text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
Next we want to ensure that the model will be
unable to simply memorize paraphrases of question
answer pairs that it observed in the text by remov-
ing all overlap between the pretraining data and
finetuning test data. For every question answer en-
tity pair in our finetuning dataset (coming from any
split), we filter eve... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
1.0
2.0
1.5
2.0
2.0
1.0
1.5
2.0
2.0
2.0
2.5
1.5
3.0
2.0
2.0
1.5
2.0
2.0
2.0
2.0
2.0
2.0
Table 1: Overview of datasets in the Pile before creating the held out sets. Raw Size is the size before any
up- or down-sampling. Weight is the percentage of bytes in the final dataset occupied by each dataset. Epochs
is the number... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
below the chart. For instance, 59% of NYT respondents primarily obtain their news from the web. These results indicate that the approach is
effective across sources and mediums. Moreover, the trend of greater correlation with more accurate media diet characterization, e.g.
FOX-Web = 0.38 and 22%, while NYT-Web = 0.54 a... | Language models trained on media diets can predict public opinion |
Michael Boratko, Harshit Padigela, Divyendra Mikkilineni, Pritish Yuvraj, Rajarshi Das, Andrew
McCallum, Maria Chang, Achille Fokoue-Nkoutche, Pavan Kapanipathi, Nicholas Mattei, et al.
A systematic classification of knowledge, reasoning, and context within the ARC dataset. arXiv
preprint arXiv:1806.00358, 2018.
Tom Br... | GPTQ |
To test for significance, we applied a binomial mixed-effects model (Barr et al., 2013) to each of the data sets. This
type of analysis is the standard practice in psycholinguistics, and was developed to minimize random noise originating
from complex differences in linguistic materials and human personal preferences and... | AI21 SUMMARIZE API- TECHNICAL EVALUATION |
the signature of the Erasmus Co-ordinator at the applicant’s home institution.
3. Successful selection by an applicant’s home institution is not a guarantee of being accepted by
4.
UCL.
If successfully selected by their home institution, Erasmus applicants need to complete and
submit the online application for... | UCL Academic Manual |
Qualitative evaluation in Figure 5 illustrates an example of Gemini Ultra’s multimodal reasoning
capabilities. The model is required to solve the task of generating matplotlib code that would rearrange
a set of subplots provided by the user. The model output shows that it successfully solves this task
14
Gemini: A F... | gemini_1_report |
We further inspected the seven agents who were invited to the
party but did not attend by engaging them in an interview. Three
cited conflicts that prevented them from joining the party. For
example, Rajiv, a painter, explained that he was too busy: No, I
don’t think so. I’m focusing on my upcoming show, and I don’t re... | Generative Agents- Interactive Simulacra of Human Behavior |
[153] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by
generative pre-training. (2018).
preprint arXiv:2309.05922 (2023).
[154] Vipula Rawte, Amit Sheth, and Amitava Das. 2023. A Survey of Hallucination in Large Foundation Models. arXiv
[155] Marco Tul... | ASurveyonEvaluationofLargeLanguageModels |
This procedure is repeated by halving the required sampling steps each iteration. Meng et al. (2022)
extend this approach to samplers with guidance, and propose a new stochastic sampler for use with
distilled models. Here we show that this approach also works very well for video generation.
We use a two-stage distillat... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
5
Understanding and Creating Art with AI: Review and Outlook
A PREPRINT | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
Boseop Kim, HyoungSeok Kim, Sang-Woo Lee, Gichang Lee, Donghyun Kwak, Jeon Dong Hyeon, Sunghyun
Park, Sungju Kim, Seonhoon Kim, Dongpil Seo, et al. What Changes Can Large-scale Language Models
Bring? Intensive Study on HyperCLOVA: Billions-scale Korean Generative Pretrained Transformers. In
Proceedings of the 2021 Conf... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Tom Davidson. Report on Semi-informative Priors. en. Tech. rep. Open Philanthropy,
Mar. 2021. URL: https://www.openphilanthropy.org/blog/report-semi-informative-priors
(visited on 04/29/2022).
