text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
Configuration Key
attention-config
attention-dropout
bias-gelu-fusion
checkpoint-activations
checkpoint-num-layers
data-impl
distributed-backend
eval-interval
eval-iters
fp16.enabled
fp16.fp16
fp16.hysteresis
fp16.initial-scale-power
fp16.loss-scale
fp16.loss-scale-window
fp16.min-loss-scale
global-batch-size
gpt-j-resid... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
These various types of methods do not necessarily conflict and can collaborate to solve the
hallucination problem in data-to-text generation.
10.4 Future Directions in Data-to-Text Generation
Given the challenges brought by the discrepancy between structure data and natural text, and the
low fault tolerance in the Da... | SurveyofHallucinationinNatural Language Generation |
Lianwei Wu, Yuan Rao, Yongqiang Zhao,
Hao Liang,
and Ambreen Nazir. 2020.
DTCA: Decision tree-based co-attention net-
works for explainable claim verification. In
Proceedings of the 58th Annual Meeting of
the Association for Computational Linguistics,
pages 1024–1035, Online. Association for
Computational Linguistics.
... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
1. Generate 62,000 interview-style programming questions by prompting (Figure 9) Llama 2 70B.
2. De-duplicate the set of questions by removing exact duplicates, resulting in ∼52,000 questions.
3. For each of these questions:
(a) Generate unit tests by prompting Code Llama 7B (Figure 10)
(b) Generate ten Python solutio... | CodeLlama2 |
o
f
n
e
u
r
a
l
n
e
t
w
o
r
k
s
S
z
e
g
e
d
y
,
C
.
,
Z
a
r
e
m
b
a
,
W
.
,
S
u
t
s
k
e
v
e
r
,
I
.
,
B
r
u
n
a
,
J
.
,
E
r
h
a
n
,
D
.
,
G
o
o
d
f
e
l
l
o
w
,
I
.
a
n
d
F
e
r
g
u
s
,
R
.
,
2
0
1
3
.
a
r
X
i
v
p
r
e
p
r
i
n
t
a
r
X
i
v
:
1
3
1
2
.
6
1
9
9
.
| Language models can explain neurons in language models |
and Acceleration. arXiv preprint arXiv:2306.00978 (2023).
[164] Zhuohan Li, Siyuan Zhuang, Shiyuan Guo, Danyang Zhuo, Hao Zhang, Dawn Song, and Ion Stoica. [n. d.]. TeraPipe: Token-Level Pipeline Parallelism for
[165] Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Xingyu Dang, and Song Han. 2023. AWQ: Activation-awa... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
this figure shows the reverse mapping f , rather than f , which is more illustrative here since it shows how the domain
abstraction partitions S1. | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
In 2015, Deep Speech 2 (Amodei et al., 2015) reported
a speech recognition system matched human-level perfor-
mance when transcribing the LibriSpeech test-clean split.
As part of their analysis they concluded: “Given this result,
we suspect that there is little room for a generic speech sys-
tem to further improve on c... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
14
M2UGen
A PREPRINT
Figure 7: Training Stage 3: The Multi-modal Understanding Adapter and Output Projection Layer are fine-tuned
while the LoRA-enabled LLaMA 2 model is trained in this stage. | M2UGen |
References
[1] Sanjeev Arora, Yingyu Liang, and Tengyu Ma. A simple but tough-to-beat baseline for sentence
embeddings. In 5th International Conference on Learning Representations, ICLR 2017, Toulon,
France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. URL
https://openreview.net/forum?id=SyK00... | E5 |
The research discovered that fact evaluation methods founded on natural language inference and
question generation answering exhibit superior performance and can complement each other.
Pezeshkpour [147] proposed a novel metric, based on information theory, to assess the inclusion of
specific knowledge in LLMs. The metr... | ASurveyonEvaluationofLargeLanguageModels |
holds for some (cid:2). The only possibility is (cid:2) = g(a). We have that h(s1) = f (s1), by definition, and t2 = h(s1) (cid:4) post(g(a)) =
h(s1) (cid:4) h(post(a)) = h(s1 (cid:4) post(a)) = h(t1) = f (t1). It follows that (cid:3) f (s1), f (t1), g(a)(cid:4) ∈ E2 and, thus, that τ is C↑.
