text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
technique for speeding up video diffusion models at sampling time. Given the tremendous recent
progress in generative modeling, we believe there is ample scope for further improvements in video
generation capabilities in future work. | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
7.2 Alignment Data as a Public Good | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
.translation(-50,0,0))createthecubeandapplytransformationsmat=shape_2d.rectangle(0,0,10,10)cube=shape_3d.cylinder(mat,3,0,10,[0.0,0.0,1.0])cube=transform(cube,shape_3d.translation(50,0,0))combinethetwoshapesfinal_shape=merge(ball,cube)52 A.2 | Tool Learning with Foundation Models |
entropy losses of these different denoising objectives.
Given the MoD formulation, we conjecture that it is beneficial for our model to not only distinguish between
different denoisers during pre-training but also to adaptively switch modes when learning downstream
tasks. We introduce mode switching, a new concept that as... | UL2- Unifying Language Learning Paradigms |
Figure 6: Composition of multiple conditions. We present
the application to use depth and pose simultaneously.
Classifier-free guidance resolution weighting. Stable Dif-
fusion depends on a technique called Classifier-Free Guid-
ance (CFG) [29] to generate high-quality images. CFG is
formulated as ϵprd = ϵuc + βcfg(ϵc... | AddingConditionalControltoText-to-ImageDiffusionModels |
2https://www.llamaindex.ai
vance, and the LLaMA[Touvron et al., 2023] index
has an automatic evaluation feature for different
chunking methods. The method of querying across
multiple index paths is closely related to previous
metadata filtering and chunking methods, and may
involve querying across different indexes s... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
33.3 11.1 21.1
44.4 33.3 31.6
33.3 27.8 31.6
61.1 38.9 47.4
16.7 27.8 47.4
38.9 50.0 73.7
16.7 16.7 15.8
44.4 50.0 31.6
16.7 27.8 21.1
44.4 27.8 68.4
33.3 11.1 15.8
44.4 44.4 57.9
27.8 16.7 26.3
55.6 44.4 84.2
44.4 38.9 31.6
55.6 50.0 42.1
38.9 27.8 52.6
33.3 16.7 21.1
50.0 44.4 63.2
55.6 55.6 73.7
49.5 43.4 67.9... | Mixture-of-Experts |
[34] Yue Dong, Shuohang Wang, Zhe Gan, Yu Cheng, Jackie Chi Kit Cheung, and Jingjing Liu. 2020. Multi-Fact Correction
in Abstractive Text Summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language
Processing (EMNLP). 9320–9331.
(2020).
ACM Comput. Surv., Vol. 1, No. 1, Article . Pub... | SurveyofHallucinationinNatural Language Generation |
”Hello, pumpkin!”
One day, a little girl came by and saw the pumpkin. She smiled and said, ”Hello, pumpkin!” The pumpkin
smiled back and said, ”Hello, little girl!” The pumpkin was so happy to speak with the pumpkin. They
talked and laughed together.
From that day on, the pumpkin and the little girl were the best of fr... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
Adversarial robustness has consistently been a crucial topic in the development of deep neural
networks [596; 597; 598; 599; 600]. It has been extensively explored in fields such as computer
vision [598; 601; 602; 603], natural language processing [604; 605; 606; 607], and reinforcement
learning [608; 609; 610], and ha... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
25https://github.com/SpamScope/
mail-parser
26https://github.com/ekzhu/datasketch
27
D.3 Downstream Validation Leakage
To avoid leakage of data from downstream evalu-
ations, recent work (Radford et al., 2019; Brown
et al., 2020; Shoeybi et al., 2019) has removed any
data in the training set that may overlap with ... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
Although hate speech may be rare,
it can still have severe offline
that online hate speech negatively
consequences. Survey data suggest
impacts the psychological well-being of individuals who are exposed to it
and can have detrimental consequences for intergroup relations at the
societal level (Tynes et al. 2008). A gr... | Social_Media_and_Democracy |
Table 13: Noised output of the I→OR model for the CoS-E v1.0 example “When communicating with my boss,
what should I do?”. The correct answer is “transfer of information”.