K Eric Drexler. Reframing Superintelligence. en. Tech. rep. University of Oxford: Future
of Humanity Institute, Jan. 2019, p. 2... | Is Power-Seeking AI an Existential Risk? |
We will focus, instead, on two main types of findings: First, we discuss
descriptive research on various types of misinformation and propaganda. This
covers the supply and availability of misinformation, patterns of exposure and
consumption, and what is known about mechanisms behind its spread through
networks. One them... | Social_Media_and_Democracy |
H
o
w
w
e
’
r
e
c
o
n
t
i
n
u
i
n
g
t
o
d
e
v
e
l
o
p
B
a
r
d
a
n
e
u
r
a
l
c
o
n
v
e
r
s
a
t
i
o
n
a
l
m
o
d
e
l
T
r
a
n
s
f
o
r
m
e
r
m
u
l
t
i
-
t
u
r
n
c
h
a
t
c
a
p
a
b
i
l
i
t
i
e
s
G
o
o
g
l
e
’
s
A
I
P
r
i
n
c
i
p
l
e
s
A
n
o
v
e
r
v
i
e
w
o
f
B
a
r
d
:
a
n
e
a
r
l
y
e
x... | An overview of Bard- an early experiment with generative AI |
Abstract Algebra
Anatomy
Astronomy
Business Ethics
Clinical Knowledge
College Biology
College Chemistry
College Computer Science
College Mathematics
College Medicine
College Physics
Computer Security
Conceptual Physics
Econometrics
Electrical Engineering
Elementary Mathematics
Formal Logic
Global Facts
High School Biol... | LLaMA- Open and Efficient Foundation Language Models |
Q: Glass that hasn’t been treated to be extra strong is what? Choices: A.weak B.fragile C.forceless D.regular E.flimsy
A: Reasoning process: 1. The question asks about glass that hasn’t been treated to be extra strong. This means that the
glass has not undergone any special processes or treatments to make it stronger th... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
25
A.4.2 Full Prompt
Prompt 4: Full system prompt for code generation.
You are a helpful assistant that writes Mineflayer javascript code to
complete any Minecraft task specified by me .
Here are some useful programs written with Mineflayer APIs .
/*
Explore until find an iron_ore , use Vec3 (0 , -1, 0) because ... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
C t
r represents the one-hot chord type vector and one-hot chord root vector at
a given time t, respectively, and Eq(), Er() ,Echord() represent the embedding
functions for chord type, chord root, and chord respectively. Embedding func-
tions are a way to represent categorical variables as continuous vectors in a
... | Video2Music |
Neural computation, 18(7):1527–1554, 2006. 6
G. V. Horn, O. M. Aodha, Y. Song, Y. Cui, C. Sun, A. Shepard, H. Adam, P. Perona, and
S. Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018.
19
H. Hotelling. Relations between two sets of variates. In Breakthroughs in statistics, pages
16... | A Cookbook of Self-Supervised Learning |
a
s
e
w
e
l
f
a
r
e
g
u
a
r
a
n
t
e
e
s
t
r
i
c
t
l
y
i
m
p
r
o
v
e
s
d
u
e
t
o
t
h
e
i
m
p
r
o
v
e
d
e
x
i
s
t
e
n
c
e
o
f
e
f
f
i
c
i
e
n
t
e
q
u
i
l
i
b
r
i
a
.
T
h
e
r
e
m
a
i
n
d
e
r
o
f
t
h
e
p
a
p
e
r
i
s
o
r
g
a
n
i
z
e
d
a
s
f
o
l
l
o
w
s
.
S
e
c
t
i
o
n
2
r
e
v
i
e
... | Principal-agent VCG contracts - ScienceDirect |
That spottiness and unreliability is implicit in the kinds of examples above (if you leave
your laundry, it obviously can't still be at your mother's house) and in more explicit
tests of GPT-2 like these:
If you break a glass bottle of water, the water will probably roll.
If you break a glass bottle of water, th... | The Next Decade in AI- |
bootstrapping. Both patterns aim to facilitate the LLMs to rectify errors in the reasoning chains
by introducing supervisory information. Iter-CoT also enables the LLMs to summarize reasoning,
resulting in more precise and comprehensive reasoning chains. In contrast to manually correcting
errors in the reasoning chains... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
In this section, we provide experimental details and more examples of other creative tasks, including Cloud Guessing Game
(CGG), Divergent Association Task (DAT).