Suppose instead that (cid:3)... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
We tested the noise robustness of Whisper models and 14
LibriSpeech-trained models by measuring the WER when
either white noise or pub noise from the Audio Degrada-
tion Toolbox (Mauch & Ewert, 2013) was added to the
audio. The pub noise represents a more natural noisy envi-
ronment with ambient noise and indistinct ch... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
fields. For evaluating language models beyond their existing capacities, BIG-bench [172] introduces
a diverse collection of 204 challenging tasks contributed by 450 authors from 132 institutions.
These tasks cover various domains such as math, childhood development, linguistics, biology,
common-sense reasoning, social ... | ASurveyonEvaluationofLargeLanguageModels |
Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze
Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven
Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin,
James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen,... | StarCoder_paper (1) |
Table 7. Long-form transcription performance improves incremen-
tally as additional decoding heuristics are employed. Details on
each intervention are described in Section 4.5.
to distinguish a segment with no speech, but combining
the no-speech probability threshold of 0.6 and the average
log-probability threshold of... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
12
Published as a conference paper at ICLR 2023
A APPENDIX
A.1 ADDITIONAL COMPARISON WITH OBQ
We now provide an additional comparison between GPTQ and OBQ on BERT-base/SQuAD Ra-
jpurkar et al. (2016) and OPT-125M/WikiText2, which is one of the largest models to which OBQ
can be reasonably applied.
Method
OBQ
GPT... | GPTQ |
Cost. LLMs have grown increasingly larger in recent years, with models such as GPT-1, GPT-2, and GPT-3 featuring
117 million, 1.5 billion, and 175 billion parameters, respectively. The cost of training an LLM is heavily influenced by its
size, with estimates suggesting that training the 11B parameter variant of T5 cost... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
A Two-Sided Discussion of Preregistration of NLP Research
Anders Søgaard Daniel Hershcovich
Miryam de Lhoneux
Department of Computer Science
{soegaard,dh}@di.ku.dk
University of Copenhagen
Department of Computer Science
KU Leuven
miryam.delhoneux@kuleuven.be
3
2
0
2
b
e
F
0
2
]
L
C
.
s
c
[
1
v
... | A Two-Sided Discussion of Preregistration of NLP Research |
[76] Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal,
Nicolas Papernot, and Ross Anderson. Model demen-
tia: Generated data makes models forget. arXiv preprint
arXiv:2305.17493, 2023. 6, 27
[77] Wenliang Dai, Junnan Li, and et al. Instructblip: Towards
general-purpose vision-language models with instruction
tu... | Let’sThinkOutsidetheBox |
provided by van Miltenburg et al. (2021). Another
concern is how NLP contributes to social and cul-
tural inequality (Hershcovich et al., 2022). If NLP
research is likely to help some more than others,
this may be reason to require preregistration. Here,
the questionnaires provided by van Miltenburg et al.
(2021) would... | A Two-Sided Discussion of Preregistration of NLP Research |
[36] Zhong Ji, Kailin Xiong, Yanwei Pang, and Xuelong
Li. Video Summarization with Attention-based Encoder-
IEEE Transactions on Circuits and
Decoder Networks.
Systems for Video Technology, 30(6):1709–1717, 2019. 3
[37] Kevin Kilgour, Mauricio Zuluaga, Dominik Roblek, and
Matthew Sharifi. Fr´echet Audio Distance: A Ref... | M2UGen |
[4] Saleema Amershi, Dan Weld, Mihaela Vorvoreanu, Adam Fourney, Besmira
Nushi, Penny Collisson, Jina Suh, Shamsi Iqbal, Paul N Bennett, Kori Inkpen,
et al. 2019. Guidelines for human-AI interaction. In Proceedings of the 2019 chi
conference on human factors in computing systems. 1–13.