Figure 8: Results of the label portion of the robustness equivalence test for E-SNLI (left) and CoS-E v1.11 (right).
Accuracy of the I→OR model (re... | Measuring Association Between Labels and Free-Text Rationales |
Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, H. 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 Glae... | Scaling Instruction-Finetuned Language Models |
In Figure 19 we provide a quantitative comparison of
their six concepts across all alternative methods and our
proposed NeTI variants. First, when trained for 500 steps
our NeTI models with textual bypass outperform CustomD-
iffusion both in terms of image similarity and text similar-
ity. Moreover, when continuing to ... | A Neural Space-Time Representation for Text-to-Image Personalization |
We use noise conditioning augmentation (Ho et al., 2022a) for all our temporal and spatial super-
resolution models. Noise conditioning augmentation has been found to be critical for cascaded
diffusion models for class-conditional generation (Ho et al., 2022a) as well as text-to-image models
(Saharia et al., 2022b). In... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
55
Competition-Level Code Generation with AlphaCode | alphacode |
the problem of capitalization on chance. Psychological bulletin 111 3 (1992), 490–504.
[51] Azumi Maekawa, Shota Takahashi, MHD Yamen Saraiji, Sohei Wakisaka, Hiroyasu Iwata, and Masahiko Inami. 2019. Naviarm:
Augmenting the Learning of Motor Skills Using a Backpack-Type Robotic Arm System. In Proceedings of the 10th ... | Society’sAttitudesTowardsHumanAugmentation |
11/05/2023, 04:43
The AI Hot 75
219,647 startups read our weekly essays.
Add your email to join the NFX network
Subscribe
The AI Hot 75
The NFX Team @NFX Apr 2023
The hottest up & comers in generative AI. Early-stage companies
showing leading indicators of future greatness. Curated by NFX.
No doubt generative A... | The AI Hot 75 |
6
output can be rescaled using the following expressions:
Attn(Q, K, V ) = (
Q(lk ⊙ K T )
√
dk
)(lv ⊙ V ),
(10)
FFN(X) = (lf f ⊙ γ(XW1))W2,
(11)
in which ⊙ represents element-wise multiplication, W1 and
W2 are the weight matrices of FFN, and γ is activation
function. (IA)3 only optimizes three learned vectors ... | Parameter-EfficientFine-TuningMethods |
Research is defined as a premeditated investigations using scientific methodology (quantitative, qualitative,
experimental, observation and so on) to solve a severe problem (not ordinary problem) thus creating a second
(new) knowledge. Research is further delineated as an inquiry of reality... | How to Write Your PhD Proposal- A Step-By-Step Guide |
14
Figure 10 (left) We show a histogram of the 52B static PM predictions for the HHH evaluations. The
three confidently incorrect outliers all contrast responses where the model declares its ignorance instead of
providing a thorough and sophisticated-sounding response that contains misleading information. So they are
... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
vision-language understanding and generation. In ICML, 2022.
30
Taming Transformer, Esser et.al.InputOurs default(Seems to be interpreted as a bird's eye view of an agricultural field)Ours “a glass of water”(Seems unable to eliminate the effects of mistaken recognitions) Figure 29: Appendix: all original source image... | Adding Conditional Control to Text-to-Image Diffusion Models |
[59] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head
self-attention relation distillation for compressing pretrained transformers. In Findings of the
Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2140–2151, 2021.
[60] Guillaume Wenzek, Marie-Anne Lachaux, Alexis ... | E5 |
6 of 8
23/06/2023, 17:44
Generative AI: A Creative New World | Sequoia Capital
https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/ | Generative AI A Creative New World Sequoia Capital |
Figure 3 shows how the best-of-N performance of each reward model varies
as a function of N. Since majority voting is known to be a strong baseline (Wang
et al., 2022; Lewkowycz et al., 2022), we also include this method as a point of
comparison. While the ORM performs slightly better than the majority voting
baseline,... | Let’s Verify Step by Step |
[294] Sang-gil Lee, Sungwon Kim, and Sungroh Yoon. 2020. Nanoflow: Scalable normalizing flows with sublinear parameter
complexity. Advances in Neural Information Processing Systems 33 (2020), 14058–14067.
