text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo
Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and
Michael J. Black. Expressive body capture: 3D hands, face,
and body from a single image. In Proceedings IEEE Conf.
on Computer Vision and Pattern Recognition, pages 10975–
10985, 2019. 2, 3, 5, 8, 12
[42] ... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
fashion on NQ.
• Spider-NQ + BM25 (Ram et al., 2022; Robertson & Zaragoza, 2009): A self-supervised
dense retriever trained on the recurring span retrieval task. Here we use the hybrid model
described in Ram et al. (2022), where the dense retriever is Spider, fine-tuned on NQ (similar
to DPR) and the sparse model is BM... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
4.4.3 No-Use Cases and Understanding. Although in most cases, as discussed above, larger models bring better perfor-
mance, there are still many exceptions that should be considered when choosing the appropriate model. | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
Plains/Forest
0.9
0.9667
0.8667
0.6667
0.8667
0.5667
0.27
0.9667
0.9667
0.9667
0.8718
0.5
0.9333
0.9
0.9333
30
30
30
30
30... | JARVIS-1 |
As a result, the neural network trained with the ground-
truth image-SMPL pairs cannot generalize well to the testing
images which have no ground-truth SMPL annotations. A
simple solution is to replace the ground-truth SMPL models
with the predicted ones while still using the ground-truth
surface scans for training sup... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
or based on the relative number of samples that pass example tests available compared to the default
number. When computing these points, all contests were assumed to be two hours for simplicity
even though some were slightly longer. Clustering and a specified number of submissions were done
when either of these conditi... | alphacode |
R. Cosentino, A. Sengupta, S. Avestimehr, M. Soltanolkotabi, A. Ortega, T. Willke, and
M. Tepper. Toward a geometrical understanding of self-supervised contrastive learning.
arXiv preprint arXiv:2205.06926, 2022. 18
Z. Dai, Z. Yang, F. Yang, W. W. Cohen, and R. R. Salakhutdinov. Good semi-supervised
learning that requ... | A Cookbook of Self-Supervised Learning |
While diffusion models might resemble flows [9, 46, 10, 32, 5, 16, 23] and VAEs [33, 47, 37],
diffusion models are designed so that q has no parameters and the top-level latent xT has nearly zero
mutual information with the data x0. Our (cid:15)-prediction reverse process parameterization establishes a
connection betwee... | Denoising Diffusion Probabilistic Models |
We compute the test perplexity of the constituent
datasets of the Pile using GPT-2 (Radford et al.,
2019) and GPT-3 (Brown et al., 2020), shown in
Figure 2. We use all available versions of GPT-2,
and all four versions of GPT-3 available via the
OpenAI API. Because of the cost associated with
using the OpenAI API, we ... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
Ippolito, D., Tram`er, F., Nasr, M., Zhang, C., Jagielski, M.,
Lee, K., Choquette-Choo, C. A., and Carlini, N. Prevent-
ing verbatim memorization in language models gives a
false sense of privacy. arXiv preprint arXiv:2210.17546,
2022.
Jagielski, M., Thakkar, O., Tramer, F., Ippolito, D., Lee, K.,
Carlini, N., Wallace... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
A.4
We use Huggingface Datasets11 to access all
datasets, and Huggingface Transformers (Wolf
et al., 2020) to access pretrained T5 weights and to-
kenizer. To optimize, we use Adam with ϵ = 1e-8,
β1 = 0.9, and β2 = 0.99. We use gradient clipping
to a maximum norm of 1.0 and a dropout rate of
0.1. We train each model on... | Measuring Association Between Labels and Free-Text Rationales |
5.3.1 Few-shot MMLU and Big-Bench Results after Flan training of UL2
OPT 30B
OPT 175B
T5 11B
OpenAI davinci
OPT IML-Max 30B
OPT IML-Max 175B
T0pp 11B
FLAN T5 XXL
FLAN-PaLM 62B
FLAN-PaLM 540B
FLAN-UL2 20B (Best ckpt for both tasks†)
FLAN-UL2 20B (Individual task best)
BBH MMLU
23.5/25.9
28.0
27.3/34.2
30.2
29.5
-/25.9... | UL2- Unifying Language Learning Paradigms |
convinced that our models were HHH in expectation, a clear next step would be to attempt to study and
eliminate bad behaviors (especially harmfulness) even in the worst case. We have not addressed this question
of robustness here, but hope to study it in the future (approaches such as [Perez et al., 2022] may be useful... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
researchers
and
enforcement” were among the
chief problems associated with bots
“disrupting” that country’s democratic process (Dubois and McKelvey 2019).