[5] John R. Anderson. 1993. Rule... | Generative Agents- Interactive Simulacra of Human Behavior |
(2) Complex processes. Secondly, a considerable portion of Oogiri game data on the Internet relies on bloggers and
website operators who disseminate the Oogiri games through translation in their respective countries. The creation of IT2T-
type Oogiri game data requires specific tools for Optical Character Recognition (... | Let’sThinkOutsidetheBox |
(FLEEK): (Bayat et al., 2023) introduce FLEEK,
an intelligent and model-agnostic tool aimed at
aiding end users, such as human graders, in fact
verification and correction. FLEEK features a
user-friendly interface capable of autonomously
identifying potentially verifiable facts within the
input text. It formulates ques... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
2https://github.com/openai/CLIP/blob/main/notebooks/
Prompt_Engineering_for_ImageNet.ipynb
use the audio descriptors as the classifier for Detic in place
of CLIP text-based ‘class’ embeddings. We use a score
threshold of 0.9 for the qualitative results in Figure 5.
Text query: ”Cooking a meal”
C. Pretraining detai... | IMAGEBIND- One Embedding Space To Bind Them A |
there has been much talk of
The multitude of definitions of misinformation speaks to the need for clarity
on what exactly we, as a scholarly community, mean when we talk about
misinformation. In an attempt to provide such structure, we compiled a wide
variety of definitions of misinformation and related terms. Looking f... | Social_Media_and_Democracy |
REFERENCES
[1] S. E. Palmer, K. B. Schloss, and J. Sammartino, ‘‘Visual aesthetics and
human preference,’’ Annu. Rev. Psychol., vol. 64, pp. 77–107, Jan. 2013.
[2] G. M. Huebner and K. R. Gegenfurtner, ‘‘Conceptual and visual features
contribute to visual memory for natural images,’’ PLoS One, vol. 7, no. 6,
Jun. 2012,... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
Andromeda is a Cerebras Wafer-Scale Cluster composed of 16 CS-2 systems. Figure 7 shows the architecture
of Andromeda, which aligns well with the large-scale parallel nature of deep learning training. Each CS-2
system contains a Cerebras Wafer-Scale Engine (WSE-2) processor, which has 40 GB of high bandwidth
SRAM and c... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Positioning ICON w.r.t. related work. ICON combines
the statistical body model SMPL with an implicit function,
to reconstruct clothed 3D human shape from a single RGB
image. SMPL not only guides ICON’s estimation, but is
also optimized “in the loop” during inference to enhance its
pose accuracy. Instead of relying on t... | ICON |
Feng, J. and Zhou, Z.-H. (2018). Autoencoder by forest. In
Proceedings of the 32nd AAAI Conference on Artificial
Intelligence.
Fernández-Delgado, M., Cernadas, E., Barro, S., and
Amorim, D. (2014). Do we need hundreds of classi-
fiers to solve real world classification problems? J. Mach.
Learn. Res., 15(90):3133–3181.
F... | Adversarial Random Forests for Density Estimation and Generative Modeling |
[12] Christian Gollier, Pierre-Fran¸cois Koehl, and Jean-Charles Rochet. 1997. Risk-taking behavior
with limited liability and risk aversion. Journal of Risk and Insurance 64, 2 (1997), 347–370.
[13] J. Green and J. J. Laffont. 1977. Characterization of Satisfactory Mechanisms for the Revela-
tion of Preferences for ... | Incomplete Information VCG Contracts for Common Agency |
resort to scraping to get the information under platform control. The decision
echoes arguments that academic researchers have themselves made about the
need, and perhaps even the right, to scrape social media data, when doing so is
in the public interest but against the terms of service of a given platform (see
Freelo... | Social_Media_and_Democracy |
Figure 3: Inception Prompt of Code Role-Playing. This shows the task specifier prompt, assistant
system prompt, and user system prompt which are used for studying the Code scenario.
datasets, and analyzed them. Moreover, we will discuss potential extensions of our framework and
highlight both the risks and opportunitie... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
(1) Guidelines
for code generation,
such as “Your function will be reused
for building more complex functions. Therefore, you should make
it generic and reusable.”;
(2) Control primitive APIs, and relevant skills retrieved from the skill library, which are
crucial for in-context learning [36–38] to work well;
4 | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
Shoveling snow.
Drone flythrough of a tropical jungle covered in snow
A beautiful sunrise on mars, Curiosity rover. High definition, timelapse, dramatic colors
A shark swimming in clear Carribean ocean.
A hand lifts a cup.
5
Figure 5: Videos generated from various text prompts.
temporally-coherent videos that are w... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
Figure 8: Ablation Study of the Deformer. Results with
loose clothing in novel poses, generated by EVA3D and our
method with different choices for the articulation module.