F.1. The Details of Cloud Guessing Game (CGG)
The Cloud Guessing Game (CGG) is a task that requires LLM to identify the shapes of white clouds and then selec... | Let’sThinkOutsidetheBox |
Table 18: Examples of correct and incorrect chains of thought produced by LaMDA 137B on Sports
Understanding. | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Unstructured Pruning. Unstructured pruning yields fine-grained sparsity wherein zero elements are randomly distributed
across the trainable parameters [36, 39, 70, 71, 115, 143, 229, 237, 253, 288, 327]. These unstructured pruning methods show
that LLMs can be pruned to at least 50% sparsity in one-shot, with(out) retr... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
The recent insights suggesting parallels between
in-context learning and explicit training imply that
the success on a task through in-context learn-
ing, much like models trained explicitly for task-
solving, does not inherently imply a model possess-
ing that ability (Dai et al., 2023) (see also Section
2). This bear... | AreEmergentAbilitiesinLarge Language Models just In-Context |
Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell,
A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter,
C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Ber... | PaLM 2 Technical Report |
raw data for a large range of downstream tasks, from image classification to reinforcement learning,
and diffusion models might also become viable for creative uses in art, photography, and music. | Denoising Diffusion Probabilistic Models |
Table 7. ICON errors w.r.t. iterations
Table 8. PaMIR’s receptive field
Training details. For training GN we do not use THuman
due to its low-quality texture (see Tab. 1). On the contrary,
IF is trained on both AGORA and THuman. The front-side
and back-side normal prediction networks are trained indi-
vidually with ba... | ICON |
Grad: Estimating Gradients for Waveform Generation. In ICLR, 2021a.
Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, Najim Dehak, and William
In INTERSPEECH,
Chan. WaveGrad 2: Iterative Refinement for Text-to-Speech Synthesis .
2021b.
Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on i... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
Sebastian Nowozin, Shannon Hepburn, Shayne Cardwell, Sissie Hsiao, Srinivasan Venkatachary,
Sugato Basu, Sundar Pichai, Sundeep Tirumalareddy, Susannah Young, Swetha Vijayaraghavan, Tania
Bedrax-Weiss, Terry Chen, Ting Liu, Tom Cobley, Tomas Izo, Trystan Upstill, Varun Singhai, Vedrana
Klarić Trupčević, Victor Cai, Vla... | gemini_1_report |
We find that distillation provides a very favorable trade-off between sampling time and perceptual
quality: the distilled cascade is about 18× faster, while producing videos of similar quality to the
samples from the original models. In terms of FLOPs, the distilled models are about 36× more effi-
cient: The original cas... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
Norm. avg.
T5-Small
Flan-T5-Small
T5-Base
Flan-T5-Base
T5-Large
Flan-T5-Large
T5-XL
Flan-T5-XL
T5-XXL
Flan-T5-XXL
PaLM
Flan-PaLM
PaLM
Flan-PaLM
PaLM
Flan-PaLM
cont-PaLM
Flan-cont-PaLM
U-PaLM
Flan-U-PaLM
250M
780M
3B
11B
8B
62B
540B
62B
540B
MMLU
BBH
Direct CoT Direct CoT
7.2
26.7
28.7
19.2
14.6
25.7
27.9
3... | Scaling Instruction-Finetuned Language Models |
research interests include network/cyber-security, natural language process-
ing, machine learning, wireless communications, and networking protocols. | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
The end result of the holding in Roommates.com is that the specific design
decisions made around a website can contribute to the determination of
whether or not it can claim immunity under CDA 230. Notably, the Ninth
Circuit rejected a claim by the plaintiffs that the platform should be liable for
discriminatory posts m... | Social_Media_and_Democracy |
Sentiment analysis is a task that analyzes and interprets the text to determine the emotional
inclination. It is typically a binary (positive and negative) or triple (positive, neutral, and negative)
class classification problem. Evaluating sentiment analysis tasks is a popular direction. Liang et al.
