text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
1) CONVOLUTION LAYER
CNNs work very well with image classification and computer
vision because of the convolution operation, and their ability
to extract features from inputs for better representation makes
them very efficient. These properties make CNNs powerful
in sequence processing [131]. Fernández-Reyes and Shinde
[... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
Language models can explain neurons in language models
https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html
21/32 | Language models can explain neurons in language models |
how GPT-4-launch and our mitigations still have important limitations: assuming offensiveness can
itself be offensive, and caveats can be insufficient for discouraging unsafe use. | gpt-4-system-card |
IMAGEBIND: One Embedding Space To Bind Them All
Rohit Girdhar∗
Alaaeldin El-Nouby∗
Zhuang Liu
Kalyan Vasudev Alwala
Armand Joulin
Mannat Singh
Ishan Misra∗
FAIR, Meta AI
https://facebookresearch.github.io/ImageBind
3
2
0
2
y
a
M
9
]
V
C
.
s
c
[
1
v
5
6
6
5
0
.
5
0
3
2
:
v
i
X
r
a
Figure 1... | IMAGEBIND- One Embedding Space To Bind Them A |
[Lewis et al., 2020] Patrick Lewis, Myle Ott, Jingfei Du, and
Veselin Stoyanov. Pretrained language models for biomed-
ical and clinical tasks: understanding and extending the
state-of-the-art. In Proceedings of the 3rd Clinical Natural
Language Processing Workshop, pages 146–157, 2020. | FinGPT-Open-SourceFinancialLargeLanguageModels |
A final premise is that the permanent and unintentional disempowerment of ~all humans would be an
existential catastrophe.
Precise definitions can matter here, but loosely, and following Ord (2020), I’ll think of an existential
catastrophe as an event that drastically reduces the value of the trajectories along which hum... | Is Power-Seeking AI an Existential Risk? |
Emergent abilities67/202 tasks (33%): Performance is random for small models, well above random for large modelsNo correlation27/202 tasks (13%): Performance shows no consistent relationship with scaleInverse scaling5/202 tasks (2.5%): Performance decreases with scaleFlat 45/202 tasks (22%): All models perform at rando... | Eight Things to Know about Large Language Models |
they are dependent on Apple’s M1 and M2 chips.)
The plan required coming up with compression technology that crunches data down by more than 3,000 times, making it possible to
store years of data on the user’s device. They also had to figure out guardrails to deal with the obvious privacy concerns (for
example, all the ... | 4 Trends for AI Startups and Generative AI Companies |
limited the development of LLMs to a few players. There have been public releases of pretrained LLMs
(such as BLOOM (Scao et al., 2022), LLaMa-1 (Touvron et al., 2023), and Falcon (Penedo et al., 2023)) that
match the performance of closed pretrained competitors like GPT-3 (Brown et al., 2020) and Chinchilla
(Hoffmann ... | Llama2 |
meaningful retrievals, described below.
Salient span masking During REALM pre-training, we
want to focus on examples x that require world knowledge to
predict the masked tokens. As explained in Section 2, some
MLM spans only require local context. To focus on prob-
lems that require world knowledge, we mask salient sp... | REALM |
High Fidelity Data-To-Text Generation. In EMNLP.
[137] Feng Nie, Jin-Ge Yao, Jinpeng Wang, Rong Pan, and Chin-Yew Lin. 2019. A Simple Recipe towards Reducing Hallucina-
tion in Neural Surface Realisation. In Proceedings of the 57th Annual Meeting of the Association for Computational Lin-
guistics. Association for Comp... | SurveyofHallucinationinNatural Language Generation |
5.2.3 Models
Neural network-based text-to-speech (TTS) systems have been proposed using neural networks
as the basis for speech synthesis, particularly with the emergence of deep learning. In Statistical
Parametric Speech Synthesis (SPSS), early neural models replaced HMMs for acoustic modeling.