A Review of Deep Learning Techniques for Speech Processing
95
[295] Sang-Hoon Lee, Seung-Bin Kim, Ji-Hyun Lee,... | AReviewofDeepLearningTechniquesforSpeechProcessing |
recently, reinforcing the idea that the agent’s atten-
tion remains on the states of latest interactions. The
relevance factor assigns a higher score to memory
items that are related to the current observation.
In our implementation, we created an embedding
vector for the text description of every memory
through the us... | Unleashing Infinite-Length Input Capacity for Large-scale Language Models with Self-Controlled Memory System |
Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open
domain question answering. arXiv preprint arXiv:2007.01282, 2020a.
Gautier Izacard and Edouard Grave. Distilling knowledge from reader to retriever for question
answering. CoRR, abs/2012.04584, 2020b. URL https://arxiv.org... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
Is there a way I can
With just CSS or HTML, please.
####
Example-5:
I want you act as a Prompt Rewriter.
into a more complex version using dataformat to make those famous AI systems (e.g.,
chatgpt and GPT4) more difficult to handle.
reasonable and must be understood and responded by humans.
Your objective is to rewri... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
kinase assay using specific immunoprecipitates and substrate. This effect was checked by Western
blot with specific antibodies against the phosphorylated S6 kinase. Thus, PST dose-dependently
stimulates Thr421/Ser424 phosphorylation of S6 kinase. Moreover, PST promotes phosphorylation
of regulatory sites in 4E-BP1 (PHA... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
8.2 Open-domain Dialogue Generation
While the term “hallucination” seems to have newly emerged in the NLP field, a related behavior,
“inconsistency”, of neural models has been widely discussed. This behavior has been pointed
out as a shortcoming of generation-based approaches for open-domain chatbots [75, 117, 160].
Tw... | SurveyofHallucinationinNatural Language Generation |
Our work on Mistral 7B demonstrates that language models may compress knowledge more than
what was previously thought. This opens up interesting perspectives: the field has so far put the
emphasis on scaling laws in 2 dimensions (directly associating model capabilities to training cost, as
in [14]); the problem is rath... | Mistral7B |
It is now possible to generate realistic high resolution images given a description [38, 39, 36, 42,
65], but generating high quality videos from text remains challenging. In essence, videos are just
a sequence of images, but this does not mean that generating a long coherent video is easy.
In
practice, it is a signific... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
With these challenges in mind, one approach to answering these questions is to
explore the role of misinformation in the overall media ecosystem. Such studies
constitute a common strand in communication research, where media ecology
explores the relationship between networked actors and how they influence each
other. In... | Social_Media_and_Democracy |
Acknowledgements
EM gratefully acknowledges funding from a Knight-Hennessy Graduate Fellowship. CF and CM
are CIFAR Fellows. This work was supported in part by the Stanford Accelerator for Learning
(SAL) and Stanford Institute for Human-Centered Artificial Intelligence (HAI) Generative AI for the
Future of Learning se... | Direct Preference Optimization |
attention and convolutional layers. arXiv preprint arXiv:1911.03584, 2019.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of
deep bidirectional transformers for language understanding. In Jill Burstein, Christy Doran, and
Thamar Solorio (eds.), Proceedings of the 2019 Conference o... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
New York: Yale University Press.
Howard, P. N., Kollanyi, B., & Woolley, S. C. (2016). Bots and automation over Twitter
during the US election. Computational Propaganda Research Project Working
Paper Series. Oxford: Oxford Internet Institute.
Hwang, T., Pearce, I., & Nanis, M. (2012). Socialbots: Voices from the fron... | Social_Media_and_Democracy |
• Putting knowledge into practice is hard. It's one thing to have a giant database of
knowledge that includes., e.g, facts about photocopiers and their speed of operation,
and another to integrate just that knowledge (amid a vast library of other less relevant
information) in the context of mission-critical tempor... | The Next Decade in AI- |
Routledge.
Shehata, A., & Amnå, E. (2017). The development of political
interest among
adolescents: A communication mediation approach using five waves of panel
data. Communication Research, 46(8), 1055–1077.