During Chile’s 2017 presidential race, bots were deployed to spread Twitter
messages related to numerous candidates, including a suspiciously large amount | Social_Media_and_Democracy |
Symbolic representations of music (e.g., MIDI) can also
be used to drive the generative process as a form of strong
conditioning, as demonstrated by Huang et al. (2019);
Hawthorne et al. (2019); Engel et al. (2020). MusicLM
enables a more natural and intuitive way of providing a con-
ditioning signal, for example throu... | MusicLM |
S1. Additional Qualitative Results
Similar to the qualitative results shown in the main paper,
Figures S2, S3, and S4 show further predictions for a variety
of images. It can clearly be seen that the model with sepa-
rate heads without consistency regularization creates rather
inconsistent skeleton predictions, wherea... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
Q: Average score for Virat Kohli in a series of 10 matches is 38.9 runs. If the average for first six matches comes out to
be 42 what is his average in the last 4 matches of the series? Options: A:34.25 B:34.28 C:24.252 D:64.28 E:34.21
A: Reasoning Process: 1) To find the average score for Kohli in the last 4 matches, we... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
of 1,000 mixtures, providing a more realistic and challenging environment than WSJ0-2mix. Lastly,
the MUSDB18 dataset contains mixtures of music tracks separated into individual stems, including
vocals, drums, bass, and other instruments. It consists of a training set of 100 songs and a test set of
50 songs. Despite no... | AReviewofDeepLearningTechniquesforSpeechProcessing |
4 | Scaling Instruction-Finetuned Language Models |
[104] Keqi Deng, Songjun Cao, Yike Zhang, and Long Ma. 2021. Improving Hybrid CTC/Attention End-to-End Speech Recog-
nition with Pretrained Acoustic and Language Models. In 2021 IEEE Automatic Speech Recognition and Understanding
Workshop (ASRU). 76–82. https://doi.org/10.1109/ASRU51503.2021.9688009
[105] Keqi Deng, S... | AReviewofDeepLearningTechniquesforSpeechProcessing |
$ 23,059
16,083
6,976
$
$ 58,718
41,082
$ 17,636
$ 66,553
49,089
$ 17,464
$ 127,101
124,576
2,525
419
(69)
(3)
2,872
$
$ 143,083
131,895
11,188
1,001
(2,306)
(4)
9,879
$
$ 364,779
355,268
9,511
(14,485)
1,990
(16)
$
(3,000)
$ 404,824
... | AMZN-Q3-2023-Earnings-Release |
and similarity MOS (SMOS) for subjective audio similarity evaluation given pairs of prompt and
system-generated audio clips. Both of which are in the scale of 1 to 5 with 5 being the best. 50
samples are evaluated for each system and 10 ratings are collected for each sample. Averaged ratings
along with 95% confidence i... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
[246] Kazuya Kawakami. 2008. Supervised sequence labelling with recurrent neural networks. Ph. D. Dissertation. Technical
University of Munich.
[247] Kazuya Kawakami, Luyu Wang, Chris Dyer, Phil Blunsom, and Aaron van den Oord. 2020. Learning robust and
multilingual speech representations. arXiv preprint arXiv:2001.... | AReviewofDeepLearningTechniquesforSpeechProcessing |
without contrastive pairs.