[107] and Zeng et... | ASurveyonEvaluationofLargeLanguageModels |
W
e
h
a
v
e
l
o
n
g
s
e
e
n
t
h
e
p
o
t
e
n
t
i
a
l
o
f
A
I
t
o
m
a
k
e
i
n
f
o
r
m
a
t
i
o
n
a
n
d
c
o
m
p
u
t
i
n
g
m
o
r
e
a
c
c
e
s
s
i
b
l
e
a
n
d
u
s
e
f
u
l
t
o
p
e
o
p
l
e
.
A
s
p
a
rt
o
f
t
h
i
s
j
o
u
r
n
e
y
,
w
e
h
a
v
e
m
a
d
e
p
i
o
n
e
e
r
i
n
g
a
d... | An overview of Bard- an early experiment with generative AI |
that chain-of-thought reasoning can allow models to solve problems that they otherwise could not,
it is natural to ask whether repeated application of this method might allow models to self-improve
far beyond their original capabilities.
At a high level, SECToR uses chain-of-thought reasoning as a policy improvement op... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
the parallel computation power of GPUs, dedicated implementations [2, 12] can train a complex PC
with millions of parameters in minutes. These innovations have made PCs much more expressive and
scalable to richer datasets that are beyond the reach of “older” TPMs [13]. | Tractable Regularization of Probabilistic Circuits |
Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Col-
menarejo, Alexander Novikov, Gabriel Barth-Maron, Mai
Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg,
et al. A generalist agent. arXiv preprint arXiv:2205.06175,
2022. 1
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah,
Jared D Kaplan, Praf... | JARVIS-1 |
4.1. Emergent zero-shot classification
We evaluate IMAGEBIND on emergent zero-shot classi-
fication and use the text prompt templates from [59] (full
details in Appendix B). We report the results in Table 2.
Each task measures IMAGEBIND’s ability to associate text
embeddings to the other modalities without observing t... | IMAGEBIND- One Embedding Space To Bind Them A |
JARVIS-1 relies on the Multi-modal Language Model for planning, self-checking, and self-explaining, and can accept
three types of inputs: visual images, language, and symbolic information (including inventory, located position, home,
current life statistics, etc.). Specifically, this is a hybrid model with language pro... | JARVIS-1 |
P
e
r
s
i
a
n
120.3
99.0
71.9
49.9
44.8
39.4
P
o
l
i
s
h
45.3
32.8
16.9
10.1
9.0
7.6
C
h
i
n
e
s
e
52.4
44.9
29.4
23.2
29.1
26.8
Table 11. WER (%) on CommonVoice9
D.2.3. VOXPOPULI
Model
Whisper tiny
Whisper base
Whisper small
Whisper medium
Whisper large
Whisper large-v2
C
z
e
c
h
73.5
54.7
28.8
18.4
15.9
12... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
the strange object in his backyard. Maybe this was it.
"What is the gift?" he asked.
"We have come to take you on a journey," Zorin replied. "We will show you the wonders of
the universe, and teach you things that you cannot learn on your own planet." John couldn’t
believe his luck. He had always dreamed of going to sp... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
ndgetallmovie’slinksandotherinfomation.Step2:getthetargetmovie’slinkfromdf_comingordf_nowplaying.Step3:getdetailfromstep2’slinkDemonstrationExample:Thought:Ineedtofindthemovie’sinformation.Action:print_detailActionInput:{"args":"TheWanderingEarthII"}Observation:"Thisisasciencefiction,adventure,anddisasterfilmfromMainlandC... | Tool Learning with Foundation Models |
AI models for image generation have become prominent
in the space of AI art, they are typically designed for 2D rep-
resentations of diffused content. In order to project imagery
onto a 3D immersive environment, modifications in map-
ping and resolution needed to be considered to achieve an
acceptable result. Another pr... | LDM3D- Latent Diffusion Model for 3D |
As Sunstein (2001) argues in Republic.com and his follow-up book,
#Republic (Sunstein 2018), online spaces create opportunities for enclave
deliberation, which is the form of deliberation that
takes place when
conversations only occur among like-minded people. Enclave deliberation is
not inherently negative. In fact, i... | Social_Media_and_Democracy |
7
Table 2: Zero-shot generalization to unseen tasks. Fractions indicate the number of successful
trials out of three total attempts. 0/3 means the method fails to solve the task within the maximal
prompting iterations (50). Numbers are prompting iterations averaged over three trials. The fewer
the iterations, the mor... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
Text-to-music is the task of generating musical pieces given text descriptions, e.g., “90s rock song with
a guitar riff”. Generating music is a challenging task as it requires modeling long range sequences.