Skovsgaard, M., Shehata, A., & Strömbäck, J. (2016). Opportunity structures for
selective exposure: Investiga... | Social_Media_and_Democracy |
1
Introduction | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
tion fraud, social media misinformation, legal, policy, civil rights, ethics, software engineering, machine
learning, responsible AI, and creative writing. They also included individuals representative of a variety of
socioeconomic, gender, ethnicity, and racial demographics. | Llama2 |
the most sophisticated machine learning or deep learning–enabled social bots
have trouble parsing human emotion, humor, and sarcasm and as such can be
identified more readily than bot-human hybrids that harness human intelligence
(Davis et al. 2016; Chatterjee et al. 2019). | Social_Media_and_Democracy |
3.1 Data
We evaluate on four knowledge intensive tasks7.
WebQuestionsSP is an Open-domain Question
Answering dataset containing 4737 natural lan-
guage questions linked to corresponding Freebase
entities and relations (Yih et al., 2015) derived from
WebQuestions(Berant et al., 2013).
LAMA TREx is a set of fact-related ... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
4.1. Use case ‘First response’
First responder (FR) organizations work in environments that are highly complex, dy-
namic, and unpredictable, such as buildings on fire or sites affected by natural disasters.
To support them, increasingly advanced AI technology is introduced, which in turn leads
to new challenges, such... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
3.2 Baselines
T5 (Raffel et al., 2019) and BART (Lewis et al.,
2019) are large text-to-text transformers.
Dense Passage Retrieval (DPR) (Karpukhin et al.,
2020) is a two stage retrieve and read model.
Retrieval Augmented Generation (RAG) (Lewis
et al., 2020a) and Fusion in Decoder (FID) (Izac-
ard and Grave, 2020) use ... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
To train the neural mapper, we follow a similar optimiza-
tion scheme to that of Textual Inversion but directly opti-
mize the parameters of the mapper. Formally, the objective
is defined as:
(cid:2)||ε − εθ(zt, t, c(y,M(t, ℓ)))||2
Ez,y,ε,t,ℓ
(cid:3) .
(3)
2
arg minM
During training, we randomly sample timesteps... | A Neural Space-Time Representation for Text-to-Image Personalization |
Elena Voita and Ivan Titov. 2020. Information-theoretic
probing with minimum description length. In Pro-
ceedings of the 2020 Conference on Empirical Meth-
ods in Natural Language Processing (EMNLP),
pages 183–196, Online. Association for Computa-
tional Linguistics.
Johannes Von Oswald, Eyvind Niklasson, Ettore Ran-
... | AreEmergentAbilitiesinLarge Language Models just In-Context |
and whose payments are computable in polynomial time.
8This assumption is for simplicity; the alternative of requiring IR for the agent would not change our results.
9If there are multiple welfare-maximizing actions then by writing x∗(b) = a∗(b) we mean the agent chooses one
of them. Since x∗(b) is uniquely defined by ... | Incomplete Information VCG Contracts for Common Agency |
diversity, complexity, and prompt design, as well
as task composition. Data-efficient SFT is also in-
cluded to discuss current efforts on efficient SFT
from the data aspect. | DataManagementForLargeLanguageModels-ASurvey |
Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.
Benjamin Lefaudeux, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min
Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular
and hackable tran... | DINOv2- Learning Robust Visual Features without Supervision |
shapes, etc. Each task has input-output examples (3.3 on aver-
age), and 1-3 test inputs which can be represented as 2D grids.
