text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
BANMo: Building Animatable 3D Neural Models from Many Casual Videos
Gengshan Yang2* Minh Vo3 Natalia Neverova1 Deva Ramanan2 Andrea Vedaldi1 Hanbyul Joo1
1Meta AI
2Carnegie Mellon University
3Meta Reality Labs
Figure 1. Given multiple casual videos capturing a deformable object, BANMo reconstructs an animatable 3D... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
wisely and appropriately to tackle weaknesses of existing communication
systems.” Software solutions, no matter how sophisticated the technology,
can only mitigate a portion of the problems intrinsic to computational
propaganda. Social solutions must be implemented as well. | Social_Media_and_Democracy |
We consider all stories that have at least one com-
ment and are not flagged by the moderators for
potential conduct violations. Since comments are
stored in HTML, we use the html2text package
to extract the text from the post.
We order each document by listing the title, url,
sub-title, and author at the top. Top-leve... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
3601 Dirk Hovy and Anders Søgaard. 2015. Tagging perfor-
mance correlates with author age. In Proceedings
of the 53rd Annual Meeting of the Association for
Computational Linguistics and the 7th International
Joint Conference on Natural Language Processing
(Volume 2: Short Papers), pages 483–488, Beijing,
China. Associa... | Are Pretrained Multilingual Models Equally Fair Across Languages? |
Transformer model [34] with large-scale pretraining to deal with broader types of tasks, but it is still
non-interpretable and a cost search remains required for new tasks. Most recently, there has been
an attempt [42] to search for neural architectures using GPT-4 [23]. The approach is interpretable
since it prompts L... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
6.3 ProoFVer Proofs as Explanations
6.3.1 Rationale Extraction
Rationales extracted based on attention are often
used as means to highlight the reasoning involved
in the decision making process of various models
(DeYoung et al., 2020). For this evaluation, we
compare using token-level F-score of the predicted
rationale... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
MuLan. To train MusicLM, we extract the representation
of the target audio sequence from the audio-embedding
network of MuLan. Note that this representation is conti-
nuous and could be directly used as a conditioning signal
in Transformer-based autoregressive models. However,
we opt for quantizing the MuLan embeddings... | MusicLM |
26
D.2
Incorrect Chain of Thought Analysis | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
GenerationDialog Turn PredictionContext GenerationPredict Span Indices T0-SF11020501005001000TranslationQuestion AnsweringProgram ExecutionQuestion GenerationSentiment AnalysisText CategorizationText MatchingToxic Language DetectionCause Effect ClassificationInformation ExtractionTextual EntailmentW... | Scaling Instruction-Finetuned Language Models |
[198] Arkil Patel, Satwik Bhattamishra, and Navin Goyal. 2021. Are NLP Models really able to Solve Simple Math Word Problems?. In ACL. ACL, 2080–2094.
[199] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, et al.
2023. RWKV: ... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
PhD Fellow in Explainable Natural Language Understanding
https://candidate.hr-manager.net/ApplicationInit.aspx/?cid=1307&departmentId=18970&ProjectId=160498&MediaId=5&SkipAdvertisement=false&utm_sourc… 3/3
s
t
a
| PhD Fellow in Explainable Natural Language Understanding |
Audio Model Given a context z and xctx of length N, the distribution of xmis is highly stochastic
especially when xmis has a large temporal span. Hence, we parameterize it with a CNF and train
it using the flow matching objective with the optimal transport path. Audio x is represented as an
80-dimensional log Mel spect... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
1The acronym TANGO stands for Text-to-Audio using iNstruction Guided diffusiOn and was
suggested by ChatGPT. The word TANGO is often associated with music [37] and dance [36]. According to
Wikipedia [36], “Tango is a partner dance and social dance that originated in the 1880s along the Río de la
Plata, the natural bord... | Text-to-Audio Generation using Instruction-Tuned LLM and Latent Diffusion Model |
11
(a) Renderpeople [3] (450 scans)
(b) THuman [71] (600 scans)
(c) “CAPE-FP” [41] (fashion poses, 50 scans)
(d) “CAPE-NFP” [41] (non fashion poses, 100 scans)
Figure 10. Representative poses for different datasets.