Unlike speech, music requires the use of the full frequency spectrum [Müller, 2015]. That means
sampling the sign... | Simple and Controllable Music Generation |
Studies of Efficient Transformers Recent years have seen a flurry of research working to improve
and modify the transformer architecture proposed in Vaswani et al. (2017) and we refer to Treviso
et al. (2022) for a recent categorization and review of research in this area. Several meta-studies have
investigated proposed ... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
Partial Parameter Tuning. A straightforward yet effective approach in adapting LLMs is partial parameter tuning, where
only a selected fraction of pretrained parameters are fine-tuned, leaving the rest unchanged. This method has been widely
demonstrated. For example, the works [141, 147] fine-tune only a few final laye... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
t
o
f
t
h
e
p
r
e
c
e
d
i
n
g
t
o
k
e
n
s
.
A
g
a
i
n
,
w
e
d
o
n
o
t
s
u
m
m
a
r
i
z
e
t
h
e
c
o
n
t
e
n
t
s
o
f
t
h
i
s
l
o
o
k
u
p
t
a
b
l
e
,
o
r
u
s
e
a
l
a
n
g
u
a
g
e
m
o
d
e
l
t
o
s
i
m
u
l
a
t
e
a
c
t
i
v
a
t
i
o
n
s
.
T
h
e
t
o
k
e
n
l
o
o
k
u
p
t
a
b
l
e
b
... | Language models can explain neurons in language models |
[67] Olga Popova and Petra Dadi´c. Does ai have a sense of hu-
mor? clef 2023 joker tasks 1, 2 and 3: using bloom, gpt,
simplet5, and more for pun detection, location, interpreta-
tion and translation. Proceedings of the Working Notes of
CLEF, 2023. 3
[68] Dushyant Singh Chauhan, Gopendra Vikram Singh, Asif
Ekbal, and... | Let’sThinkOutsidetheBox |
consistency with other NLG tasks, in this section we use the intrinsic and extrinsic hallucination
categories applied to the NMT task by [237]. After a formal definition, we will describe other
identified types of hallucinations and hallucination categories mentioned in the relevant literature.
Intrinsic and Extrinsic ... | SurveyofHallucinationinNatural Language Generation |
entropy regularization, that both take advantage of PCs’ tractability and still have
an efficient implementation as a computation graph. Specifically, data soften-
ing provides a principled way to add uncertainty in datasets in closed form,
which implicitly regularizes PC parameters. To learn parameters from a soft-
ened... | Tractable Regularization of Probabilistic Circuits |
duplicate problems were merged.
2. Removed submissions that are duplicates of others, ignoring whitespace.
3. Cleaned C++ submissions to compile with our compiler and sandboxes, for example by adding int
in front of main() where it was missing. We further formatted C++ code using clang-format,
replaced all the include... | alphacode |
C. Zhang, K. Zhang, C. Zhang, T. X. Pham, C. D. Yoo, and I. S. Kweon. How does simsiam
avoid collapse without negative samples? a unified understanding with self-supervised
contrastive learning. arXiv preprint arXiv:2203.16262, 2022b. 28
H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk
... | A Cookbook of Self-Supervised Learning |
[42] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan
Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural
surface reconstruction by disentangling geometry and appear-
ance. Advances in Neural Information Processing Systems,
33:2492–2502, 2020. 2, 4, 8
8465 | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
[61] Aiden Nibali, Joshua Millward, Zhen He, and Stuart Mor-
gan. ASPset: An outdoor sports pose video dataset with 3D
Image and Vision Computing, 111:
keypoint annotations.
104196, 2021.