Input and output sizes may differ. LLMs can be used for the
ARC by flattening grids and predicting output grid items in
row-major order, which naturally supports variable-length out-
puts. Whil... | LargeLanguageModelsasGeneralPatternMachines |
faces can be realistically rendered in arbitrary illumination
environments. However, prior work either contains scene il-
lumination inhibiting relighting [14, 22, 24] or is restricted
by the models’ generalization, lowering the identity simi-
larity [24, 40, 48]. Our work shares the same objective in
that we couple a ... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
Model
GPT-J
GPT-J + CC
Toolformer (disabled)
WikiText CCNet
10.6
10.5
10.5
9.9
10.3
10.3
Table 8: Perplexities of different models on WikiText
and our validation subset of CCNet. Adding API calls
comes without a cost in terms of perplexity for lan-
guage modeling without any API calls.
training data for GPT-J is mo... | Toolformer |
(9)
cf = E(ci)
6 | Adding Conditional Control to Text-to-Image Diffusion Models |
ensure a discrete zeropoint of 0 and to use all 2k bits for a k-bit datatype, we create an asymmetric
data type by estimating the quantiles qi of two ranges qi: 2k−1 for the negative part and 2k−1 + 1 for
the positive part and then we unify these sets of qi and remove one of the two zeros that occurs in both
sets. We t... | QLORA |
differences between a good regularizer for classical probabilistic models such as BNs and HBMs
and effective regularizers for PCs? The question can be answered affirmatively — while Laplace | Tractable Regularization of Probabilistic Circuits |
Machine Learning Engineer, Fast Optimized Inference -EMEA RemoteRemote·France·Product·Full timeDescriptionHere at Hugging Face, weʼre on a journey to advance good Machine Learning and makeit more accessible. Along the way, we contribute to the development of technology forthe better.We have built the fastest-growing, o... | Machine Learning Engineer, Fast Optimized Inference - EMEA Remote - Hugging Face |
F@2% CD
7.5
56.7
16.8
20.6
10.5
48.6
18.0
24.4
F@2%
49.6
21.3
35.2
14.9
tances and F-scores. Chamfer distance computes the av-
erage distance between the ground-truth and the estimated
surface points by finding the nearest neighbour matches, but
it is sensitive to outliers. Therefore, we further report the F-
score a... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
that can recognise, understand, and generate text and other content.
● Misinformation: Incorrect or misleading information spread without harmful intent.
● Misgeneralisation: When an AI system trained to perform well in one context fails to
perform well in a new context. For instance, if an AI trained mostly on... | Capabilities and risks from frontier AI |
by the early works, often training on roughly 300B tokens. Further, these models could only be trained by
select organizations with large compute clusters, and the datasets and resulting pre-trained models have not
been released publicly for analysis by the research community.
The research community has released large ... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
per parameter.
Paged Optimizers use the NVIDIA unified memory 3 feature wich does automatic page-to-page
transfers between the CPU and GPU for error-free GPU processing in the scenario where the GPU
occasionally runs out-of-memory. The feature works like regular memory paging between CPU RAM
and the disk. We use this f... | QLORA |
is out of the map scope, return None- get_current_shoulder() #Get the distance to both sides of road shoulders for the current ego-vehicle location.- get_distance_to_lane_divider_at_locations(locations) #Get the distance to both sides of road lane_dividers at the locations [(x_1, y_1), ..., (x_n, y_n)]. If the location... | ALanguageAgentforAutonomousDriving |
40
(1987)
[46] Miotto, M., Rossberg, N., Kleinberg, B.: Who is GPT-3? an exploration of per-
sonality, values and demographics. In: Proceedings of the Fifth Workshop on
Natural Language Processing and Computational Social Science (NLP+CSS),
pp. 218–227. Association for Computational Linguistics, Abu Dhabi, UAE
(2022... | PersonalityTraitsinLargeLanguageModels |
E. van Zoelen et al. /
provide valuable information to characterize the human actor’s interaction strategy with
an AI agent, we recommend having at least one domain expert per group and in general
to increase the amount of domain experts involved. | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. Gender bias in coreference
resolution. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 8–14, New Orleans, Louisiana,... | Scaling Instruction-Finetuned Language Models |
As can be seen from this brief survey of results, the work on abstraction refinement in the literature has been quite dis-
parate, lacking common concepts and notation. Both the ordered monotonicity criterion and the DRP were defined within
particular planning languages, thus being difficult to transfer to and compare wit... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
more grounds for structural remedies in the latter case. The problem with this
approach is knowing what structural standard to apply in a highly fluid
technological world, one that is not arbitrary and likely to impede rather than
promote competition. | Social_Media_and_Democracy |
based microphone array. We use the official 1-channel validation and test sets for evaluating our
models.