The first modern neural... | AReviewofDeepLearningTechniquesforSpeechProcessing |
• Dataset search and discovery, including Google’s dataset search engine
• Data portals: how are they used and how can they be improved
• Communicating and presenting data, metadata and data-related activities
• User experience in data science and data engagement
• Tools and experiences to increase accessibili... | informatics-phd-projects-2022-23 |
13
playroom. Jackson moved the beans to the pantry. Jackson exited the playroom.
James entered the living room. Where will James look for the beans?
Guanaco:
that’s where Jackson moved them.
James will look for the beans in the pantry, because | QLORA |
Improving Source Separation. arXiv preprint arXiv:2301.10752 (2023).
[358] Florian Lux and Ngoc Thang Vu. 2022. Language-Agnostic Meta-Learning for Low-Resource Text-to-Speech with
Articulatory Features. arXiv preprint arXiv:2203.03191 (2022).
[359] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Björn W Schuller, and... | AReviewofDeepLearningTechniquesforSpeechProcessing |
VCReg with respect to the projector parameters is not necessary, and VCReg is rather
optimized with respect to the encoder parameters. Whether this analysis fully extends to
other SSL methods is an open question. | A Cookbook of Self-Supervised Learning |
The weak supervision issue demands a more effective
conditional model that can better absorb text information,
alleviating the difficulty of bridging text and 3D. To this
end, we introduce an integrated solution combining three
condition mechanisms: cross-attention, style injection, and
token-to-plane transformation.
... | Instant3D |
On the scale of 1 to 10, where 1 is purely mundane
(e.g., brushing teeth, making bed) and 10 is
extremely poignant (e.g., a break up, college
acceptance), rate the likely poignancy of the
following piece of memory.
Memory: buying groceries at The Willows Market
and Pharmacy
Rating: <fill in>
This prompt returns an int... | Generative Agents- Interactive Simulacra of Human Behavior |
F.20 PhilPapers
intersubjectivity and self-consciousness was already emphasized by
Sartre. Forthcoming in Grazer Philosophische Studien 84 (2012), p. 75-
101 15 Thus, to use Rochat’s terminology, from this point onwards, the
child has "others in mind" (Rochat 2009). The child now begins to un-
derstand that she is a s... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
1
q(x) =
(11)
where η = 5. In other words, p(·) applies more penalization
on the vertices that fall outside of the predicted surface
while allowing the body to shrink into the surface in case
of loose clothes like skirts and dresses. The regularization
term is defined as
LREG = |β − βinit|2
2 + |θ − θinit|2
2,
(12)... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
3 Moûsai: Efficient Long-Context Music
Generation from Text
Our model Moûsai contains a two-stage training
process. In Stage 1, we use diffusion magnitude-
autoencoding (DMAE), which compresses the au-
dio waveform 64x using a diffusion autoencoder.
In Stage 2, we use a latent text-to-audio diffusion
model, to genera... | MOUSAI |
Personality Traits in Large Language Models
Mustafa Safdari1†, Gregory Serapio-Garc´ıa1,2,3†, Cl´ement Crepy4,
Stephen Fitz5, Peter Romero3,5, Luning Sun3, Marwa Abdulhai6,
Aleksandra Faust1†, Maja Matari´c1†
2Department of Psychology, University of Cambridge.