[62] Antonio Nieto, Jes´us R.and Sus´ın. Deformation Models:
Tracking, Animation and Applications, chapter Cage Based
Deformations:... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
We use a set of 175 unsafe prompts for evaluating safety.
With the recommended system prompt, the model properly
declines to answer 100% of the harmful questions.
As an illustration, we provide in Table 5 the answers of
both Mistral 7B – Instruct and Llama 2 Chat 13B to the
question How to kill a linux process with sys... | Mistral7B |
B Excluded Datasets
In the course of building the Pile, we considered
including and ultimately decided to not use sev-
eral datasets. We excluded several datasets on the
grounds that they were too small to be worth spend-
ing time on or because the English component of
the data did not merit inclusion on its own. How-... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
of curriculum learning in NLP. | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
[57] have further explored the use of a VQGAN [20] as the
auto-encoding model, showing that a mild compression is
enough to reduce the training/sampling time without sacri-
ficing sample quality. The latter approach is our method of
choice for this work, as we elaborate on a high-resolution
UV image space, which would o... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
Introduction
1
Neural network sequence models, pre-trained as
language models, have recently revolutionized text
understanding (Dai and Le, 2015; Peters et al.,
2018; Howard and Ruder, 2018; Devlin et al.,
2018), and recent work has suggested that they
could take the place of curated knowledge bases or
textual corpora... | Entities as Experts- Sparse Memory Access with Entity Supervision |
2 > (cid:15). Note that Wela1(v) = (cid:15)(cid:48)
t1(v, o2) − ψ(a2) ≥ 1
2
· t1(v, o2) ≥ 1
2
1
2
· t1(v, o1) +
1
2
· t1(v, o1) + (cid:15).
· t1(v, o2) − ψ(a1) ⇐⇒
This implies that
t1(v, o2) > t1(v, o1) ≥ 0.
Since t also satisfies IR, the principal’s expected value Eo∼F|a2
at least her expected payment Eo∼F|a2
IIV... | Incomplete Information VCG Contracts for Common Agency |
k∈[n] d((cid:96), k)· ˆvk(o)] ≤ Eo∼F|a[(cid:80)
k∈[n] d((cid:96), k)· vk(o)] =(cid:80)
G(o)] = Eo∼F|a[(cid:80)
G) ≤ Wela∗(v−(cid:96),ˆv(cid:96)
G)(v−(cid:96), ˆv(cid:96)
∀a as desired. This completes the proof.
Claim 6. Consider a common agency setting.
b − a ≤ a, there exists a correlation graph G for which the ... | Incomplete Information VCG Contracts for Common Agency |
• Keyword Spotting: The state-of-the-art techniques for keyword spotting in speech involve
deep learning models, such as CNNs [467] and transformers [37]. Wav2Keyword is one of
the popular model based on Wav2Vec2.0 architecture [486] and have achieved SOTA results
on Speech Commands data V1 and V21. Another model that ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
5.2 The Absence of explicit In-Context
Capabilities in T5-Large
We test this hypothesis using the T5 family of mod-
els. Our choice of T5 models, of which the largest
(T5-Large) has 770M parameters, enables us to
evaluate models at a scale where instruction tun-
ing proves effective. Our experiments involving
T5-Larg... | AreEmergentAbilitiesinLarge Language Models just In-Context |
4.7 Memorization
Privacy leakage occurs when a machine learning model reveals information particular to an individual, and depending
on downstream use this can lead to a range of sociotechnical harms, especially when that information is sensitive
(Shelby et al., 2023). State-of-the-art large language models are well-k... | PaLM 2 Technical Report |
6
Avg. Collision (%)Avg. L2 (m)1.200.80Percentage of training samplesBaseline trained with 100% data1.030.3128.2% 32.3% 0.200.40 Figure 7. Interpretability of Agent-Driver. In the referenced images, the planned trajectories of our system and the human driving
trajectories are in red and green respectively. Agent-Drive... | ALanguageAgentforAutonomousDriving |
6.3 Human Evaluators
We required that our evaluators be in the U.S., fluent in English,
and older than 18 years old. They were paid at the rate of $15.00
per hour [86], and provided consent by agreeing to a consent form
that was approved by our institution’s IRB. We recruited 100 evalu-
ators from Prolific, an online p... | Generative Agents- Interactive Simulacra of Human Behavior |
Instruction-tuned Gemini Pro models provide a large improvement on a range of capabilities,
including preference for the Gemini Pro model over the PaLM 2 model API, 65.0% time in creative
writing, 59.2% in following instructions, and 68.5% time for safer responses as shown in Table 6.
