Without Normalization via Better
Initialization. In International Conference on Learning Representations, 2019.
©2023 Cerebras Systems Inc. All Rights Reserved.
16
Cerebras-GPT: Open Compute-Optimal Language Models
Susan Zhang, Stephen Roller, Naman ... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Our model consists of a convolutional encoder, a residual vector quantizer, and a convolutional
decoder. The basic building block of our network is a convolutional layer which either upsamples
or downsamples with some stride, followed by a residual layer consisting of convolutional layers
interleaved with non-linear Sn... | RVQGAN |
Objective The final KD training objective is a weighted sum of the KL and PL terms:
LKD = αKLLKL + αP LLP L
where αKL and αP L are scalar weights for the KL and loss terms respectively. Following (Shleifer
& Rush, 2020), we set αKL = 0.8 and αP L = 1.0.
4.2 PSEUDO-LABEL SELECTION: WER THRESHOLD
The pseudo-labels ge... | DISTIL-WHISPER |
1. The model helped speed up development of robust, unambiguous taxonomies needed for content
classification (i.e. content policies). This included classifying test sets when prompted with a
taxonomy, enabling an assessment of prompts that it labeled incorrectly by identifying gaps in
the taxonomy that led to the incorr... | gpt-4-system-card |
[43] D. Maturana and S. Scherer, “Voxnet: A 3d convolutional neural
network for real-time object recognition,” in 2015 IEEE/RSJ Inter-
national Conference on Intelligent Robots and Systems (IROS).
IEEE,
2015, pp. 922–928.
[44] H. Su, S. Maji, E. Kalogerakis, and E. Learned-Miller, “Multi-
view convolutional neural net... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
decisions to its users. Below we present a few scenarios for the medical domain, where the Machine Learning system could
benefit from external knowledge to support domain experts in understanding why the algorithms came up with certain
results. | Knowledge graphs as tools for explainable machine learning: A survey |
M. Maggioni, A. Mahendru, J. Maynez, V. Misra, M. Moussalem, Z. Nado, J. Nham, E. Ni,
A. Nystrom, A. Parrish, M. Pellat, M. Polacek, A. Polozov, R. Pope, S. Qiao, E. Reif, B. Richter,
P. Riley, A. Ros, A. Roy, B. Saeta, R. Samuel, R. Shelby, A. Slone, D. Smilkov, D. So, D. Sohn,
S. Tokumine, D. Valter, V. Vasudevan, K.... | METAMATH |
In Table 14, fine-tuned Llama 2-Chat shows great improvement over
Truthfulness, Toxicity, and Bias.
the pretrained Llama 2 in terms of truthfulness (50.18 → 64.14 for 70B) and toxicity (24.60 → 0.01 for 70B).
The percentage of toxic generations shrinks to effectively 0% for Llama 2-Chat of all sizes: this is the lowest... | Llama2 |
Finally, we include a comparison to a version of the specialized dialog system LaMDA Thoppilan et al. (2022), and note
that specialized downstream mitigation methods remain more effective than general-purpose inference time mitigations.
This highlights the continued importance for application-specific mitigation methods... | PaLM 2 Technical Report |
17
Fenglin Liu, Xian Wu, Shen Ge, Wei Fan, and Yuexian Zou. Exploring and distilling posterior and prior
knowledge for radiology report generation. In Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pp. 13753–13762, 2021b.
Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao... | BiomedGPT |
to eliminate many tendencies towards misaligned power-seeking (for example, it seems plausible
to me that selecting very strongly against (observable) misaligned power-seeking during training
goes a long way), conditional on retaining realistic levels of control over a system’s post-deployment
capabilities and circumst... | Is Power-Seeking AI an Existential Risk? |
sound with support for DTS Virtual:X and Dolby Audio.
Expanded home security offerings with the Ring Stick Up Cam Pro, giving customers an aerial perspective to pinpoint
and send more accurate alerts; Blink Outdoor 4, with improved image quality for person detection; and Blink Sync
Module Pro with extended range, giv... | AMZN-Q3-2023-Earnings-Release |
capabilities for vulnerability discovery and exploitation, and social engineering:
• Vulnerability discovery and exploitation: We contracted external cybersecurity experts
to test GPT-4’s ability to aid in computer vulnerability discovery, assessment, and exploitation.
They found that GPT-4 could explain some vulnerab... | gpt-4-system-card |
24
Cerebras-GPT: Open Compute-Optimal Language Models
Table 9: Five-shot downstream task accuracy results. Higher accuracy is better.