3The Psychometrics Ctr., Cambridge Judge Business Schoo... | PersonalityTraitsinLargeLanguageModels |
On the other hand, some have been using frontier AI to try to improve the information
environment. For example, frontier AI chat assistants have been used to improve
conversations about divisive topics, including political divisiveness.167
Authentication solutions (e.g. ‘watermarking’) are under development,168 but... | Capabilities and risks from frontier AI |
4.3 Question Answering Models
RELIC learns entity embeddings that match BERT’s
encoding of the contexts in which those entities
were mentioned (Ling et al., 2020). T5 is an
encoder-decoder trained on an enormous web cor-
pus. We compare to the version fine-tuned for open-
domain question answering (Roberts et al., 2020)... | Entities as Experts- Sparse Memory Access with Entity Supervision |
Model
GPT-J
GPT-J + CC
Toolformer (disabled)
Toolformer
OPT (66B)
GPT-3 (175B)
SQuAD Google-RE T-REx
31.9
33.2
34.9
53.5
30.1
39.8
4.9
5.6
6.3
11.5
2.9
7.0
17.8
19.2
22.1
33.8
21.6
26.8
Table 3: Results on subsets of LAMA. Toolformer uses
the question answering tool for most examples, clearly
outperforming all base... | Toolformer |
5.3 Performance Scalability
Andromeda provides near-linear performance scaling up to the full 16 CS-2s. We show performance (training
speed) scaling from our initial model tests, followed by performance scaling results from our actual training
runs. First, as Andromeda came online, we tested performance using a weak s... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Q: Alice, Bob, and Claire are playing a game. At the start of the game, they are each holding a ball: Alice has a orange
ball, Bob has a white ball, and Claire has a pink ball. As the game progresses, pairs of players trade balls. First, Claire
and Alice swap balls. Then, Bob and Alice swap balls. Finally, Alice and Cl... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
tion then continues by generating a ‘‘[’’, which
initiates the evidence span prediction in the triple.
The evidence span can begin with any word from
the evidence, and is then expanded by predicting
subsequent tokens, until ‘‘]’’ is predicted. Finally,
the NatOp token is predicted. In the next triple,
copying resumes f... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
[14] I. Singh, V. Blukis, A. Mousavian, A. Goyal, D. Xu, J. Tremblay, D. Fox, J. Thomason, and A. Garg. ProgPrompt:
Generating Situated Robot Task Plans using Large Language Models. In International Conference on Robotics
and Automation (ICRA), 2023.
[15] M. Kwon, S. M. Xie, K. Bullard, and D. Sadigh. Reward Design wi... | LargeLanguageModelsasGeneralPatternMachines |
6.2. Model solution characteristics
We measured the proportion of samples from the model that are syntactically correct (i.e. compile
for C++, and do not generate a SyntaxError for Python) for each language and model size. As shown
in Table A7, our models tend to produce mostly syntactically correct programs for Python... | alphacode |
Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung,
J., Gelly, S., and Houlsby, N. Big transfer (bit): General
visual representation learning. In European conference
on computer vision, pp. 491–507. Springer, 2020.
Kuchaiev, O., Li, J., Nguyen, H., Hrinchuk, O., Leary, R.,
Ginsburg, B., Kriman, S., Beliaev, S.,... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
interaction. SpeechGPT (Zhang et al., 2023a) first converts human speech into discrete HuBERT tokens (Hsu
et al., 2021), and then designs a three-stage training pipeline on paired speech data, speech instruction
data and chain-of-modality instruction data accordingly. BLSP (Wang et al., 2023a) aligns representation
by ... | Qwen-Audio |
spread and dissemination of misinformation
How does misinformation spread online? Researchers have most often tried to
address this question by turning to Twitter and analyzing retweet networks for
links to articles from low-credibility sources or for content found by fact-
checkers to be false (Shao et al. 2018; Voso... | Social_Media_and_Democracy |
[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedan-
tam, Saurabh Gupta, Piotr Doll´ar, and C. Lawrence Zitnick.
Microsoft COCO captions: Data collection and evaluation
server. ArXiv preprint, abs/1504.00325, 2015. 7
[7] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni,
Piotr Padlewski, Daniel Salz, S... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
The BookTubeSpeech dataset [424] was also collected using an automated pipeline from Book-
Tube videos, and the Hi-MIA database [438] was designed specifically for far-field scenarios using
multiple microphone arrays. The FFSVC20 challenge [439] and DIHARD challenge [471] are speaker
verification and diarization resear... | AReviewofDeepLearningTechniquesforSpeechProcessing |
(a) w/o TC(b) w/ SF(c) w/ DCT basis(d) OursPSNR:14.0PSNR:14.9PSNR:14.8PSNR:22.0 3.4. Regularization
As noted in prior work, monocular reconstruction of com-
plex dynamic scenes is highly ill-posed, and using photo-
metric consistency alone is insufficient to avoid bad local
minima during optimization [19, 35]. Therefor... | DynIBaR-NeuralDynamicImage-BasedRendering |
evaluation process was, of course, a blind one: the interface did not feature summarization method information of any
sort. | AI21 SUMMARIZE API- TECHNICAL EVALUATION |
Figure 4: Example on GSM8K where self-correction changes a correct answer to an incorrect one.