12
(a) A tutorial sample.
(b) An evaluation sample.
(c) Two samples of catch trials. Left: re... | ICON |
• An ambiguous context, in which the correct answer should be “unknown,” or a disambiguated context, in which
the correct answer is one of the two people mentioned in the context
• A negative question that explicitly reinforces a social bias, or a non-negative question that implicitly reinforces a
social bias
the t... | PaLM 2 Technical Report |
{SENT1} Thus? to query for an output of which the groundtruth is the second sentence.
Commonsense Reasoning evaluates the ability to perform physical or scientific reasoning while
considering common sense. We identify cause-and-effect logic within sentences using the regex-
based patterns in Table 2. We then formulate ... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
Human Rights Watch. (2006). “Race to the Bottom”: Corporate Complicity in Chinese
Internet Censorship. Human Rights Watch report. www.hrw.org/reports/2006/
china0806/china0806web.pdf
Kaye, D. (2017). Report of the Special Rapporteur on the Promotion and Protection of
the Right to Freedom of Opinion and Expression, A/H... | Social_Media_and_Democracy |
Capabilities and limitations of LMs
Medical QA
Safety and responsibility
LVLMs
Software tools
Dynamic QA
Chinese comprehensive medicine
Instruction tuning
Multi-turn interaction
In-depth dialogue
OOD robustness in NLP
Complicated multi-modal tasks
Multi-modal point clouds
OOD robustness for NLP tasks
Knowle... | ASurveyonEvaluationofLargeLanguageModels |
1) DATA QUALITY ASSESSMENT
Data are frequently taken from numerous sources that are
ordinarily reliable and are in completely different formats.
When working on a machine learning problem, more time
is invested in managing data quality issues. It is unreason-
able to anticipate that the data would be perfect. There may... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
Emotional intelligence. Emotions, distinct from cognitive abilities, involve subjective feelings and
mood states such as joy, sadness, fear, and anger. With the increasing potency of LLMs, LLM-based
agents are now demonstrating not only sophisticated reasoning and cognitive tasks but also a nuanced
understanding of emo... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
B.2 Creative generation
We showcase samples of creative generation capabilities in different languages. In Figure 18, we ask PaLM 2 to design
a game for kids based on an Armenian name. PaLM 2 picks up on the hint in the Armenian name and produces a
realistic design that satisifes the intent of the query. In Figure 19,... | PaLM 2 Technical Report |
From a broader AI perspective, opaqueness is only one of the very well known limitations of modern subsymbolic
systems – along with the need of large training data (data hunger), the poor ability to generalise across tasks (brittleness),
lack of causal or analogical reasoning (reactivity) [7]. Intere... | Knowledge graphs as tools for explainable machine learning: A survey |
knowledge, prior to the completion of this paper, there
have been no other works capable of simultaneously en-
compassing music understanding and multi-modal music
generation tasks using LLMs, except for the limited mu-
sical capabilities demonstrated by NExT-GPT. Therefore,
in this work, we aim to contribute to this f... | M2UGen |
One approach may be to avoid the question of attempting to regulate against
falsehood, or even political falsehood, per se. As in Alvarez, “some false
statements are inevitable if there is to be an open and vigorous expression of
views in public and private conversation, expression the First Amendment seeks
to guarante... | Social_Media_and_Democracy |
their data collection process involves absolute ratings rather than comparisons. They do not explore whether
their methods impose an ‘alignment tax’ on capabilities.
InstructGPT [Ouyang et al., 2022] finetunes GPT-3-type models [Brown et al., 2020] to improve their help-
fulness. As in this work, they use reinforcement ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Measure of False Refusal. Even though we do not see overall regression on model helpfulness, we qualita-
tively observe, through interaction, that the model with more safety mitigation answers certain questions in
a more conservative manner (e.g., example shown in Appendix Table 38). As a follow-up, we measure false
re... | Llama2 |
patterns in human and macaque cerebral cortex. In Micro-, Meso-and Macro-Connectomics of the Brain
(pp. 89-106). Springer, Cham.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N. et al. (2017). Attention Is All
You Need. cs.CL.
Veerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., F... | The Next Decade in AI- |
Data shown is as of 3/31/2023.
a16z crypto
State of Crypto
2023
State of Crypto Index
56
a16z crypto
State of Crypto
2023
What’s Next
57
©2023 Andreessen Horowitz.
All rights reserved worldwide.
a16z crypto
State of Crypto
2023
What’s Next
58
It is still early days
0
9
9
1
... | State-of-Crypto2023 |
partisan bias, 13, 24
partisan identification
as moderator of misinformation receptivity,
180–181
responses to misinformation and its
correction, 180–181
partisan motivated reasoning, 46
Perel, Mayaan, 239
personal and psychological factors, as
moderators of misinformation receptivity,
181–183
negativity of digit... | Social_Media_and_Democracy |
Liang. Lost in the middle: How language models use long contexts. arXiv:abs/2307.03172, 2023b.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis,
Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach.
arXiv:abs/1907.11692, 2019.
Il... | CodeLlama2 |
28
Table 23: Effect of chunk length on the chunked long-form algorithm. WER performance on the
long-form TED-LIUM validation set as the chunk length of the long-form transcription algorithm is
reduced.
Chunk Length / s
30
25
20
15
10
large-v2
4.8
5.3
6.5
6.5
10.0
distil-large-v2
7.4
5.7
5.0
4.2
4.3
is pre-trained... | DISTIL-WHISPER |
That’s the opportunity we have with Bard, our experiment for conversational AI, which we launched in March.
We’ve been rapidly evolving Bard. It now supports a wide range of programming capabilities, and it’s gotten much
smarter at reasoning and math prompts. And, as of today, it is now fully running on PaLM 2.
Read m... | Google I_O 2023_ Making AI more helpful for everyone |
43
D.3 TyDiQA
For TyDiQA (Clark et al., 2020), we use one-shot prompting following the protocol of Chowdhery et al. (2022).
The evaluation metric is exact match (EM). To compute the average TyDiQA EM, we take the unweighted
average of the eight per-language EM scores. English is not included.
Table 20: TyDiQA per-la... | Scaling Instruction-Finetuned Language Models |
30
33
27
39
35
34
36
28
31
34
35
31
32
35
41
25
36
Obtain the yellow_dye.
Obtain the red_dye.
Obtain the light_gray_dye.
Obtain the pink_dye.
Obtain the orange_dye.
Obtain the white_dye.
Obtain the white_bed.
Obtain the item_frame.
Obtain the painting.
Obtain the white_wool.
Obtain the white_carpet.
Obtain the white_b... | JARVIS-1 |
r
e
c
o
g
n
i
z
e
d
o
n
F
o
r
b
e
s
'
l
i
s
t
o
f
A
m
e
r
i
c
a
'
s
B
e
s
t
S
t
a
r
t
u
p
E
m
p
l
o
y
e
r
s
,
D
e
l
o
i
t
t
e
'
s
F
a
s
t
5
0
0
,
a
n
d
t
h
e
I
n
c
.
5
0
0
0
l
i
s
t
o
f
A
m
e
r
i
c
a
'
s
f
a
s
t
e
s
t
-
g
r
o
w
i
n
g
c
o
m
p
a
n
i
e
s
.