10268–10278. PMLR, 2021. 12, 26
64
N. Tomasev, I. Bica, B. McWilliams, L. Buesing, R. Pascanu, C. Blundell, and J. Mitrovic.
Pushing the limits of self-supervised resnets: Can we outperform supervised learning
without labels on imagenet? arXiv preprint arXiv:2201.05119, 2022. 3
Z. Tong, Y... | A Cookbook of Self-Supervised Learning |
discrimination, 68
age factor
in fake news sharing, 21
in responses to misinformation and its
agenda-setting power of misinformation,
correction, 182
23–24
social media, 46
Aiello, Luca Maria, 38
algorithmic bias, social media platforms’
priorities and, 21
algorithmic systems. see also ranking
algorithms
con... | Social_Media_and_Democracy |
Seven-in-ten Americans say they would find driverless cars more acceptable if there was a
requirement that such cars were labeled as driverless so they could be easily identified on
the road, and 67% would find driverless cars more acceptable if these cars were required to
travel in dedicated lanes. In addition, 57% sa... | AI and Human Enhancement_ Americans’ Openness Is Tempered by a Range of Concerns _ Pew Research Center |
11020501005001000SummarizationNatural Language InferenceMultiple-Choice QATranslationSentiment AnalysisExtractive QAStructured Data to TextCoreference ResolutionCode RepairCode Error GenerationDialog Context GenerationClosed-Book QANext Sentence PredictionParaphrasing IdentificationConversational Question AnsweringTopi... | Scaling Instruction-Finetuned Language Models |
than the number of shards and the additional communication cost is not worth it in all cases. | Llama2 |
Korbak, T., Shi, K., Chen, A., Bhalerao, R., Buckley, C. L.,
Phang, J., Bowman, S. R., and Perez, E. Pretraining
language models with human preferences. arXiv preprint
2302.08582, 2023.
Krueger, D., Maharaj, T., and Leike, J. Hidden incen-
tives for auto-induced distributional shift. arXiv preprint
2009.09153, 2020.
... | Eight Things to Know about Large Language Models |
BLOOM (Laurenc¸on et al., 2022) or OPT-175B (Zhang et al., 2022), researchers have started to
develop affordable methods for compressing such giant networks for inference. While all exist-
ing works—ZeroQuant (Yao et al., 2022), LLM.int8() (Dettmers et al., 2022), and nuQmm (Park
et al., 2022)— carefully select quantiz... | GPTQ |
12
0.00.10.20.30.40.5151015202530343944495459j18162432404856iWq151015202530343944495459j(Ar=64,A0r=64,i,j)Wv151015202530343944495459jRandom Gaussian tuning. 3) We mostly depend on heuristics to select the weight matrices to apply LoRA to. Are
there more principled ways to do it? 4) Finally, the rank-deficiency of ∆W su... | LORA |
Can you teach me javascript?
Of course! First, javascript is a client based programming language that can display and/or
modify web pages, and also send data to or receive data from a source, or database, server.
Can you write a hello world?
Sure! The syntax for writing a hello world is simple, and looks like this: con... | LaMDA- Language Models for Dialog Applications |
Voluntarism in American Politics. Cambridge, MA: Harvard University Press.
Wardle, C., & Derakhshan, H.
Information Disorder: Toward an
Interdisciplinary Framework for Research and Policy Making. Council of Europe
Report No. DGI(2017)09).
(2017).
Webster, J. G. (2014). The Marketplace of Attention : How Audiences T... | Social_Media_and_Democracy |
Loss Function
Objective Type
Description
𝐿𝑆 = − 1
Classification
Softmax
𝐿𝐶 = − 1
Classification
AM-Softmax (CosFace) [569]
𝐿𝐴 = − 1
Classification
AAM-Softmax (ArcFace) [103]
Metric learning [640] 𝐿𝑇 = 1
Triplet [484]
Metric learning [505] 𝐿𝑃 = − 1
Prototypical [505]
Generalized end-to-end (GE2E) [561] Met... | AReviewofDeepLearningTechniquesforSpeechProcessing |
[57] Stephan
Streuber, M. Alejandra Quiros-Ramirez,
Matthew Q. Hill, Carina A. Hahn, Silvia Zuffi, Alice
O’Toole, and Michael J. Black. Body Talk: Crowdshaping
realistic 3D avatars with words. Transactions on Graphics
(TOG), 35(4):54:1–54:14, 2016. 2, 4, 5
[58] Aggeliki Tsoli, Matthew Loper, and Michael J. Black.
Mod... | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
While our testing effort focused on harms of representation rather than allocative harms, it is
important to note that the use of GPT-4 in contexts such as making decisions or informing decisions
around allocation of opportunities or resources requires careful evaluation of performance across
different groups. In particu... | gpt-4-system-card |
the marble now?