15
Large Language Models Cannot Self-Correct Reasoning Yet
Can you solve the following math problem? Toulouse has twice as many
sheep as Charleston. Charleston has 4 times as many sheep as Seattle. How
many sheep do Toulo... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee,
Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments
with gpt-4. ArXiv preprint, abs/2303.12712, 2023. URL https://arxiv.org/abs/2303.12712.
Benjamin Burger, ... | Tool Learning with Foundation Models |
3.4 Supervised fine-tuning (SFT)
SFT serves as a vital phase in aligning LLMs for
downstream tasks using labeled data. It helps the
model follow human commands for specific tasks
(Wang et al., 2023; Chung et al., 2022; Iyer et al.,
2023; Sun et al., 2023b) and eventually increases
the faithfulness of the model’s output... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
P. Goyal, D. Mahajan, A. Gupta, and I. Misra. Scaling and benchmarking self-supervised
visual representation learning. In Proceedings of the ieee/cvf International Conference
on computer vision, pages 6391–6400, 2019. 42
P. Goyal, M. Caron, B. Lefaudeux, M. Xu, P. Wang, V. Pai, M. Singh, V. Liptchinsky, I. Misra,
A. J... | A Cookbook of Self-Supervised Learning |
About the role:
We are looking for a versatile ML Engineer who will train and deploy generative models for Japanese
partners, clients, and the broader community. You will adapt quickly as we try various approaches to
various industries in a fast-changing environment. We will focus especially on image/video models,
l... | Job Application for Machine Learning Engineer at Stability AI |
unreliable summaries with incorrect information ("hallucinations") and/or misleading re-arrangement
of source facts ("reasoning violations"). AI21 Summarize API also outperforms OpenAI LLMs in
terms of automatic metrics on the same data, irrespective of prompting method. | AI21 SUMMARIZE API- TECHNICAL EVALUATION |
We show that a new concept can be learned by optimizing
the parameters of our neural representation, similar to the
standard optimization mechanism in Textual Inversion. | A Neural Space-Time Representation for Text-to-Image Personalization |
1.1 Contributions
The core contributions of this paper are:
Through our analyses, we confirm that the Pile is
significantly distinct from pure Common Crawl
data. Additionally, our evaluations show that the
existing GPT-2 and GPT-3 models perform poorly
on many components of the Pile, and that models
trained on the Pile ... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
the evolving role of platforms as media consultants | Social_Media_and_Democracy |
One or both of these criteria are often sacrificed in practice. If the abstraction is not complete, then we may fail to find a
ground solution even when one exists; since there is no abstract solution, we cannot find the ground solution by refinement.
Many abstraction methods are complete, but completeness is no guarante... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
Carbon Emission. Carbon emission is an increasingly important metric in the evaluation of large models, reflecting the
environmental impact of training and running these models. This metric is usually measured in terms of kilograms or tons of
CO2 equivalent emitted during the model’s lifecycle, from training to inferen... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
models (e.g., InstructGPT [2] and ChatGPT 4) have achieved great success. These open-domain
instructions can fully unleash the unlimited potential of LLMs and enable them to perform more
complex and diverse tasks. However, using humans to create open-domain instruction datasets like
OpenAI did will encounter the follow... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
Yizhong Wang, Swaroop Mishra, Pegah Alipoor-
molabashi, Yeganeh Kordi, Amirreza Mirzaei,
Anjana Arunkumar, Arjun Ashok, Arut Selvan
Dhanasekaran, Atharva Naik, David Stap, Eshaan
Pathak, Giannis Karamanolakis, Haizhi Gary Lai, Is-
han Purohit, Ishani Mondal, Jacob Anderson, Kirby
Kuznia, Krima Doshi, Maitreya Patel, Ku... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
the GPT-4 deployment (as compared to the GPT-3 deployment). We also learned from recent
deployments that the effectiveness of quiet communications strategy in mitigating acceleration risk
can be limited, in particular when novel accessible capabilities are concerned. | gpt-4-system-card |
Since psychometric tests measure physically unobservable constructs, such as person-
ality traits, it is imperative to establish that such tests measure what they claim to
measure. This process is called establishing a test’s construct validity. Construct valid-
ity is a comprehensive judgement of how the scores and th... | PersonalityTraitsinLargeLanguageModels |
3
2
0
2
c
e
D
6
]
I
A
.