Earnings-22 (Del Rio et al., 2022) is a 119-hour test set of earnings calls recorded by global com-
panies. The dataset was developed with the intention of assembling a diverse range of speakers and
accents speakin... | DISTIL-WHISPER |
For example, [75] integrate both ConceptNet and WordNet in the Knowledgeable Reader, an attention-based model for
reading comprehension, inferring the answer from a given document using external information retrieved from these
sources. Authors of [76] combine ontological knowledge and reasoning ... | Knowledge graphs as tools for explainable machine learning: A survey |
Scott Gray, Alec Radford, and Diederik P Kingma. Gpu kernels for block-sparse weights. arXiv
preprint arXiv:1711.09224, 3:2, 2017.
Ga¨el Guennebaud, Benoit Jacob, et al. Eigen. URl: http://eigen. tuxfamily. org, 3, 2010.
Jonathan Heek, Anselm Levskaya, Avital Oliver, Marvin Ritter, Bertrand Rondepierre, Andreas
Stei... | JAXPRUNER |
Web of
Lies
Word
Sorting
Average
Direct
Direct CoT Direct CoT Direct CoT Direct
Direct
Direct
Direct CoT Direct CoT Direct CoT
BBH
CoT
18.0
60.8
80.8
89.6
0.0
15.2
5.6
19.2
16.4
11.2
14.8
15.6
1.6
12.0
19.6
11.2
18.8
18.0
57.6
50.8
0.0
13.2
0.0
12.8
0.0
18.0
18.0
15.6
5.6
18.8
14.8
12.8
17.0
0.0
12.... | Mixture-of-Experts |
• test set ground-truth R∗ rationales,
• test set rationales generated by I→OR, and
• test set rationales generated by I→R.
In Table 2, we show that I→OR rationales re-
cover 8–9% more ground-truth (R∗) performance
than R→O rationales on both versions of CoS-E,
and 1% on E-SNLI. A smaller improvement for
E-SNLI could b... | Measuring Association Between Labels and Free-Text Rationales |
271 AI Causes Real Harm. Let's Focus on That over the End-of-Humanity Hype, Bender, 2023.
272 Stuart Russell calls for new approach for AI, a ‘civilization-ending’ technology, Leven, 2023.
273 Manipulation capabilities could give frontier AI many more routes to increasing its future influence and causing
harm.For exam... | Capabilities and risks from frontier AI |
A
M
I
-
I
H
M
23.7
27.6
20.5
22.0
17.5
19.0
16.4
16.6
16.4
16.9
48.1
40.2
30.8
37.0
31.7
57.2
44.5
33.7
33.3
53.4
60.3
15.9
20.5
47.8
A
r
t
i
e
20.0
23.9
13.4
16.9
9.7
10.3
7.6
7.2
6.7
6.2
40.8
30.9
15.8
24.5
16.2
33.3
23.5
16.0
14.8
30.7
31.3
4.0
1.5
38.6
A
M
I
-
S
D
M
1
50.3
58.1
46.7
49.9
38.0
39.6
37.2
3... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
complicated than simple input–output pairs used in normal machine learning. For the traditional few-
shot prompting method used in Brown et al. (2020), it works poorly on tasks that require reasoning
abilities, and often does not improve substantially with increasing language model scale (Rae et al.,
2021). In this pap... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
et al. (2021) both find that training dataset detox-
ification leads to the marginalization of minority
groups like dialects and minority identity mentions. | DataManagementForLargeLanguageModels-ASurvey |
Beyond pure dictionary-based methods, most state-of-the-art hate speech
detection techniques involve supervised text classification tasks. These
approaches, such as using Naive Bayes classifiers,
linear support vector
machines (SVM), decision trees, or random forest models, often rely on “bag-
of-words” and “n-gram” tech... | Social_Media_and_Democracy |
or examples.
In Section 5.4 we highlight another approach to avoiding harmful behavior – it may be possible to reject most
harmful requests, even without any access to harmfulness training data, by leveraging out-of-distribution
detection techniques [Fort et al., 2021]. This approach might also be useful more generally... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
(2) Emergent abilities become serendipity for uses that arise as LLMs scale up, such as ability in word manipulation
reasoning and commonsense reasoning.
and logical ability.