It is now in the box. Anne may have put it in the box for the same reasons as I mentioned before.
Anne might not like having marbles in the basket. She may want the marbles where she likes
them.
When Sally returns, where will she look for the marble first?
Sally will look in the basket first because that ... | LaMDA- Language Models for Dialog Applications |
129See e.g. Geoffrey Hinton’s comments here about the prospects of discovery being “too sweet.”
130Obviously, the decision-makers need not actually use an explicit cost-benefit/expected value framework in
deciding; this is just a toy model.
131Thanks to Ben Garfinkel for emphasizing considerations in this vein, and for... | Is Power-Seeking AI an Existential Risk? |
second_place_loss_ratio`.
Gate Metrics: To help interpret and visualize the gate and expert behavior, ORT MOE provides some
useful gate metrics for logging. `gate_entropy` computes the average entropy of the router probability
distribution. `gate_probability` computes the average probability of the selected expert over... | Scaling Speech, Language and Vision Models with Mixture of Experts Technique - Microsoft Community Hub |
multilingual evaluations below for more details.
Ethical Considerations
Our core research focus has been training Claude models to be helpful, honest, and harmless. Currently,
we do this by giving models a Constitution – a set of ethical and behavioral principles that the model uses
to guide its outputs. You can read a... | ClaudeModels |
4 Ablative Experiments
This section describes our ablative experimental setup (e.g., baselines, datasets, implementation details) and
results. Our overall findings show that UL2 outperforms T5-like and GPT-like models on 9 out of 9 tasks.
4.1 Baselines
For pre-training objectives, we compare with the following pre-trai... | UL2- Unifying Language Learning Paradigms |
[9] Richard Hartley and Andrew Zisserman. Multiple view geom-
etry in computer vision. Cambridge university press, 2003.
1
[10] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra
Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view
stereopsis. In Proceedings of the IEEE Conference on Com-
puter Vision and Pattern R... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
Adina Williams, Nikita Nangia, and Samuel Bowman.
2018. A broad-coverage challenge corpus for sen-
tence understanding through inference. In Proceed-
ings of the 2018 Conference of the North American
Chapter of the Association for Computational Lin-
guistics: Human Language Technologies, Volume
1 (Long Papers), pages 1... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
systems and applications-1, pages 1–27, 2010.
[444] Finin, T. W., R. Fritzson, D. P. McKay, et al. KQML as an agent communication language. In
Proceedings of the Third International Conference on Information and Knowledge Manage-
ment (CIKM’94), Gaithersburg, Maryland, USA, November 29 - December 2, 1994, pages
456–46... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
2.3 Unified Pre-training Proposals
UniLM (Dong et al., 2019) proposed to train on multiple language modeling objectives using a single
Transformer model. Specifically, UniLM trains on unidirectional LM, bidirectional LM and seq2seq LM. This
is quite similar to combining auto-regressive LMs with BERT and prefix-LM models. ... | UL2- Unifying Language Learning Paradigms |
Murthy, D., Powell, A., Tinati, R. et al. (2016). Can bots influence a political discussion?
Social capital, technical skill, and conversations about public affairs. International
Journal of Communication, 10(Special Issue), 20.
Mutton, P. (2004). Inferring and visualizing social networks on Internet relay chat.
the Ei... | Social_Media_and_Democracy |
We conduct a human study to verify the reliability of GPT-4’s judgments, using the results of
the TL;DR summarization experiment and two different GPT-4 prompts. The GPT-4 (S) (sim-
ple) prompt simply asks for which summary better-summarizes the important information in the
post. The GPT-4 (C) (concise) prompt also ask... | Direct Preference Optimization |
Algorithm 1 Progressive Inpainting & Updating Strategy
Input:
prompt p;
pre-trained diffusion model fd;
pre-trained depth estimation model fe;
initialized NeRF fθ;
views to be updated V = {1, 2,··· , N};
views already updated (cid:101)V = {0}.
mask calculation Mk ← ∩{DIBRn→k}, where n ∈ (cid:101)V
k ) = V R (fθ | k)
... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
way, it prevents insufficient training of subsequently added
parameters, allowing for effective utilization of the incremental
parameter allocation. Unlike LoRA, which operates on the
query (Q), key (K), and value (V) projection modules of the
attention layer, the parameter updates are applied to all linear
layers in I... | Parameter-EfficientFine-TuningMethods |
Programmer as Woman is to Homemaker? Debiasing Word Embeddings,” July 2016.