5. Conclusion
details can be found in Sup. Mat.
4.3. Ablation Study
Normal Discriminator: Our normal discriminator serves
an important role in im... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten
Bosma, brian ichter, Fei Xia, Ed H. Chi, Quoc V
Le, and Denny Zhou. 2022b. Chain of thought
prompting elicits reasoning in large language mod-
els. In Advances in Neural Information Processing
Systems.
Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei,
Nathan Scales, Xue... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
r
i
l
y
w
i
t
h
i
n
c
r
e
a
s
i
n
g
m
o
d
e
l
s
i
z
e
.
B
e
c
a
u
s
e
l
a
r
g
e
r
m
o
d
e
l
s
h
a
v
e
m
o
r
e
l
a
y
e
r
s
,
t
h
e
s
e
t
r
e
n
d
s
t
o
g
e
t
h
e
r
m
e
a
n
t
h
a
t
e
x
p
l
a
n
a
t
i
o
n
s
c
o
r
e
s
d
e
c
l
i
n
e
w
i
t
h
i
n
c
r
e
a
s
i
n
g
m
o
d
e
l
s
i
z
e
.
R
a
... | Language models can explain neurons in language models |
model after pseudo-labelling using either greedy or beam-search, and so we opted to pseudo-label
the training data with greedy decoding for its faster inference speed. | DISTIL-WHISPER |
the recognition accuracy decreases. For this test we
used a training data set composed of 2925 and a
testing test of 608 silouhettes. For a single neigh-
bour (N = 1), with std = {0,1,2,3}, the recognition
rate is respectively RR = {98.81,96.43,74.6,44.84}.
But, if we augment the number of N assumption re-
turned by th... | VISAPP_HumanPoseEstimation |
2https://github.com/facebookresearch/luckmatters/tree/main/ssl
26
method, DirectCLR, that does not require a trainable projector. They show regularizing
the representation in DirectCLR by applying the InfoNCE SimCLR objective on sub-vectors
of the representation without a trainable projector is sufficient to outperfor... | A Cookbook of Self-Supervised Learning |
The views expressed here are those of the individual AH Capital Management, L.L.C. (“a16z”) personnel quoted and are not the views of a16z or its affiliates. Certain information contained in
here has been obtained from third-party sources, including from portfolio companies of funds managed by a16z. While taken from sou... | State-of-Crypto2023 |
4
Mehrish et al.
Although deep learning has made remarkable progress in speech processing, it still faces certain
challenges that need to be addressed. These challenges include the requirement for substantial
amounts of labeled data, the interpretability of the models, and their robustness to different
environmental ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
References
[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine
Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch,
Katie Millican, Malcolm Reynolds, et al. Flamingo: a vi-
sual language model for few-shot learning. arXiv preprint
arXiv:2204.14198, 2022. 1, 2
[2] Jean-Baptiste Alayrac, Adria Recasens, R... | IMAGEBIND- One Embedding Space To Bind Them A |
to generate plausible audio. However, the generated sam-
ples might not necessarily adhere to the text description pro-
vided as conditioning.
We report the FAD based on two audio embedding models,
both of which are publicly available: (1) Trill2 (Shor et al.,
2020), which is trained on speech data, and (2) VGGish3,
(H... | MusicLM |
trigrams, or lists of length “n” (Burnap and Williams 2016; Waseem and
Hovy 2016; Badjatiya et al. 2017; Davidson et al. 2017). More recent work | Social_Media_and_Democracy |
3 APPROACH | DATASET DISTILLATION |
a network in that they link to one another and engage with each other’s content.