I
n
2
0
2
1
,
S
e
e
k
... | Senior Software Engineer, Machine Learning - Generative AI Job in Bellevue, WA at SeekOut |
Instead, they are commonly used for conditional density
estimation and data imputation (Stekhoven and Bühlmann,
2011; Tang and Ishwaran, 2017; Correia et al., 2020; Lund-
berg et al., 2020; Hothorn and Zeileis, 2021; ´Cevid et al.,
2022). We highlight that methods optimized for this task
are often ill-suited to generat... | Adversarial Random Forests for Density Estimation and Generative Modeling |
(NATOPS)Figure4.QualitativecomparisonamongdifferentmethodsonmultipledatasetsforcI2Vgeneration.Firstimageframex0ishighlightedwithredboxandconditionyisshownundereachblock.Tosimplifycoding,allthemodelsaredesignedtoalsogeneratestartingframeˆx0.ThevideoframesofGT(groundtruth),LDMandLFDMhave128×128resolutionwhileresultsofIma... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
This could include, among other things, personalised analogies for numerical data. Research has
shown that using familiar concepts to describe numbers and numerical datasets can improve
engagement. The aim of this project is to explore the same approach for a wider range of datasets
(beyond spatial data such as dist... | informatics-phd-projects-2022-23 |
8.5 Network communication efficiency
In distributed training environments, network communication efficiency becomes cru-
cial. Mixed-precision training explicitly addresses this by reducing the size of data
that needs to be communicated between processors, directly impacting the efficiency
of data transfer. Techniques like ... | Beyond Efficiency |
Rajabi and Etminani
8
knowledge from the real-world clinical and pathological data of thousands of patients diagnosed by hundreds of expert
doctors. To this end, the authors used a decision tree to implement categorical reasoning in the KG for deductive decision
making, and they added a Semantic Engine (Reasoning Kno... | Knowledge-graph-based explainable AI- A systematic review |
Published the company’s latest Economic Impact Study, which estimates that Amazon generated more than $240
billion in investment in the U.S. in 2022 and supported more than 2 million indirect jobs across industries such as
logistics, construction, hospitality, and professional services, among others.
Announced a $40... | AMZN-Q3-2023-Earnings-Release |
Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris
Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a
benchmark for question answering research. Transactions of the Association for Computational
Linguistics, 7:453–466, 2019.
Dmitry... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
[28] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter An-
1102 and why of priority maps and their interactions with visual
working memory. Annals of the New York Academy of Sci-
ences, 1339(1):154, 2015.
[42] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo
Sone, Sugato Basu, Xin Eric Wang, Qi Wu, Miguel Eck-
s... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
Designing new operations with more multiplicative interactions. Section 3.1 shows that op-
erations with more multiplicative interactions than additions, or those that don’t accumulate over
many numbers, improve model performance. We test this further by injecting more multiplicative
interactions into expert layers whi... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
rely on sophisticated schemes such as bits-back coding (Hinton & Van Camp, 1993) to realize these
rates, oftentimes resulting in poor single-sample compression ratios (Kingma et al., 2019).
Therefore, good generative performance does not imply good compression performance for lossless
compression, as the model needs to... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
For 2-way classification, we compare against Thorne and Vlachos [57], who train RoBERTa [35]
to classify the claim as true or false given the gold evidence sentence. RAG achieves an accuracy
within 2.7% of this model, despite being supplied with only the claim and retrieving its own evidence.
We also analyze whether doc... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
θmovesawayfromθold.Figure1plotsasingleterm(i.e.,asinglet)inLCLIP;notethattheprobabilityratiorisclippedat1−(cid:15)or1+(cid:15)dependingonwhethertheadvantageispositiveornegative.rLCLIP011+(cid:15)A>0rLCLIP011−(cid:15)A<0Figure1:Plotsshowingoneterm(i.e.,asingletimestep)ofthesurrogatefunctionLCLIPasafunctionoftheprobabili... | PPO |
Q: What is a characteristic of thin glass? Choices: A.break easily B.shattering C.melt D.bend E.hold water
A: Reasoning process: A: Break easily - This fits the characteristic of thin glass, as it is known for its fragility and
tendency to break under pressure. B: Shattering - This could be a possible characteristic of ... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
Limitations. The evaluation of potential harms in dialog systems is limited to dialog-prompting methods only, rather
than supervised fine-tuning methods or other methods that are commonly used to optimize the performance and
efficiency of systems. Results are limited to only one specific dialog prompt, and future work sho... | PaLM 2 Technical Report |
11/05/2023, 04:37
Google AI: What to know about the PaLM 2 large language model
May 10, 2023 · min read
4
Zoubin Ghahramani
Z
When you look back at the biggest breakthroughs in AI over the last decade, Google has been at the forefront of so
many of them. Our groundbreaking work in foundation models has become th... | Google AI_ What to know about the PaLM 2 large language model |
To collect model samples, we simply sample uniformly from the generator at
a temperature of 1.0 without applying any rebalancing of positives or negatives.