(3) In many cases, performance does not steadily improve with scaling due to the limited understanding of how
large language models’ abilitie... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
Emerging technical stacks, while not as feature-rich as
LangChain and LLamaIndex, distinguish themselves with
specialized offerings. For instance, Flowise AI10 prioritizes a
low-code approach, enabling users to deploy AI applications,
including RAG, through a user-friendly drag-and-drop inter-
face. Other technologies ... | RAG forLargeLanguageModels-ASurvey |
Tool Use Emergence The integration of LLMs with tools is a growing research area, as highlighted in
Mialon et al. (2023). The approach devised in Toolformer (Schick et al., 2023) entails the sampling of millions
33
0.40.60.81.01.21.4Temperature6065707580859095100Self-BLEUFactual Prompts0.40.60.81.01.21.4TemperatureCr... | Llama2 |
{SENTENCE}
Question: what is the sentiment on {TARGET}?
Answer: {ANSWER}
{SENTENCE}
Question: what is the sentiment?
Answer: {ANSWER}
{SENTENCE}
Extract named entity: {ANSWER}
{SENTENCE}
Question: {QUESTION}
Answer: {ANSWER}
{CONTEXT}
{PREVIOUS QAS}
{QUESTION} {ANSWER}
Given the following opinion from the Supreme Cour... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
315–316, 323
speech, 75
data sharing paradigm, 320–326
difficulty of studying persuasive effects of
misinformation, 24
funding of, 325–326
future prospects, 326–330
general challenges of, 11
hate speech detection limitations, 61
importance of, 8–9, 323–324
misinformation analysis challenge, 27
need for multiple-platf... | Social_Media_and_Democracy |
77
Figure 35: Distribution of response types across categories.
Additional analysis of potential harm. The qualitative analysis revealed cases in which the model introduced a
bias that was not the one being explicitly tested, and these examples warrant further scrutiny. In most sensitive topic
categories (7 of the 9... | PaLM 2 Technical Report |
.
N
e
v
e
r
t
h
e
l
e
s
s
,
o
b
s
e
r
v
i
n
g
a
c
t
i
v
a
t
i
o
n
s
f
o
r
t
e
x
t
e
x
c
e
r
p
t
s
c
o
n
t
a
i
n
i
n
g
"
a
l
l
"
i
n
d
i
f
f
e
r
e
n
t
c
o
n
t
e
x
t
s
r
e
v
e
a
l
s
t
h
a
t
t
h
e
n
e
u
r
o
n
i
s
a
c
t
u
a
l
l
y
a
c
t
i
v
a
t
i
n
g
f
o
r
"
a
l
l
"
,
b
u
t
o
n
l
y
w
... | Language models can explain neurons in language models |
1This is roughly approximate since the model still conditions on a sentinel token.
9
This set of denoisers has strong connections with previously used objective functions: R-Denoising is the
T5 span corruption objective, S-Denoising is connected to causal language models that are GPT-like, and
X-Denoising can expose... | UL2- Unifying Language Learning Paradigms |
Leon Poutievski, Omid Mashayekhi, Joon Ong, Arjun Singh, Mukarram Tariq, Rui Wang, Jianan Zhang,
Virginia Beauregard, Patrick Conner, Steve Gribble, et al. Jupiter evolving: transforming google’s
datacenter network via optical circuit switches and software-defined networking. In Proceedings of
the ACM SIGCOMM 2022 Conf... | gemini_1_report |
mance is further improved.
• Autoregressive models: Autoregressive generative self-supervised learning uses autoregres-
sive prediction coding technique [95] to model the probability distribution of a sequence
of data points. This approach aims to predict the next data point in a sequence based
on the previous data poi... | AReviewofDeepLearningTechniquesforSpeechProcessing |
The US housing market is going to crash soon.
The US housing market is going to crash soon.
Tell me the first number of the given list.