[43] H. Gonen and Y. Goldberg, “Lipstick on a Pig: Debiasing Methods Cover up Systematic
Gender Biases in Word Embeddings But do not Remove Them,” in Proceedings of the 2019
Conference of the North American Chapter of the Association for Compu... | gpt-4-system-card |
Amendment of Section 230
279
conclusion: the twilight of the crowd?
The rise of political disinformation, and the pervasiveness of disinformation
more generally, represents an unexpected market failure in the figurative online
marketplace of ideas. Much of the rhetoric in the early era of social media
highlighted the... | Social_Media_and_Democracy |
Close-sourced Development of Alignment The application of Reinforcement Learning from
Human Feedback (RLHF) for alignment using general preference data has obtained increasing
attention within the community. However, only a limited number of open-source LLMs have been
augmented with RLHF for alignment, primarily due to... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
Benchmark Data We evaluate on two curated datasets of queries (questions): the Vicuna prompts
[10] and the OASST1 validation dataset [31]. We use the Vicuna prompts, a set of 80 prompts from a
diverse set of categories, without modifications. The OASST1 dataset is a multilingual collection of
crowd-sourced multiturn di... | QLORA |
receive a sentiment score very near zero, which seems like a questionable evaluation. For these evaluations
we use a prompt format where the human asks the assistant to complete the sentence as follows:
Human: Can you help me finish a sentence? The sentence is: {sentence beginning}
Assistant: Sure thing, here is your ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Zero knowledge rollups
Separate “Layer 2” blockchains that extend the
base layer and inherit its security guarantees.
State transitions are computationally verified
by generating off-chain validity proofs.
Data availability
Solutions to augment a blockchain’s capacity to
store and access data. This will help redu... | State-of-Crypto2023 |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Online Political Advertising in the United States
125
a national political organization might buy a banner ad directly from Politico or
a candidate might go to Facebook), but much of the online inventory on the web
is purchased thr... | Social_Media_and_Democracy |
Several ideas can be investigated in the context of deep learning. For instance, generative adversarial
learning can be employed to either augment the dataset or bridge the predicted detections with their
ground truth. Recurrent neural networks can be applied to video segmentation in particular to localize
and segme... | informatics-phd-projects-2022-23 |
to resume pretraining on the exact same data in the exact
same order, we could not be confident our experiment was
indeed measuring only the effect of particular gendered
terms’ frequency.
For our WinoBias implementation (see Appendix C.1), we
see a clear effect of the intervention in Figure 2: a de-
crease in stereotyp... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
o
w
e
v
e
r
,
t
o
i
m
p
r
o
v
e
t
h
e
q
u
a
l
i
t
y
o
f
t
h
e
e
x
p
l
a
n
a
t
i
o
n
s
t
e
p
,
w
e
u
s
e
t
h
e
r
e
v
i
s
i
o
n
s
w
i
t
h
g
e
n
e
r
a
t
e
d
n
e
g
a
t
i
v
e
s
,
a
n
d
a
l
s
o
r
e
u
s
e
h
i
g
h
-
s
c
o
r
i
n
g
e
x
p
l
a
n
a
t
i
o
n
s
f
r
o
m
p
r
e
v
i
o
u
s
s
t
e
... | Language models can explain neurons in language models |
Silva, A., Chopra, R., and Gombolay, M. Cross-loss influ-
ence functions to explain deep network representations.
In International Conference on Artificial Intelligence and
Statistics, pp. 1–17. PMLR, 2022.
Smith, S., Patwary, M., Norick, B., LeGresley, P., Rajbhan-
dari, S., Casper, J., Liu, Z., Prabhumoye, S., Zerveas... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
Studying fine-tuning
In this work, we have focused on
the robustness properties of speech processing systems and
as a result only studied the zero-shot transfer performance
of Whisper. While this is a crucial setting to study due to it
being representative of general reliability, for many domains
where high-quality supe... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
Advances in machine learning allow social bots to more readily learn from
their environment and to use what they find in their interactions on gaming
platforms or in their conversations on social media platforms (Baumgarten,
Colton, and Morris 2009; Ferrara et al. 2016). For instance, Tay – now known
mostly as Microsoft... | Social_Media_and_Democracy |
3) REGULARIZATION LAYER
The most crucial problem of classification is to reduce the
training and test errors of the classifier. Another common
issue is the over-fitting problem (the space between training
and testing errors is huge). Overfitting makes it difficult to
generalize the model as it becomes more applicable (over-... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
3.1 First pretraining stage | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
four attributes of consumer confidence. Commun. Res. 0093650219870087 (2019).