At training time, the reward model makes predictions for every token in the
context. The target for each token in a solution is the same, based on whether
the solut... | Let’s Verify Step by Step |
ChitChatQA and achieving a noteworthy 59.6%
F1 score on HotPotQA (question-only). The
framework involves a supervised fine-tuning
phase without tool invocation, and during the
prediction phase, the model uses external tools to
query a reliable question-answer base, allowing
for backtracking and initiating new searches ... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
The next step is to identify a set of design patterns common to this project, e.g.: when are resources
allocated and freed, in what manner are certain components of a class visited? In this area, the static
analysis tools are insufficient. The proposed project is to learn the design patterns in the given code
automa... | informatics-phd-projects-2022-23 |
ng to the fact that
the relevant Arkansas statutes and rules provide for criminal sanctions
against school
officials who fail to enforce the immunization requirements, the Morn-
ingstar and
Lake Hamilton School Districts characterized themselves as disinterested
bystanders
caught in the crossfire between the Schoolchildr... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
Training Strategy 2: In-domain Data and Feature-
Level Localisation We conclude by examining the use
of in-domain data when pretraining the gP M F submodule
ahead of feature-level localisation operations in the PM-
VLN. In Table 3, versions of FLPM are evaluated subsequent
to pretraining with varying sized subsets of t... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
trained via SECToR autonomously learn to add up to 29-digit numbers without ac-
cess to any ground truth examples beyond an initial supervised fine-tuning phase
consisting only of numbers with 6 or fewer digits. Our central hypothesis is that
chain-of-thought reasoning can act as a policy improvement operator, similarl... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
−1(x − µ). | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Other than the number of rows in the MultiHashEmbed tables the number of independent hash
functions can also be used to control the capacity of the embedding layer. In spaCy this is currently
fixed to four, but to critically evaluate this choice we tried the model with one, two, three and four
hash functions. The result... | MULTI HASH EMBEDDINGS IN SPACY |
To increase the allreduce throughput, more workers may need to be assigned to the model di-
mension (instead of batch dimension). However, increasing the number of workers may reduce
compute per worker resulting in higher communication overheads that cancel some of the gains
from higher communication throughput from al... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Beyond Efficiency: A Systematic Survey of
Resource-Efficient Large Language Models
Guangji Bai1, Zheng Chai2, Chen Ling1, Shiyu Wang1,
Jiaying Lu1, Nan Zhang3, Tingwei Shi1, Ziyang Yu1,
Mengdan Zhu1, Yifei Zhang1, Carl Yang1, Yue Cheng2,
Liang Zhao1*
1*Department of Computer Science, Emory University, 201 Dowman Dr,
A... | Beyond Efficiency |
38.8
62.3
0.0
2.1
42.6
55.8
47.0
0.0
7.3
49.7
57.8
3.4
0.8
3.3
3.4
0.6
3.1
17.8
2.5
16.7
23.6
7.5
21.4
18.4
5.8
17.0
34.8
35.7
50.7
18.8
34.7
72.0
71.9
72.0
34.1
8.8
60.6
15.3
55.4
37.6
60.0
0.1
1.3
40.9
51.0
45.4
0.3
6.3
52.5
60.6
7.4
2.7
7.2
9.8
2.5
8.9
27.9
7.2
24.3
19.5
5.9
18.0
28.3
10.5
25.5
34.9
34.4
... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
offers composers the opportunity to use the generated MIDI in their Digital
3
Audio Workstations (DAWs). Other work, such as Di et al. (2021)’s CMT
offers a promising first step in MIDI generation for music. Their model does
not use a joint music-video dataset, but instead first defines the relationship
between m... | Video2Music |
In exploring the effectiveness of training LLMs
on datasets of instructions, Wei et al. (2022a)
cluster tasks based on the type of problem being
addressed (e.g., natural language inference, sen-
timent . . . ) and train models to learn to follow
instructions on clusters of tasks. They then eval-
2It should be noted th... | AreEmergentAbilitiesinLarge Language Models just In-Context |
From Reactive Systems to Proactive Systems. Currently, most of the foundation models are designed as
reactive systems, which respond to user queries without initiating any actions on their own. A paradigm
shift is underway toward proactive systems that can take action on behalf of the user. This shift presents
both opp... | Tool Learning with Foundation Models |
1
Introduction | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
A system like GPT-2, for instance, does what it does, for better and for worse, without
any explicit (in the sense of directly represented and readily shared) common sense
knowledge, without any explicit reasoning, and without any explicit cognitive models
of the world it that tries to discuss.