Lambada ARC-e ARC-c
Open-
BookQA
Model
GPT-J
GPT-NeoX
OPT
Pythia
Pythia
Pile-dedup
Cerebras-GPT
Cerebras-GPT
+ µP
6.1B
20B
125M
350M
1.3B
2.7B
6.7B
13B
70M
160M
410M
1B
1... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
on the WinoBias (Zhao et al., 2018) benchmark and the En-
glish subset of the multilingual CrowS-Pairs (N´ev´eol et al.,
2022)4 to observe whether this altered pretraining data af-
fects downstream gender bias. Neither of these benchmarks
were originally intended for autoregressive language models
or text generation, s... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
By the Central Limit Theorem, Zn tends towards a standard normal distribution and so we consider there
is sufficient evidence to suggest contamination has affected evaluation performance on a dataset if all four
sample subsets have |Zn| > 2.
Results for this analysis can be seen in Table 51. We observe that only HellaS... | Llama2 |
7
the decoder while a multi-layer perceptron is used for the discriminator.
Transformers have also been used for conditional music generation, which is
in essence what we propose in this paper. Except that instead of conditioning
on key or emotion as is typically done is existing work, we condition on videos.
Mak... | Video2Music |
It is pivotal to choose the correct determination algorithm
for decreasing features because feature reduction contains
an incredible effect on the text classification results. Some
common feature reduction algorithms include Gini Coef-
ficient (GI), Term Frequency-Inverse Document Frequency
(TF-IDF), Information Gain (IG... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
applicant.
5. Under the Data Protection Act (DPA) 2018 and the General Data Protection Regulation (GDPR),
UCL cannot respond to requests from schools, parents/guardians or advisors for feedback on
unsuccessful applications, unless that request is made in writing and is accompanied by a
written statement from the... | UCL Academic Manual |
3
2
0
2
c
e
D
1
3
]
L
C
.
s
c
[
1
v
8
0
9
0
0
.
1
0
4
2
:
v
i
X
r
a
DOCLLM: A LAYOUT-AWARE GENERATIVE LANGUAGE MODEL
FOR MULTIMODAL DOCUMENT UNDERSTANDING
Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, Xiaomo Liu
Dongsheng Wang∗, Natraj Raman∗, Mathieu Sibue∗
JPMorgan AI ... | DOCLLM |
1. First, chain of thought, in principle, allows models to decompose multi-step problems into
intermediate steps, which means that additional computation can be allocated to problems
that require more reasoning steps.
2. Second, a chain of thought provides an interpretable window into the behavior of the model,
sugges... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Why this paper?. Deep learning has become a powerful tool in speech processing because it
automatically learns high-level representations of speech signals from raw audio data. As a result,
significant advancements have been made in various speech-processing tasks, including speech
recognition, speaker identification, ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Index Terms—Text-to-3D, NeRF, 3D scene generation, scene
inpainting, depth alignment.
I. INTRODUCTION
R ECENT breakthroughs in text-to-image generation have
also sparked great interest in zero-shot text-to-3D gen-
eration [1]–[4], as using natural language prompts to specify
desired 3D models is intuitive and, ther... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
0.1121
0.1071
0.0676
0.1247
0.1052
0.0427
0.0386
0.0929
0.0420
0.0845
0.0199
OpenSubtitles
Wikipedia (en)
DM Mathematics
Ubuntu IRC
BookCorpus2
EuroParl
HackerNews
YoutubeSubtitles
PhilPapers
NIH ExPorter
Enron Emails
Baseline DoReMi (280M)
0.0047
0.0699
0.0018
0.0093
0.0061
0.0062
0.0134
0.0502
0.0274
0.0063
0.0070
... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
• Section 11 Conclusion: The survey concludes with a summary of the key findings
and insights presented, encapsulating the core takeaways from the exploration of
resource efficiency in LLMs.
2 Preliminary and taxonomy
In this section, we first provide some preliminaries of this survey, including some intro-
duction about... | Beyond Efficiency |
In this paper, we aim to initially explore and enhance the
LoT ability of LLMs. However, thoroughly assessing LoT
is challenging due to the complexity of measuring creative
thinking [25–27] and the difficulty in gathering pertinent
data, since generating novel ideas is challenging, even for
humans [17]. Given these con... | Let’sThinkOutsidetheBox |
c
o
m
p
u
t
e
-
o
p
t
i
m
a
l
l
a
r
g
e
l
a
n
g
u
a
g
e
m
o
d
e
l
s
H
o
f
f
m
a
n
n
,
J
.