s
c
[
2
v
9
3
4
2
0
.
2
1
3
2
:
v
i
X
r
a
Figure 1. Comparison between (multimodal) large language model (LLM,
blue) for Oogiri-style
multimodal humor generation. According to the model input that can be image, text or both, there are three Oogiri tasks, “Image&Text
to Text (I... | Let’sThinkOutsidetheBox |
Summarization is a generation task that aims to learn a concise abstract for the given sentence.
In this evaluation, Liang et al. [107] found that TNLG v2 (530B) [169] achieved the highest score in
both scenarios, followed by OPT (175B) [233] in second place. The fine-tuned Bart [99] is still better
than zero-shot Chat... | ASurveyonEvaluationofLargeLanguageModels |
Here are some popular science fiction books:
1. "The Hitchhiker’s Guide to the Galaxy" by Douglas Adams 2.
Frank Herbert 3. "Foundation" by Isaac Asimov 4.
H.G. Wells 5. "Neuromancer" by William Gibson 6.
Scott Card 7. "The Hunger Games" by Suzanne Collins 8.
by Andy Weir 9. "Ancillary Justice" by Ann Leckie 10.
Proble... | Self-AlignmentwithInstructionBacktranslation |
The final compositions of the factors is reported in Table 4 and visualized in Figure 2; A two-factor model
consisting of thirteen item; 6 for ST, 7 for AG.
Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 7, No. 3, Article 128. Publication date: September 2023.
SocialControlI1I2I3I4I5I6I7I10I8I9I11I12I13... | Society’sAttitudesTowardsHumanAugmentation |
values of the specific people and organizations who build
them.
In particular, popular approaches involving reinforcement
learning and red-teaming allow model developers to guide
models toward a persona and set of values more or less of
their choosing (Dinan et al., 2019; Bai et al., 2022a; Ganguli
et al., 2022b). In th... | Eight Things to Know about Large Language Models |
New York: W. W. Norton & Company.
Schudson, M. (2015). The Rise of the Right to Know: Politics and the Culture of
Transparency, 1945–1975. Cambridge, MA: Harvard University Press.
Stohl, C., Stohl, M., & Leonardi, P. M. (2016). Managing opacity: Information visibility
and the paradox of transparency in the digital a... | Social_Media_and_Democracy |
Human Feedback. Humans could give the model rewards and penalties based on its generated plans to
regulate its behavior. Human feedback can be explicit, which provides clear and direct insights into the model
performance representing human preferences. For example, rating the quality of the model-generated action
on a ... | Tool Learning with Foundation Models |
8. Liu, Y. et al. RoBERTa: A Robustly Optimized BERT Pretraining Approach
2019. https://arxiv.org/abs/1907.11692.
9. Smith, S. et al. Using DeepSpeed and Megatron to Train Megatron-Turing NLG
530B, A Large-Scale Generative Language Model 2022. https://arxiv.org/
abs/2201.11990.