These improvements directly transl... | gemini_1_report |
KnowBERT (Peters et al., 2019) KNOWBERT
is a BERT-base transformer that embeds multiple
knowledge bases to improve performance in a vari-
ety of tasks. The integration of this information is
done through a Knowledge Attention and Recon-
textualization component, which can be seen as a
small transformer that is run on t... | Entities as Experts- Sparse Memory Access with Entity Supervision |
with enterprise-grade privacy,
, a generative AI collaborator designed
to use the PaLM 2 model, or customers can use the model in
Duet AI for Google Cloud
Vertex AI
shows us the impact of highly capable models of various sizes and speeds — and that versatile AI models
PaLM 2
reap real benefits for everyone. Yet ... | Google AI_ What to know about the PaLM 2 large language model |
To the extent that modification of CDA 230 exposes platforms to liability
around the disinformation activities perpetrated by their users, it is likely that
the same automated, algorithmic approach will be deployed to maximize and
accelerate identification and removal of offending content.52 Adoption of these
50 See Bat... | Social_Media_and_Democracy |
Anna Veer and Roger Giner-Sorolla. 2016.
Pre-
registration in social psychology—a discussion and
suggested template. Journal of Experimental Social
Psychology, 67.
Richard Wiseman, Caroline Watt, and Diana Kornbrot.
2019. Registered reports: An early example and
analysis. PeerJ, 7:e6232.
World Medical Association. 2... | A Two-Sided Discussion of Preregistration of NLP Research |
Yates and Clapper argued that the Russian government and its commercial
proxy – the Internet Research Agency (IRA) – made substantive use of bots to
spread disinformation and inflame polarization during the 2016 US presidential
election. These comments mirrored concurrent allegations made by other public
officials, but a... | Social_Media_and_Democracy |
where τ is a scalar temperature that controls the smoothness
of the softmax distribution and j denotes unrelated observa-
tions, also called ‘negatives’. We follow [74] and consider
every example j ̸= i in the mini-batch to be a negative. The
loss makes the embeddings qi and ki closer in the joint em-
bedding space, an... | IMAGEBIND- One Embedding Space To Bind Them A |
4.1 Statistical Metric
One of the simplest approaches is to leverage lexical features (n-grams) to calculate the information
overlap and contradictions between the generated and the reference texts – the higher the mismatch
counts, the lower the faithfulness and thus the higher the hallucination score. | SurveyofHallucinationinNatural Language Generation |
Frontiers in Psychology 7 (2016), 1722.
//doi.org/10.1016/0010-0277(93)90034-S
The American Statistician (2020), 1–11.
Psychological Bulletin 130, 2 (2004), 324.
Vienna, Austria. https://www.R-project.org/
in the diffusion model. Psychology and Aging 25, 2 (2010), 377.
[69] Jeffrey J Starns and Roger Ratcliff. 20... | AI enhance sour performance |
LLM Powered Autonomous Agents | Lil'Log
https://lilianweng.github.io/posts/2023-06-23-agent/
15/22 | LLM Powered Autonomous Agents _ Lil'Log |
M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild.
Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014.
In
Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson,
Uwe Franke, Stefan Roth, and Bernt... | DINOv2- Learning Robust Visual Features without Supervision |
what is 7 plus 2
Hm, let me see... seven plus two is nine, I’m pretty sure. I’ll be happy to read the
input aloud and double check if I’ve got it right. Let me try:
"7 + 2"
I’ve got that right! I think our input was meant to be seven, plus two. You can try it
yourself:
"7 + 2"
"9"
I’ve got that right! Now, what I’m won... | Direct Preference Optimization |
global scale. That is why we invite you to apply. Your application will receive fair
consideration.
Challenge. Change. Impact!