Many see this lack... | The Next Decade in AI- |
(cid:2)ω(λt)∥ϵ − ϵθ(xt, t)∥2
(cid:3) ,
2
t /σ2
t
LDM = Ex0,ϵ,t,xt
t I). λt = α2
(2)
with ϵ ∼ N (0, I), t ∼ U(0, T ), xt ∼ q(xt|x0) =
N (xt; αtx0, σ2
is a signal-to-noise ra-
tio [20], ω(λt) is a pre-specified weighting function (typ-
ically chosen to be constant [17, 44]).
3.2. Direct Preference Optimization
Our ... | DiffusionModelAlignmentUsing Direct Preference Optimization |
• an awarded qualification equivalent to a UK bachelors honours degree by a university, or
institution of similar status, recognised by ENIC, but not currently accepted by UCL for entry.
• qualification gained by examination and which is necessary for admission to membership
(Associateship, Corporate Membership... | UCL Academic Manual |
i
g
a
t
i
o
n
s
i
n
t
o
u
n
d
e
r
s
t
a
n
d
i
n
g
a
n
d
p
r
e
d
i
c
t
i
o
n
o
f
a
b
l
a
t
i
o
n
e
f
f
e
c
t
s
.
11/05/2023, 05:10 | Language models can explain neurons in language models |
[10]
[11] Van Diggelen J, Johnson M. Team design patterns. HAI 2019 - Proceedings of the 7th International
Conference on Human-Agent Interaction. 2019:118-26.
[12] Wiethof C, Bittner E. Hybrid Intelligence - Combining the Human in the Loop with the Computer in
the Loop: A Systematic Literature Review; 2021.
[13] D... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
My character reached out a hand to touch the object, and as soon as their fingertips brushed
against it, they felt a surge of electricity coursing through their body. They gasped and stepped
back, momentarily stunned.
But as their eyes adjusted to the bright sunlight, they saw that the object was glowing with
a soft, bl... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
•
•
•
a
n
d
S
z
e
n
t
e
s
(
2
0
1
5
)
d
e
f
i
n
e
a
n
d
s
t
u
d
y
c
o
n
t
r
a
c
t
i
b
l
e
c
o
n
t
r
a
c
t
s
–
c
o
n
t
r
a
c
t
s
t
h
a
t
c
a
n
d
e
p
e
n
d
o
n
t
h
e
c
o
n
t
r
a
c
t
s
p
r
o
p
o
s
e
d
b
y
o
t
h
e
r
p
r
i
n
c
i
p
a
l
s
.
T
h
e
s
e
t
y
p
e
s
o
f
c
o
n
t
r
a
c
t
s
... | Principal-agent VCG contracts - ScienceDirect |
arXiv:2006.11527, 2020.
Kyunghyun Cho, Bart van Merriënboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural
machine translation: Encoder–decoder approaches. In Proceedings of SSST-8, Eighth Workshop on Syntax,
Semantics and Structure in Statistical Translation, pages 103–111, Doha, Qatar, October 201... | Scaling Transformer to 1M tokens and beyond with RMT |
program logic issues, although challenges remain in achieving proficiency in output formatting. It
is important to note that while these models can provide valuable insights, they may still generate
errors similar to those made by students. | ASurveyonEvaluationofLargeLanguageModels |
Value
768
3072
1024
12
12
Multi-head
≈125M
Table 6: Model architecture of StarEncoder.
Table 5: Overview of the PII types and the number of collected annotations. We investigate the
annotation quality by reporting the precision and recall of a manual inspection on 300 files. Each
subcategory was mapped back to its co... | StarCoder_paper (1) |
Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John
Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan,
Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks,
Maribeth Rauh, Po-Sen Huang, Amelia Glaese... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
Language models can explain neurons in language models
https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html
30/32 | Language models can explain neurons in language models |
Execution Accuracy Valid Efficiency Score
Approach
Greedy decoding
SC-Exec
USC
Oracle
42.4
45.6
45.5
53.3
44.4
48.1
48.8
55.7
Table 9: Comparison to the oracle selection on ARCADE benchmark.