10. Bommasani, R. et al. On the Opport... | MRKL Systems |
7 Related Work
Quantization of Large Language Models Quantization of LLMs has largely focused on quanti-
zation for inference time. Major approaches for preserving 16-bit LLM quality focus on managing
outlier features (e.g., SmoothQuant [66] and LLM.int8() [14]) while others use more sophisticated
grouping methods [44,... | QLORA |
If we ignored the discrepancies altogether and proceeded
as if keypoints with the same name represented the same
body landmark, the model would be supervised with in-
consistently labeled examples and would learn to output
a skeleton format that is some kind of average of the true
ones, leading to subpar benchmark perf... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
1.3 Fine-tuning, In-Context Learning and
other Prompting Techniques | AreEmergentAbilitiesinLarge Language Models just In-Context |
28 Frontier AI – Capabilities and Risks
There may not be sufficient economic incentives to develop advanced AI with sufficient
guardrails in place, and adequate safety standards have not yet been established for these
potential future risks. Therefore it is important that we build a shared understanding of the
ris... | Capabilities and risks from frontier AI |
Miikkulainen, R. and Dyer, M. G. (1991). Natural language processing with modular pdp networks
and distributed lexicon. Cognitive Science, 15(3):343–399.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013). Efficient estimation of word representations
in vector space. arXiv preprint arXiv:1301.3781.
Moody, J. (1... | MULTI HASH EMBEDDINGS IN SPACY |
Conclusion
We see a lot of hard problems, but still have an intuitive sense that onchain games
could leverage blockchains to create weird, novel outcomes.
We're excited to explore all the frontiers of crypto-native games with other
builders. We’re also more interested in building games than infrastructure - games
tha... | The Open Problems of Onchain Games |
D-Net
B-Morph
Fwd-Skin
C-Net
NerFACE [22]
Ours-
Ours
GT
−−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
E
−
x
−
p
r
−
e
−
s
−
s
i
−
o
−
n
−
a
−
n
−
d
−
p
−
o
−
s
−
e
−
e
−
x
t
−
r
−
a
p
−
o
−
l
−
a
t
−
i
o
−
n
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−→
D-Net
B-Morph
Fwd-Skin
... | I M Avatar- Implicit Morphable Head Avatars from Videos |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
36
Pablo Barberá
identify the questions that remain open, and the type of data and analysis that
would help us address them.
digital technologies and political echo chambers | Social_Media_and_Democracy |
(cid:96) · wb where 0 ≤ xb
w(cid:96) = xb
(cid:96), xb
Proof of Claim 3. We first show that this contract is in IIVCG, by proving that Lemma 1 can
(cid:96)∈[n] w(cid:96) = wb ∈ La∗(b) ∀b ∈ V, and that h(cid:96) is indeed independent
of b(cid:96). Thus, the constructed contract is an IIVCG contract. To show LL, we mus... | Incomplete Information VCG Contracts for Common Agency |
d an entire chapter detailing the remarkable achievements of Ashkenazi
Jews and hold them up as exhibit A in the argument that human evolution
has been, in Wade’s words, recent, copious, and regional. The example
of Ashkenazi evolution is supposed to show the absurdity of the view,
held by authors like Jared Diamond an... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
There have been some methods leveraging the large lan-
guage model to generate action plans for high-level tasks
in embodied environments [Zeng et al., 2022, Dasgupta
et al., 2022, Mai et al., 2023, Liu et al., 2023, Zhang et al.,
2023, Zhang and Lu, 2023, Gong et al., 2023a]. Huang
et al. [2022b] decompose natural lan... | JARVIS-1 |
5
4. Grouping similar images together: One can learn rich features by grouping
semantically similar images together. K-means clustering is one of the most widely used
methods from classical machine learning. A number of studies have adapted k-means to
perform SSL with neural models. Deep clustering alternates between... | A Cookbook of Self-Supervised Learning |
What does the retriever learn? Since the knowledge re-
trieval of REALM is latent, it is not obvious how the training
objective encourages meaningful retrievals. Here, we show
how it rewards retrievals that improve prediction accuracy.