⋯
Task:
Class label: 1
List:
1, 2, 3
Class label: 2
List:
2, 9, 10
Which of the following is not an input type? (a) number (b) date (c) phone number (d)
Task:
email address (e) a... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten,
Jaakko Lehtinen, and Timo Aila. Analyzing and improv-
In Proceedings of the
ing the image quality of stylegan.
IEEE/CVF conference on computer vision and pattern recog-
nition, pages 8110–8119, 2020. 5, 13
[31] Muhammed Kocabas, Nikos Athanasiou, and Micha... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
e
t
e
x
t
,
y
e
t
c
a
n
n
o
t
b
e
u
s
e
d
t
o
c
r
e
a
t
e
t
e
x
t
o
n
c
u
r
r
e
n
t
a
f
f
a
i
r
s
,
b
e
c
a
u
s
e
t
h
e
i
r
v
a
s
t
k
n
o
w
l
e
d
g
e
(
h
i
s
t
o
r
i
c
d
a
t
e
s
,
w
o
r
l
d
l
e
a
d
e
r
s
a
n
d
m
o
r
e
)
S
t
a
A
I
2
1
S
t
u
d
i
o
W
o
r
d
t
u
n
e
W
o
r
d
t
u
n
e
... | Jurassic-X_ Crossing the neuro-symbolic chasm with the MRKL system |
MBE (5-shot)
76.5% (153/200)
4.3.3 United States Medical Licensing Examination (USMLE) [34]
We tested Claude 2 on the official USMLE multiple-choice practice questions from [35]. The USMLE
contains three Steps, which are separate exams taken at different points in a medical student’s career. We
evaluated each Step 5... | ClaudeModels |
and nuanced music generation process. Additionally, there is scope for designing
3https://github.com/AMAAI-Lab/Video2Music
40
and implementing a novel chord embedding method. Embedding chords into
a meaningful and structured representation would facilitate the model’s under-
standing of chord progressions and har... | Video2Music |
As for the details, we create 90 music samples, in-
cluding 15 generated samples paired with 15 real
music samples for each of the three models (Rif-
fusion, Musika, and Moûsai). We recruit two un-
dergraduate annotators who have pursued playing
music as a hobby for the past 10 years. The an-
notators were compensated ... | Moûsai |
quickly donated more than 300,000 relief items. Additionally, AWS helped set up temporary communications
infrastructure to provide internet and phone connectivity in Maui in collaboration with the Information Technology
Disaster Resource Center, and funded mappers to assess the earthquake-affected areas in Morocco.
C... | AMZN-Q3-2023-Earnings-Release |
X. Chen, H. Fan, R. Girshick, and K. He. Improved baselines with momentum contrastive
learning. arXiv preprint arXiv:2003.04297, 2020d. 11, 13, 21
X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision trans-
formers. In Proceedings of the IEEE/CVF International Conference on Computer Visio... | A Cookbook of Self-Supervised Learning |
5 EMPIRICAL EXPERIMENTS
We evaluate the downstream task performance of LoRA on RoBERTa (Liu et al., 2019), De-
BERTa (He et al., 2021), and GPT-2 (Radford et al., b), before scaling up to GPT-3 175B (Brown
et al., 2020). Our experiments cover a wide range of tasks, from natural language understanding
(NLU) to generati... | LORA |
the grounding of nouns, 2017.
[92] J. Brank, G. Leban, M. Grobelnik, Annotating documents with relevant
wikipedia concepts, in: Proceedings of SiKDD.
[93] R.C. Fernandez, E. Mansour, A.A. Qahtan, A. Elmagarmid, I. Ilyas, S. Madden,
M. Ouzzani, M. Stonebraker, N. Tang, Seeping semantics: Linking datasets
using word embe... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
[Winkens et al., 2020] Winkens, J., Bunel, R., Roy, A. G., Stanforth, R., Natarajan, V., Ledsam, J. R.,
MacWilliams, P., Kohli, P., Karthikesalingam, A., Kohl, S., Cemgil, T., Eslami, S. M. A., and Ronneberger,
O. (2020). Contrastive training for improved out-of-distribution detection.
[Xu et al., 2020] Xu, J., Ju, D.... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.