Finally, mainstream social media platforms (Twitter, Facebook, YouTube) are
used by members of these groups to spread disinformation and conspiracy
theories to larger numbers of people and seed topics for journalists. Producers
in this bro... | Social_Media_and_Democracy |
find that when RLHF is applied to large language models, the answer seems to be an almost-categorical
no. Our RLHF-trained models tend to perform better than their raw, generative counterparts on virtually all
evaluations, as summarized in Figure 3. We also argue that one can mix specialized skills with alignment-
relat... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Facebook or other platforms. Although preventing researchers
from
deliberately uncovering a particular individual’s age without user consent
seems a perfectly reasonable barrier to protect privacy, preventing aggregate
analysis of different age cohorts of users on the same basis would necessarily
prevent us from unders... | Social_Media_and_Democracy |
13 %
N/A
10 %
N/A
343 %
N/A
N/A
N/A
103 %
N/A
N/A
244 %
237 %
77 %
76 %
(1)
(2)
(3)
(4)
(5) | AMZN-Q3-2023-Earnings-Release |
Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black,
and Yulia Tsvetkov. 2019. Quantifying social bi-
ases in contextual word representations. In 1st ACL
Workshop on Gender Bias for Natural Language
Processing.
Tom Kwiatkowski, Jennimaria Palomaki, Olivia Red-
field, Michael Collins, Ankur Parikh, Chris Alberti,
Danie... | LLaMA- Open and Efficient Foundation Language Models |
Notice to Affected Individuals
Platforms also provide potentially useful information to individuals affected by
takedown requests. In particular, they may (1) respond to a person who
requested removal, letting them know if the request was honored; (2) notify
the user whose content was taken down; or (3) “tombstone” mis... | Social_Media_and_Democracy |
Self-supervised learning defines a pretext task based on unlabeled inputs to produce
descriptive and intelligible representations [Hastie et al., 2009, Goodfellow et al., 2016].
In natural language, a common SSL objective is to mask a word in the text and predict
the surrounding words. This objective of predicting the c... | A Cookbook of Self-Supervised Learning |
Table 1: Model performance on general benchmarks.
57.17
67.01
52.15
68.68
70.21
85.36
81.71
95.28
7.94
8.99
90.60
-
-
5 | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
Table 6: Training hyperparameters of Qwen-Audio
Configuration
Audio encoder init.
LLM init.
SpecAugment Policy
Optimizer
Optimizer hyperparameter
Peak learning rate
Minimum learning rate
Audio encoder learning rate decay
Learning rate schedule
Weight decay
Gradient clip
Training steps
Warm-up steps
Global batch size
G... | Qwen-Audio |
7.3 Robustness Evaluation
Beyond general tasks, it is crucial for LLMs to maintain robustness against a wide variety of inputs
in order to perform optimally for end-users, given their extensive integration into daily life. For
instance, the same prompts but with different grammars and expressions could lead ChatGPT
and... | ASurveyonEvaluationofLargeLanguageModels |
Preventing AI systems from pursuing unintended goals is an unsolved research problem,
known as the “specification problem”. It is generally not possible to completely express
complex behaviours, concepts, or goals directly in code, and so teaching AI which behaviours
are desirable or undesirable must be done indirec... | Capabilities and risks from frontier AI |
John: Good morning Eddy. Did you sleep well?
Eddy: Good morning dad. Yeah, I slept great.
John: That’s good. What are you working on today?
Eddy: I’m working on a new music composition for my
class. It’s due this week, so I’m trying to get it finished.
But I’m having so much fun with it!
John: That sounds great!
Soon ... | Generative Agents- Interactive Simulacra of Human Behavior |
general knowledge for instruction following and COCO
[45] for aligning the image encoder. In addition to uti-
lizing existing resources, we also collect our own dataset.
We adopt an automated approach to overcome the labor-
intensive and time-consuming nature of manual data col-
lection. Specifically, inspired by previ... | M2UGen |
that addresses incomplete results, as described below.