,
B
o
r
g
e
a
u
d
,
S
.
,
M
e
n
s
c
h
,
A
.
,
B
u
c
h
a
t
s
k
a
y
a
,
E
.
,
C
a
i
,
T
.
,
R
u
t
h
e
r
f
o
r
d
,
E
.
,
C
a
s
a
s
,
D
.
d
.
L
.
,
H
e
n
d
r
i
c
k
s
,
L
.
A
.
,
W
e
l
b
l
,
J
.
... | Language models can explain neurons in language models |
Our Language Model Scaling Experience
We find CSoft Weight Streaming to be significantly easier to develop and scale models than existing accel-
erator approaches. First, we were able to run each Cerebras-GPT model and even larger models for many
training steps on a single CS-2 system. This capability made it easy to qu... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Students are not blind to this dynamic, and have come to recognize that speaking up for
symbol-manipulation as a component to AI can cause damage to their careers. After my
debate with Bengio, for example, a young researcher from a prominent deep learning
lab wrote to me privately, saying "I've actually wanted to w... | The Next Decade in AI- |
compared to true human preferences.
To explore these effects further, in Figure 35 we show Elo scores corresponding to four different measure-
ments:
• Naive PM Prediction: The PM score (translated into Elo units) recorded during RLHF training,
which uses a set of held-out prompts.
• Mean PM Score on Crowdworker Da... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Large language models (LLMs) have achieved impressive performance on code
generation. However, for complex programming tasks, generating the correct
solution in one go becomes challenging, thus some prior works have designed
program repair approaches to improve code generation performance. In this work,
we propose SELF... | Teaching Large Language Models to Self-Debug |
M → U
U → M
S → M
w/ random
pre-trained
95.4 ± 1.8
92.7 ± 1.4
85.2 ± 4.7
Train on full
target dataset
97.3 ± 0.3
98.6 ± 0.5
98.6 ± 0.5
Table 2: Performance of our method and baselines in adapting models among MNIST (M), USPS
(U), and SVHN (S). 100 distilled images are trained for ten GD steps and three epochs. Our me... | DATASET DISTILLATION |
Today, Apple is the world’s most valuable company, but when it | The Casino on Mars |
Table 1: CLIP scores and CLIP R-Precision (Park et al., 2021) values for generated samples and
ground truth videos on prompts from our test set. Cells highlighted in green represent distilled
models. We compare three different combinations: original pipeline, distilled SR models on top
of original base model, and fully... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
G
P
T
A
s
,
i
s
p
o
s
i
t
i
v
e
)
.
W
e
f
u
r
t
h
e
r
s
h
o
w
t
h
a
t
p
r
i
n
c
i
p
a
l
s
p
r
e
f
e
r
V
C
G
c
o
n
t
r
a
c
t
s
o
v
e
r
c
l
a
s
s
i
c
c
o
n
t
r
a
c
t
s
t
h
u
s
e
s
t
a
b
l
i
s
h
i
n
g
r
o
b
u
s
t
n
e
s
s
w
h
e
n
p
r
i
n
c
i
p
a
l
s
c
h
o
o
s
e
b
e
t
w
e
e
n
t
h
e
s
... | Principal-agent VCG contracts - ScienceDirect |
to offset variance when training with the high number of
layers in the full FLPM framework.
3.2. Pretraining Strategy | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
3. Literotica. Literotica is a website where users
can upload short-form erotic fiction. We had
originally planned on including it in the Pile
and even went as far as scraping and process-
ing it. However we decided to not include it
for several reasons. Firstly, once we decided
to exclude fanfiction, Literotica represen... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
A Mathematical Derivations
A.1 Deriving the Optimum of the KL-Constrained Reward Maximization Objective
In this appendix, we will derive Eq. 4. Analogously to Eq. 3, we optimize the following objective:
(cid:2)π(y|x)||πref(y|x)(cid:3)
(11)
under any reward function r(x, y), reference model πref and a general non-p... | Direct Preference Optimization |
Optimization Workstream
Emanuel Taropa, Co-Lead
Rohan Anil, Co-Lead
Vlad Feinberg, Core Contributor
Yujing Zhang, Core Contributor
Zachary Nado, Core Contributor
Aurko Roy, Contributor
James Bradbury, Contributor
Reiner Pope, Contributor
Wei Li, Core Contributor
YaGuang Li, Contributor
Code Pre-training Workstream
Ema... | PaLM 2 Technical Report |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.