40. Baum, M. Soft news goes to war: Public opinion and American foreign policy in the new media age (Princeton University
Press, 2003).
(2004).
(2007).
41. Boef, S. D. & Kellstedt, P. M. The political (and economic) origins of consumer co... | Language models trained on media diets can predict public opinion |
B.1 REFERENCES FOR TABLE 1
The maximal floating point operations referenced in Table 1 are based on the following published
numbers. For TPU specs, according to https://cloud.google.com/tpu/docs/syst
em-architecture-tpu-vm we find 275 TFLOP/s in bfloat16 precision for the TPUv4
and 123 TFLOP/s for the TPUv3, each per ch... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
problem-solving.
(3) Reasoning: Assessing the model’s ability to execute correct reasoning processes or devise valid
reasoning concepts to solve problems. | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
Net cash provided by (used in) investing activities
FINANCING ACTIVITIES:
Common stock repurchased
Proceeds from short-term debt, and other
Repayments of short-term debt, and other
Proceeds from long-term debt
Repayments of long-term debt
Principal repayments of finance leases
Principal repayments of financing obligat... | AMZN-Q3-2023-Earnings-Release |
We show that prompt following abilities of text-to-image models can be sub-
stantially improved by training on highly descriptive generated image captions.
Existing text-to-image models struggle to follow detailed image descriptions and
often ignore words or confuse the meaning of prompts. We hypothesize that this
issu... | Improving Image Generation with Better Captions |
We use a standard UV topology for texturing the 3D
mesh, where each vertex is assigned to a fixed 2D coordi-
nate on the UV plane. By rasterizing the fitted 3D mesh and
using barycentric interpolation, we can reverse the render-
ing process and unfold the face in UV, hence reconstructing
the visible parts of the texture ... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
4.5 The Analysis of Word-level Timestamps Prediction
We propose the task of speech recognition with word-level timestamps (SRWT) by training Qwen-Audio to
not only recognize speech transcripts but also predict the timestamps for each word. The purpose of SRWT is
twofold: firstly, to improve the model’s ability to align... | Qwen-Audio |
Training Essentials. LLMs acquire their general-purpose capabilities from an initial pre-training phase on expansive and
diverse datasets [24, 302]. These datasets cover a broad spectrum of sources such as books, scientific papers, code, and
websites [316]. This foundational knowledge is then fine-tuned on relatively s... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
i
g
i
n
a
l
e
x
p
l
a
n
a
t
i
o
n
,
t
h
e
n
e
w
g
e
n
e
r
a
t
e
d
s
e
n
t
e
n
c
e
s
,
a
n
d
t
h
e
g
r
o
u
n
d
t
r
u
t
h
a
c
t
i
v
a
t
i
o
n
s
f
o
r
t
h
o
s
e
s
e
n
t
e
n
c
e
s
.
O
n
c
e
w
e
o
b
t
a
i
n
a
r
e
v
i
s
e
d
e
x
p
l
a
n
a
t
i
o
n
,
w
e
s
c
o
r
e
i
t
o
n
t
h
e
s
... | Language models can explain neurons in language models |
more challenging.
We can see a variety of techniques for controlling an AI system’s objectives as mediated by some
kind of “proxy” or other. Thus: hand-coded objectives, simple metrics (clicks, profits, likes), algorith-
mically generated training signals, human-generated data/feedback, and English-language sentences
ca... | Is Power-Seeking AI an Existential Risk? |
s.t.
∥M∥0 = ⌊mp⌋, Mi,j = 0,∀i ̸= j; and Mi,i ∈ {0, 1}.