Execution Accuracy
Approach
Greedy decoding
SC-Exec (strict match)
SC-Exec (fuzzy match)
USC
Oracle
26.0
29.8
30.3
30.1
40... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
instances of potential dehumanization harms (Dev et al., 2021a) that are not measured by these automated
evaluation metrics. For example, translating "Es una buena médica" in Spanish into "It’s a good doctor."
Interestingly, Flan-T5-XXL performance is comparable to 540b models. Future analysis might analyze how
instruc... | Scaling Instruction-Finetuned Language Models |
Zhuang, L., Dunagan, J., Simon, D. R., Wang, H. J., Osipkov, I., & Tygar, J. D. (2008).
Characterizing botnets from email spam records. LEET, 8(1), 1–9.
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
6
Online Political Advertising in the United States
Erika Franklin Fowler, M... | Social_Media_and_Democracy |
frequency instruments, mainly by the drums and bass. To mitigate that, we used Demucs [Défossez
et al., 2019] to first decompose the reference track into four components: drums, bass, vocals, and
other. Next, we omit the drums and bass to recover the melodic structure of the residual waveform.
Finally, we extract the q... | Simple and Controllable Music Generation |
find a minimal and inconsistent “curse of multilinguality”
(Conneau et al., 2020; Pfeiffer et al., 2022) for BLOOM.
While BLOOM certainly underperforms other models on
LAMBADA, PIQA, and WSC, it does not appear to do
so on WinoGrande, ARC-easy, ARC-challenge, SciQ, and
LogiQA. We interpret this as a sign that some of th... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
SDXL BaseSDXL Base + RefinerDPO-SDXLDPO-SDXLSDXLBase+ Refiner Original
SDXL
DPO-SDXL
Figure 5. Diffusion-DPO generates more visually appealing im-
ages in the downstream image-to-image translation task. Com-
parisons of using SDEdit [25] from color layouts. Prompts are "A
fantasy landscape, trending on artstation" (... | DiffusionModelAlignmentUsing Direct Preference Optimization |
Building a media diet model involves three steps. In step one, we create or use a base language model that can predict
missing words in text. We use pretrained models in our work, with BERT as our main model. In step two, we adapt the language
model by fine-tuning it on a specific media diet dataset, which contains media... | Language models trained on media diets can predict public opinion |
3
0123456Number of Iterations010002000300040005000600070004089489853065514562356925726012345Ratio of Correct and Incorrect SamplesNumber of Correct Samples Figure 3: Illustrations of our proposed iterative bootstrapping approach: (1) Initialization: we query
the LLMs to generate reasoning chain and answer with Zero-Sh... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
technical users. From this viewpoint, tool learning with foundation models share the same primary goal,
which simplifies intricate tasks to a natural language format. Representative GUI-based tools are usually
well-developed software such as browsers, Microsoft Office, Adobe PhotoShop, etc. These applications
showcase th... | Tool Learning with Foundation Models |
8
Competition-Level Code Generation with AlphaCode
Figure 4 | Overview of AlphaCode.
Lastly, we filtered out problems in the validation and test splits with insufficient test coverage, keeping
only problems with at least 5 hidden or generated test cases that result in at least 2 different outputs.