Addressing Model Failures and Prompting We inspected StarCoder-generated programs on these
benchmarks and found that there were several cases where the model produces what are effectively
empty solutions, e.g., pass or a comment Insert code here. We also observed ... | StarCoder_paper (1) |
importance to OpenAI is the risk of racing dynamics leading to a decline in safety standards, the
diffusion of bad norms, and accelerated AI timelines, each of which heighten societal risks associated
with AI. We refer to these here as acceleration risk.”24 This was one of the reasons we spent
eight months on safety res... | gpt-4-system-card |
19
A: Loose clothingB: Anthropomorphous inputC: HPS failureInputReconstruction from 4 viewpointsHPS References
[1] 3DPeople. 3dpeople.com, 2018. 9
[2] HumanAlloy. humanalloy.com, 2018. 9
[3] RenderPeople. renderpeople.com, 2018. 2, 5, 9, 12
[4] Thiemo Alldieck, Marcus A. Magnor, Bharat Lal Bhatna-
gar, Christian Theob... | ICON |
exercises) to also provide qualitative assessment of safety capabilities of the model. Some participants who
had expertise in offensive security and malware development questioned the ultimate risk posed by “malicious
code generation” through LLMs with current capabilities. | CodeLlama2 |
sent 2 teams, each with 5 players. This means each school has sent 2 * 5 = 10 players. Each school has also sent 2 coaches. This means each school has sent 10 + 2 = 12 people. There are 4 schools, so in total all of the schools have sent 4 * 12 = 48 people. The answer is 48. (Correct)62B Model Output62B Model Output62B... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Ecker, U. K. H., Lewandowsky, S., Cheung, C. S. C., & Maybery, M. T. (2015). He did
it! She did it! No, she did not! Multiple causal explanations and the continued
influence of misinformation. Journal of Memory and Language, 85, 101–115.
https://doi.org/10.1016/j.jml.2015.09.002
Ecker, U. K. H., Lewandowsky, S., Fenton... | Social_Media_and_Democracy |
end, we begin by delving into crucial background information (§ 2). In particular, we commence by
tracing the origin of AI agents from philosophy to the AI domain, along with a brief overview of the | TheRiseandPotentialofLargeLanguageModel BasedAgents |
PAMI, 2022. doi: 10.1109/TPAMI.2022.3170353. Early ac-
cess.
[16] John P Cunningham and Zoubin Ghahramani. Linear dimen-
sionality reduction: Survey, insights, and generalizations.
JMLR, 16(1):2859–2900, 2015.
[17] Qi Dang, Jianqin Yin, Bin Wang, and Wenqing Zheng. Deep
learning based 2D human pose estimation: A surv... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
now is that UL2 drops the SuperGLUE suite against T5 (1B). However, this is compensated by not only
out-performing on 7 out of 8 tasks but also improving performance by 2-4 times on one-shot evaluation. The
gains on supervised fine-tuning is smaller, but still noticeable across the board on XSUM, SGD and TOT.
Table 7: E... | UL2- Unifying Language Learning Paradigms |
Manuscript submitted to ACM, 2023,
Table 3. Relative frequency and frequency difference of responses from the different usage scenario categories and analysis of
Bayesian logistic regression models, for novice vs. competent, proficient, and expert users. Based on paired differences of 4000 draws
from the posterior dis... | Adoptionand AppropriationofLLMs |
We demonstrate the efficacy of Diffusion-DPO by fine-
tuning state-of-the-art text-to-image diffusion models, such
as Stable Diffusion XL (SDXL)-1.0 [30]. Human eval-
uators prefer DPO-tuned SDXL images over the SDXL-
(base + refinement) model 69% of the time on the Par-
tiPrompts dataset, which represents the state-of... | DiffusionModelAlignmentUsing Direct Preference Optimization |
trying to be everything to everyone. They will likely first integrate deeply into applications for leverage and distribution and later attempt to replace the incumbent applications with AI-native workflows. It will take time to build these applications the right way to accumulate users and data, but we
believe the best o... | Generative AI A Creative New World Sequoia Capital |
process during training, particularly when a limited number of inference steps are available. This
improvement results in faster and higher-quality sampling. SpecGrad [264] introduces adaptations
to the time-varying spectral envelope of diffusion noise based on conditioning log-mel spectrograms,
drawing inspiration fro... | AReviewofDeepLearningTechniquesforSpeechProcessing |
depth values and 2) convert them to disparity for scale nor-
malization. This dataset is only used in training, so we do
not use any metadata or class labels.
SUN Depth-only (SUN-D). We use only the ∼5K depth
maps from the val split of the SUN RGB-D [67] dataset
and denote them as SUN Depth-only. This dataset is only
u... | IMAGEBIND- One Embedding Space To Bind Them A |
Answer:
(B) Sunlight is the source of energy for nearly all ecosystems.
Choice 2
Question: Which statement best explains why photosynthesis is the foundation
of most food webs?
Choices: (A) Most ecosystems are found on land instead of in water.
(B) Sunlight is the source of energy for nearly all ecosystems.