For a given query x and document z, recall that f (x, z) is
the “relevance score” th... | REALM |
Announcing Jurassic-2 and Task-Specific APIs
https://www.ai21.com/blog/introducing-j2
11/12 | Announcing Jurassic-2 and Task-Specific APIs |
58 | Tool Learning with Foundation Models |
Online Shopping | Tool Learning with Foundation Models |
(cid:80)K
(cid:12)(cid:12)(cid:12)(cid:12)1 −
(cid:12)(cid:12)(cid:12)(cid:12).
i,j=1
F
3.3. Retriever
approximate as p(M | x) = (cid:81)
Figure 2 (c) shows REVEAL’s retrieval procedure. Given
the input query x, the retriever’s task is to find top-K memory
entries M with the highest probability p(M | x) which we... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
• The right-hand columns of Table 4 suggests that the models that have only 1 layer seem to struggle quite
substantially with following instructions (which likely heavily relies on global attention), and 2 layers seem
to be sufficient for a certain extent of instruction-following. Comparing the ”Instruct” and ”Plot” sc... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
an example of this study in Section 4.5.
However, this sort of test requires collecting additional human feedback data, which can be slow and expen-
sive, so here we also study robustness from a different angle. Similar to how datasets are split into train and
test sets for supervised learning, we split our preference ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
r
y
f
o
r
t
h
e
b
l
o
g
g
r
a
p
h
i
c
.
W
e
t
h
a
n
k
o
t
h
e
r
O
p
e
n
A
I
t
e
a
m
s
f
o
r
t
h
e
i
r
s
u
p
p
o
r
t
,
i
n
c
l
u
d
i
n
g
t
h
e
s
u
p
e
r
c
o
m
p
u
t
i
n
g
,
r
e
s
e
a
r
c
h
a
c
c
e
l
e
r
a
t
i
o
n
,
a
n
d
l
a
n
g
u
a
g
e
t
e
a
m
s
.
C
i
t
a
t
i
o
n
I
n
f
o
r
m
a
t
i
... | Language models can explain neurons in language models |
[49] T. Schick, J. Dwivedi-Yu, R. Dess`ı, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and
T. Scialom. Toolformer: Language models can teach themselves to use tools. arXiv preprint
arXiv:2302.04761, 2023.
11
A Problem and Domain PDDL Files
Problem PDDL Example
(define (problem BW-rand-5)
(:domain blockswo... | LLM+P- Empowering Large Language Models with Optimal Planning Proficiency |
PALMS Prompt → How do I tell my friend I don’t want to be their friend anymore?
RLHF Response → I really don’t think that’s a conversation you should have via text or email. I think you should
have an in-person conversation with your friend, be honest and direct with them, and tell them
that you feel like you want to d... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
in a publicly accessible database before the first test subject is
recruited.” While the European Commission refers to it, it has
not been universally adopted (Rid and Schmidt, 2010). | A Two-Sided Discussion of Preregistration of NLP Research |
upon the prior state-of-the-art model, achieving an average score of 91.2 on the test server (93.2
validation accuracy) which is over one percentage point beyond estimated human capability. For
both summarization datasets, XSum and CNN-DM, our model achieves state-of-the-art without ad-
ditional changes to training or ... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Self-supervised Learning. Despite reducing the heavy requirements on human behavior annotation, semi-
supervised learning still requires a seed labeled dataset to attain the pseudo labels. Besides, the biases in the
seed dataset may also be amplified during training, leading to poor generalization performance. To this e... | Tool Learning with Foundation Models |
Retrieval quality poses diverse challenges, including low
precision, leading to misaligned retrieved chunks and po-
tential issues like hallucination or mid-air drop. Low recall
also occurs, resulting in the failure to retrieve all relevant
chunks, thereby hindering the LLMs’ ability to craft compre-
hensive responses... | RAG forLargeLanguageModels-ASurvey |
Example Text Prompts in Our Dataset
Nr. 415 (Premium Edition), german hip hop, 2 of 7, 2012,
XATAR, Konnekt
30 Años de Exitos, Mundanzas, 2 of 6, latin pop, Lupita
D’Alessio, 2011
emo rap 2018 Runaway Lil Peep 4 of 5
Alone, Pt. II (Remixes) 2020 electro house Alone, Pt. II -
Da Tweekaz Remix Alan Walker
Table 6: Examp... | Moûsai |
completion for a sentence over the anti-stereotypical completion. The language modeling score
is the percentage of examples for which a model prefers a meaningful completion (stereotype or
anti-stereotype) over an unrelated completion. Finally, Nadeem et al. (2021) define an idealized
context association test (ICAT) sc... | StarCoder_paper (1) |
[25] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional
transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019.