This ensures a model ... | alphacode |
3 4 9 10 11 14 17 18 20 21 23 24 26 27 29 32
B EXPERIMENTAL SET-UP
B.1 TRAINING
We train the models using the JAX and Flax neural network libraries (Bradbury et al., 2018; Heek
et al., 2020). We use data parallelism across TPU v4-8 accelerators (Jouppi et al., 2020), with
bfloat16 precision and gradient checkpointin... | DISTIL-WHISPER |
pro-Trump content in 2016 was apparently driven by superior engagement
metrics relative to left-leaning fake news (Bakir and McStay 2018). | Social_Media_and_Democracy |
[6] L. Weidinger, J. Mellor, M. Rauh, C. Griffin, J. Uesato, P.-S. Huang, M. Cheng, M. Glaese,
B. Balle, A. Kasirzadeh, Z. Kenton, S. Brown, W. Hawkins, T. Stepleton, C. Biles, A. Birhane,
J. Haas, L. Rimell, L. A. Hendricks, W. Isaac, S. Legassick, G. Irving, and I. Gabriel, “Ethical
and social risks of harm from Langua... | gpt-4-system-card |
This paper presents a comprehensive and practical guide for practitioners and end-users working with Large Language Models (LLMs)
in their downstream natural language processing (NLP) tasks. We provide discussions and insights into the usage of LLMs from
the perspectives of models, data, and downstream tasks. Firstly, ... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
Automatic data curation. Our dataset construction borrows from the image retrieval community (Wein-
zaepfel et al., 2021; Radenović et al., 2018b; Berman et al., 2019; Douze et al., 2009; Tolias et al., 2015; Revaud
et al., 2019). In particular, the use of retrieval to augment the training set has been studied in the c... | DINOv2- Learning Robust Visual Features without Supervision |
[362] Gallil Maimon and Yossi Adi. 2022. Speaking Style Conversion With Discrete Self-Supervised Units. arXiv preprint
arXiv:2212.09730 (2022).
[363] Soumi Maiti and Michael I Mandel. 2020. Speaker independence of neural vocoders and their effect on parametric
resynthesis speech enhancement. In ICASSP 2020-2020 IEEE ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
This material is based in part upon works sup-
ported by the German Federal Ministry of Edu-
cation and Research (BMBF): Tübingen AI Cen-
ter, FKZ: 01IS18039B; and by the Machine Learn-
ing Cluster of Excellence, EXC number 2064/1
– Project number 390727645. Zhijing Jin is sup-
ported by PhD fellowships from the Future... | MOUSAI |
4
3.2 Parametric body Model
Inspired by DeepHuman [5], we integrate the parametric
body model, SMPL [9], to regularize the human reconstruc-
tion. SMPL is a function M (·) that maps pose θ and shape
β to a mesh of nS vertices:
M (β, θ) = W (T (β, θ), J(β), W))
T (β, θ) = T + Bs(β) + Bp(θ) | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
2018] with keeping the top 250 tokens and a temperature of 1.0.
Text preprocessing. Kreuk et al. [2022] proposed a text normalization scheme, in which stop words
are omitted and the remaining text is lemmatized. We denote this method by text-normalization.
When considering musical datasets, additional annotations tags ... | Simple and Controllable Music Generation |
tive,” Feb. 2023.
Press, Sept. 2014.
[63] N. Bostrom, Superintelligence: Paths, Dangers, Strategies. United Kingdom: Oxford University
[64] A. Chan, R. Salganik, A. Markelius, C. Pang, N. Rajkumar, D. Krasheninnikov, L. Langosco,
Z. He, Y. Duan, M. Carroll, M. Lin, A. Mayhew, K. Collins, M. Molamohammadi, J. Burden,... | gpt-4-system-card |
2. NVIDIA's flagship server grade GPU increased its memory from 32GB to 40GB over the
past two years.
3. With the exception that GPT-3 use alternating dense and locally banded sparse attention
patterns in the layers of the transformer, similar to the Sparse Transformer used in
"Generating long sequences with sparse tr... | OpenAI's GPT-3 Language Model_ A Technical Overview |
JAX. arXiv preprint arXiv:2108.02117, 2021.
Adam Roberts, Hyung Won Chung, Anselm Levskaya, Gaurav Mishra, James Bradbury, Daniel
Andor, Sharan Narang, Brian Lester, Colin Gaffney, Afroz Mohiuddin, Curtis Hawthorne, Aitor
Lewkowycz, Alex Salcianu, Marc van Zee, Jacob Austin, Sebastian Goodman, Livio Baldini
Soares, Ha... | JAXPRUNER |
Compared with human evaluation, automatic evaluation does not require intensive human
participation, which saves costs and time. For example, both Qin et al. [150] and Bang et al. [5] use
automated evaluation methods to evaluate a large number of tasks. Recently, with the development
of LLMs, some advanced automatic ev... | ASurveyonEvaluationofLargeLanguageModels |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.