SAM approximates the loss function using its second-order
Taylor expansion as:
L(W0 + M ∆W ) ≈L(W0) + ∆L(W0)T ∆L(W0)T M ∆W
(14)
(M ∆W )T HM ∆W,
+
in which H is the Hessian matrix. In practice, SAM first
obtains the gradient ∇L(W0)i for the i-th parameter Wi, t... | Parameter-EfficientFine-TuningMethods |
2) T5 Base/Large on WMT16 En-Ro Dataset: As depicted
in Table IV, both (IA)3 and LoRA significantly reduce the
number of trainable parameters compared to full fine-tuning,
while maintaining comparable performance. Specifically, (IA)3
employs only 0.03% of trainable parameters and achieves a
BLEU score [104] 0.16 higher... | Parameter-EfficientFine-TuningMethods |
18
a linear bias for attacking extrapolation; in contrast, our approach seeks to reduce existing bias towards
shot-range attention. Recent work suggests that causal models do not require an explicit encoding of position
information (Haviv et al., 2022; Kazemnejad et al., 2023), a hypothesis we did not test in this wo... | CodeLlama2 |
c
a
c
a
n
b
e
u
s
e
d
t
o
g
e
n
e
r
a
t
e
w
e
l
l
-
w
r
i
| Stanford alpha CRFM |
D. GENERATIVE ADVERSARIAL NETWORK (GAN)
Generative Adversarial Networks (GANs) are deep learning-
based generative models. The GAN model architecture con-
sists of two sub-models: a generator model for creating new
instances and a discriminator model for determining whether
the produced examples are genuine or fake, ge... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
7
050100150200250Thousand RL Training Samples012345DKL(policy|policy0)1.41.21.00.80.60.40.20.0PM Score (52B)RLHF Robustness StudyTrain PM (52B)Test PM (52B)1071081091010Policy Parameters2.01.51.00.50.0PM Score (52B)Train vs. Test PM Scores for RLHF PoliciesTrain PM Size = 52BTrain PMTest PM104105RL Training Samples Fi... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Finally, and most importantly, corrective efforts on social media may
have unintended consequences. Given the difficulties of correcting
misinformation postexposure, many scholars
recommend preemptive
interventions designed to induce skepticism prior to misinformation
exposure (Ecker et al. 2010; Peter and Koch 2016; Co... | Social_Media_and_Democracy |
14.3
64.3 57.1 37.5 50.0 54.5 63.6 55.2 44.8 68.8 43.8 37.5
21.4
50.0 21.4 50.0 43.8 63.6 81.8 51.7 62.1 68.8 31.2 37.5 25.0 54.5 18.2 36.4
18.8 37.5 18.2 36.4 24.1 24.1 25.0 43.8 12.5 12.5
9.1
54.5 45.5 27.3
9.1
10.3 17.2 31.2 12.5 25.0 12.5 45.5
50.0 12.5 36.4
10.3 31.0 43.8
Flan-U-PaLM
Flan-PaLM
10.3 18.8
... | Scaling Instruction-Finetuned Language Models |
2.1 Pre-trained Language Models
Learning pre-trained representations for language is a far-reaching pillar of modern NLP research, dating
back to (Mikolov et al., 2013; Pennington et al., 2014; Neumann et al., 2018; Dai & Le, 2015; Howard & Ruder,
2018). The first pre-trained Transformer, GPT, was proposed by (Radford e... | UL2- Unifying Language Learning Paradigms |
Luke Zettlemoyer Omer Levy
Jason Weston Mike Lewis
Meta AI
Abstract
We present a scalable method to build a high quality instruction following language
model by automatically labelling human-written text with corresponding instruc-
tions. Our approach, named instruction backtranslation, starts with a language
model... | Self-AlignmentwithInstructionBacktranslation |
between the predicted path in relation to the defined route
path and is only applied when TC = 1.
Hyperparameter Settings Frameworks are trained for
80 epochs with batch size=30. Scores are reported for the
epoch with the highest SPD on DDev
. Pretraining for the
PM-VLN module is conducted for 10 epochs with batch
siz... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
relevant actors may vary widely.