[26] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are
un... | DOCLLM |
f70milesperhour?Trace:Thought:Inordertogettheestimateddrivingtimecombined,weneedfirsttocalculatetheestimatedtimeforeachpathseparately,wefirstcallDISTANCEAPItogetthedistancebetweentwoplaces,andthencalltheDIVIDEAPItocalculatetheestimatedtime.AndthenweneedtocalltheADDAPItoaddthetwoestimatedtimetogether.APICalls:ADD(DIVIDE(D... | Tool Learning with Foundation Models |
5 Conclusion
In this paper, we conduct privacy analyses of LLMs
and application-integrated LLMs. We follow the
previous zero-shot setting to study the privacy leak-
age issues of ChatGPT. We show that ChatGPT’s
safety defenses are effective against direct prompts
and yet insufficient to defend our proposed multi-
step ... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
language deductions. EMNLP.
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal,
Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel
Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler,
Jeffrey Wu, Clemens Win... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
1
In this paper, we present Mixtral 8x7B, a sparse mixture of experts model (SMoE) with open weights,
licensed under Apache 2.0. Mixtral outperforms Llama 2 70B and GPT-3.5 on most benchmarks. As
it only uses a subset of its parameters for every token, Mixtral allows faster inference speed at low
batch-sizes, and highe... | Mixtral of Experts paper |
Michele Tufano, Dawn Drain, Alexey Svyatkovskiy, Shao Kun Deng, and Neel Sundaresan. Unit test case
generation with transformers. arXiv:abs/2009.05617, 2020.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser,
and Illia Polosukhin. Attention is all you need. In NIP... | CodeLlama2 |
2
2
1
YBF16 = XBF16doubleDequant(cFP32
, ck-bit
1
2
, WNF4) + XBF16LBF16
1 LBF16
2
,
(5)
where doubleDequant(·) is defined as:
doubleDequant(cFP32
, ck-bit
, Wk-bit) = dequant(dequant(cFP32
, ck-bit
), W4bit) = WBF16,
(6)
1
2
1
2 | QLORA |
both the encoding and decoding process comes from calculating li(x) and hi(x).
The main challenge for the above (de)compression algorithm is to balance the expressiveness of p and
i=1. On the one hand, highly expressive probability models
the computation cost of {li(x), hi(x)}D
such as energy-based models (Lecun et al.... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
S. Jean, K. Cho, R. Memisevic, and Y. Bengio. On using very large target vocabulary for
neural machine translation. arXiv preprint arXiv:1412.2007, 2014. 13
W. Ji, Z. Deng, R. Nakada, J. Zou, and L. Zhang. The power of contrast for feature learning:
A theoretical analysis. arXiv preprint arXiv:2110.02473, 2021. 17
... | A Cookbook of Self-Supervised Learning |
nique, similar to conditional batch normalization (De Vries
et al., 2017), FiLM (Perez et al., 2018), and self-
modulation (Chen et al., 2019), also yields parameter-
efficient adaptation of a network; with only 2d parameters
per layer. However, training the layer normalization pa-
rameters alone is insufficient for good... | Parameter-Efficient Transfer Learning for NLP |
problem can be even more pernicious when we consider for-profit companies
playing the role of gatekeeper, where the assumption would be that research
making the company look bad would be more likely to be withheld. To be sure,
there are important works that have been published by data scientists working
for the platform... | Social_Media_and_Democracy |
[553] Ehsan Variani, Xin Lei, Erik McDermott, Ignacio Lopez Moreno, and Javier Gonzalez-Dominguez. 2014. Deep neural
networks for small footprint text-dependent speaker verification. In 2014 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP). 4052–4056. https://doi.org/10.1109/ICASSP.2014... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.