(C) Carbo... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
hensively explaining their origins, hold substantial
implications for the discussion surrounding safety
and security when utilising LLMs. The ability of
LLMs to perform well above the random baseline
on tasks that cannot be solved through memorisa-
tion and are indicative of certain “abilities”, without
explicit traini... | AreEmergentAbilitiesinLarge Language Models just In-Context |
BaselineFLAN-ECBASE
Freeze-GateFLAN-ECBASE
Freeze-ExpertFLAN-ECBASE
Freeze-MoEFLAN-ECBASE
37.8
38.2
37.3
36.9
38.1
36.2
Balance-lossFLAN-ECBASE
Table 2: Ablations on different finetuning strategies of FLAN-ECBASE and FLAN-STBASE.
Freeze-GateFLAN-STBASE
Freeze-ExpertFLAN-STBASE
Freeze-MoEFLAN-STBASE
Balance-lossFLAN... | Mixture-of-Experts |
INDEX TERMS Convolutional neural networks, image aesthetics, image memorability, fine art, visual
sentiment.
I. INTRODUCTION
Deep learning techniques have been successfully employed
for resolving a wide variety of tasks in many different
areas. With the rise of digitized and online available fine art
collections, new pe... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
43
Gan, C., Huang, D., Chen, P., Tenenbaum, J. B., & Torralba, A. (2020). Foley
music: Learning to generate music from videos. In Computer Vision–ECCV
2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Pro-
ceedings, Part XI 16 (pp. 758–775). Springer.
Goel, K., Vohra, R., & Sahoo, J. K. (2014). Pol... | Video2Music |
tasks, utilizing questions sourced from the Chinese Gaokao examination. On the other hand,
SOCKET [21] serves as an NLP benchmark designed to evaluate the performance of LLMs in
learning and recognizing social knowledge concepts. It consists of several tasks and case studies
to assess the limitations of LLMs in social ... | ASurveyonEvaluationofLargeLanguageModels |
Optimists might point, for example, to the development of new technologies
of differential privacy that might help us out of the privacy versus access trade-
off. These new methods, which have met with mixed success as part of the
research effort of Social Science One, usually add statistical noise to datasets in
such ... | Social_Media_and_Democracy |
MultiHashEmbed that parametrizes the number of hash functions. We believe that this parameter
can lead to significant improvements in speed and power usage.
Given these findings, we recommend spaCy users to: | MULTI HASH EMBEDDINGS IN SPACY |
9
As a result, this joint training enables the model to generate interesting video dynamics in different
styles. See Fig. 8 for such examples.
2.6.1 CLASSIFIER FREE GUIDANCE
We found classifier free guidance (Ho & Salimans, 2021) to be critical for generating high fidelity
samples which respect a given text prompt. T... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
have been developed (Higgins et al., 2017; Arjovsky et al.,
2017), including some designed for mixed data in the tabu-
lar setting (Choi et al., 2017; Jordon et al., 2019; Xu et al.,
2019). While the evidence lower bound of a VAE approxi-
mates the data likelihood, there is no straightforward way
to compute this quanti... | Adversarial Random Forests for Density Estimation and Generative Modeling |
Criter.
Valid.
Indep.
Conc.
Shapeable
Shapeable
+
−
+
+
++
N/A
+
+
++
++
N/A
−−
−
++
−
Table 6: Summary of results for psychometric test-based experiments across models. Results
for construct validation experiments are summarized left-to-right in terms of structural, con-
vergent, discriminant, and criterion va... | PersonalityTraitsinLargeLanguageModels |
from information stored in context vs in weights. arXiv:2210.05675, 2022.
[51] R. N. Shepard and J.-J. Chang. Stimulus generalization in the learning of classifications. Journal of Experimental
Psychology, 65(1):94, 1963.
[52] F. G. Ashby and J. T. Townsend. Varieties of perceptual independence. Psychological Review... | LargeLanguageModelsasGeneralPatternMachines |
𝐹𝑥ABC𝑥BDE𝑓Figure3.ThetrainingframeworkofLFDM.Ontheleftisstageonefortraininglatentflowauto-encoderwhileontherightisstagetwofortrainingdiffusionmodel.Instagetwo,theencoderΦistheonealreadytrainedinstageone,andthelatentflowsequencefK1andocclusionmapsequencemK1areestimatedbetweenx0andeachframeingroundtruthvideoxK1usingthe... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
Ethical theories can be broadly classified into three categories:
1. Deontological ethical theories - which hold that certain actions are inherently right or
wrong, regardless of the consequences or intentions.