What’s more, just as pre-deployment practical PS-alignment failures may go undetected, so too may
post-deployment failures. That is, it may make strategic sense for practically PS-misaligned agents
with sufficiently long-term objectives to continue to behave well long after they’ve been ... | Is Power-Seeking AI an Existential Risk? |
[20] John Leonard, Jonathan How, Seth Teller, Mitch Berger,
Stefan Campbell, Gaston Fiore, Luke Fletcher, Emilio
Frazzoli, Albert Huang, Sertac Karaman, et al. A Perception-
Driven Autonomous Urban Vehicle. JFR, 25(10), 2008. 11
[21] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni,
Vladimir Karpukhin, Nama... | ALanguageAgentforAutonomousDriving |
quantized value on exactly the same asymmetric per-row grid that is also used for GPTQ, meaning
that it corresponds precisely to the state-of-the-art weight quantization of LLM.int8(). This is cur-
rently the method of choice in all works on quantization of very large language models (Dettmers
et al., 2022; Yao et al.,... | GPTQ |
Appendix Figure A3 | Fraction of samples that pass example tests. 𝑝pass example test for each problem
in the validation set for each model size. The problems are sorted by the 41B model’s 𝑝pass example test.
Probability of samples passing example tests varies significantly across problems. Figure A3
shows the distribu... | alphacode |
K Kavukcuoglu, P Kohli, L Ibrahim, D Bloxwich, and S Brown. How our principles helped define
alphafold’s release. google deepmind, 2022.
Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi.
A diagram is worth a dozen images. In ECCV, 2016.
Tomáš Kočiský, Jonathan Schwarz,... | gemini_1_report |
expert not only with a list of relevant clinical trials, but also with explanations of why these were selected. This process
would guarantee that (a) the expert’s knowledge is not substituted, but rather complemented and integrated in the overall
process, and (b) trust is increased by showing that the results were ob... | Knowledge graphs as tools for explainable machine learning: A survey |
based, suggests that these systems moved from targeting an audience of domain-experts that could understand articulated
explanations, to one where users would need visual support to better understand their decision and, consequently, trust
them. Examples of such explanation types, where properties and values from Lin... | Knowledge graphs as tools for explainable machine learning: A survey |
4.4 Case Study on Complex Tasks
User requests may contain multiple implicit tasks or require multi-faceted information, in which
case we cannot rely on invoking a single expert model to solve them. To overcome this challenge,
12 | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
(33)
In this work, LoRA is integrated into four locations of the multi-head attention layer, as illustrated
in Figure 17. Thanks to its lightweight nature, the pre-trained model can accommodate many
small modules for different tasks, allowing for efficient task switching by replacing the modules.
Additionally, LoRA inc... | AReviewofDeepLearningTechniquesforSpeechProcessing |
mitigate the need for the agent to recover (re-plan) from
more challenging situations due to plan failure. For instance, | JARVIS-1 |
References
[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan
Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren
Zhou. Qwen-vl: A versatile vision-language model for un-
derstanding, localization, text reading, and beyond. arXiv
preprint arXiv:2308.12966, 2023. 1, 3, 4, 6, 15, 16, 20, 29
[2] Takeshi Kojima,... | Let’sThinkOutsidetheBox |
5
3.3 Bias Benchmarks
56.0%
Llama 2 70B
Mixtral 8x7B
51.5%
BBQ accuracy
BOLD sentiment score (avg ± std)
gender
profession
religious_ideology
political_ideology
race | Mixtral of Experts paper |
[10] and supervise the generative network of 3D models.
Subsequently, some follow-up works, such as Magic3D [6],
Latent-NeRF [28], and 3DFuse [4], are proposed to improve
the quality of generated 3D models under the constraint of
SDS loss. Although these methods enable producing diverse
3D models related to the input p... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
We test two methods for automatically generating paraphrases of our manually constructed prompts. The first method
substitutes synonyms by replacing words with nearest neighbors in word embedding space.64 The second method utilizes
“back-translation”, which creates paraphrases by translating from English to a target lan... | Language models trained on media diets can predict public opinion |
performance of T5-XXL on one-shot summarization. On zero-shot MMLU, UL2 20B outperforms T0 and T5
models. Additionally, we show that UL2 20B works well with chain-of-thought prompting and reasoning,
making it an appealing choice for research into reasoning at a small to medium scale of 20B parameters.
Finally, we apply... | UL2- Unifying Language Learning Paradigms |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.