2. Teleological ethical theories - which hold that the morality of an action depends on the
outcome or result ... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
2.5.3. DECODER
where T denotes the total number of layers in the discrim-
inator and Dl outputs the feature map of the l-th layer of
The decoder is essentially the HiFi-GAN V1 genera-
tor (Kong et al., 2020). It is composed of a stack of trans-
Conditional Variational Autoencoder with Adversarial Learning for End-t... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
[47] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dy-
namicFusion: Reconstruction and tracking of non-rigid scenes
in real-time. In Proc. Computer Vision and Pattern Recogni-
tion (CVPR), 2015.
[48] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and
Andreas Geiger. Differentiable volumetric rendering: Learn... | DynIBaR-NeuralDynamicImage-BasedRendering |
Acknowledgments.
We thank Mathilde Caron for initial discussions that led to this work. We thank Olivia Joulin for the horse
drawing used in Fig. 10. We also thank the rest of FAIR and Meta AI for feedback on this work through
the entire project.
References
Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. S... | DINOv2- Learning Robust Visual Features without Supervision |
in neural information processing systems 27 (2014).
[179] Xiangru Tang, Arjun Nair, Borui Wang, Bingyao Wang, Jai Desai, Aaron Wade, Haoran Li, Asli Celikyilmaz, Yashar
Mehdad, and Dragomir Radev. 2021. CONFIT: Toward Faithful Dialogue Summarization with Linguistically-Informed
Contrastive Fine-tuning. arXiv preprint ... | SurveyofHallucinationinNatural Language Generation |
Arc-c HellaS MMLU
Arc-c HellaS MMLU
Arc-c HellaS MMLU
33B
70B
13B
3.2 Long range performance
To assess the capabilities of Mixtral to tackle long context, we evaluate it on the passkey retrieval
task introduced in [23], a synthetic task designed to measure the ability of the model to retrieve a
passkey inserted ran... | Mixtral of Experts paper |
Fig. 10. Effectiveness validation of support set. Without the support set,
although NeRF achieves good rendered image in the training view due to
overfitting, it cannot produce a clear result in a novel inpainting view. By
contrast, the case with support set enable to obtain images with desired quality
in both training ... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
[31] K. Shu, D. Mahudeswaran, S. Wang, D. Lee, and H. Liu, ‘‘FakeNewsNet:
A data repository with news content, social context, and spatiotemporal
information for studying fake news on social media,’’ Big Data, vol. 8,
no. 3, pp. 171–188, Jun. 2020.
[32] M. Amjad, G. Sidorov, A. Zhila, H. Gómez-Adorno, I. Voronkov, and... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
C Math reasoning results
D Infilling
Degradation in random span infilling in SPM format. As observed in Section 3.2 and Table 14,
random span infilling performance on HumanEval infilling tasks (Bavarian et al., 2022) degrades in our
models in suffix-prefix-middle (SPM) format compared to prefix-suffix-middle (PSM) fo... | CodeLlama2 |
8 DISCUSSION
In this section, we reflect on applications, future work and limita-
tions, and ethical and societal risks of generative agents.
8.1 Applications of Generative Agents
Generative agents have vast potential applications that extend be-
yond the sandbox demonstration presented in this work. For in-
stance, s... | Generative Agents- Interactive Simulacra of Human Behavior |
Since the dataset was noisy and diverse, we generated the data using the same TTS model and vocoder
that was employed for the Conversational dataset. This was done to ensure consistency and eliminate
any potential variations in the dataset that could impact the results of the study. For evaluation, we
used the CVSS Jia... | Translatotron3 |
2 Methods
2.1 Using the APIs
In this section we describe the methods of evaluation, including the datasets and metrics we used.
AI21 Summarize API is designed to address summarization with a simple interface2. The request contains a source
(input) text and the response contains its summary, generated by the underlyi... | AI21 SUMMARIZE API- TECHNICAL EVALUATION |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.