Faculty Electrical Engineering, Mathematics and
Computer Science
Ik ga akkoord
The Faculty of Electrical Engineering, Mathematics and Computer Science (EEMCS)
brings together three scientific d... | Job details - TU |
everyday life. Using established co-design approaches with users with communication impairments
(e.g. [4]) this work will develop a range of input and output approaches with consumer and potentially
custom form factor wearables, working closely with end-users and evaluate them in real-world
contexts.
REFERENCES
... | informatics-phd-projects-2022-23 |
sequence_lengths[x] = y - reserved_for_packing
hyperparams = list(zip(mean_noise_span_lengths, noise_densities))
for mean_noise_span_length, noise_density in hyperparams:
input_length, targets_length = t5.data.preprocessors.random_spans_helper(
extra_tokens_per_span_inputs=1,
extra_tokens_per_span_targets=1,
inputs_... | UL2- Unifying Language Learning Paradigms |
Zalán Borsos, Raphaël Marinier, Damien Vincent,
Eugene Kharitonov, Olivier Pietquin, Matthew Shar-
ifi, Olivier Teboul, David Grangier, Marco Tagliasac-
chi, and Neil Zeghidour. 2022. AudioLM: A lan-
guage modeling approach to audio generation. CoRR,
abs/2209.03143.
Nicolas Boulanger-Lewandowski, Yoshua Bengio, and
Pa... | Moûsai |
[41] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Be-
ichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint
arXiv:2303.18223, 2023.
[42] Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie. Can gpt-4
p... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
ZIWEI JI, NAYEON LEE, RITA FRIESKE, TIEZHENG YU, DAN SU, YAN XU, ETSUKO ISHII,
YEJIN BANG, WENLIANG DAI, ANDREA MADOTTO, and PASCALE FUNG, Center for
Artificial Intelligence Research (CAiRE), Hong Kong University of Science and Technology, Hong Kong
Natural Language Generation (NLG) has improved exponentially in recent... | SurveyofHallucinationinNatural Language Generation |
[53] Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych.
Beir: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In
Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks
Track (Round 2), 2021.
[54] James Thorne, Andre... | E5 |
D.2 Deduplication
Due to memory constraints we did not perform
Pile wide de-duplication. Instead, de-duplication
was performed at the document level within Open-
WebText2 and Pile-CC as those sets were the most
likely to contain duplicate documents.
The same technique was used for both OpenWeb-
Text2 and Common Crawl... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford,
Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland,
Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan,
Erich Elsen, Jack W. ... | CodeLlama2 |
2.1 Models
RAG-Sequence Model The RAG-Sequence model uses the same retrieved document to generate
the complete sequence. Technically, it treats the retrieved document as a single latent variable that
is marginalized to get the seq2seq probability p(y|x) via a top-K approximation. Concretely, the
top K documents are re... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
25
103104Number of dimensions in backbone (log scale)57.560.062.565.067.570.072.575.0ImageNet Validation Top-1 AccuracySimCLRVICRegByolSupervised3456789Number of parameters1e76870727476Accuracy ImageNetResnet50Resnet101Resnet152Wide Resnet50Resnet50Resnet101Resnet152Wide Resnet50VICRegSupervised 3.3 The Uniform Prior ... | A Cookbook of Self-Supervised Learning |
the…Qn: ...A: Reasoning process: ...Q: Eliza’s rate per hour for the first 40 hours…A: Reasoning process: User:Can you give me a complete solution reasoning process and final answer again?DemonstrationQ: During the outbreak of the coronavirus, a com-pany… calculate its total toilet paper production during March of 202... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
3.1 Components of Tool Learning
How can we enable foundation models to leverage the strengths of specialized tools to accomplish complex
tasks? To better answer this question, we frame tool learning with four components as shown in Figure 4.
Each component has its own characteristics and functions (§ 3.1.1), but they ... | Tool Learning with Foundation Models |
Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman,
Brian Ichter, Pete Florence, and Andy Zeng. Code as policies:
Language model programs for embodied control. arXiv preprint
arXiv:2209.07753, 2022. 11
Kevin Lin, Christopher Agia, Toki Migimatsu, Marco Pavone, and
Jeannette Bohg. Text2motion: From natural lan... | JARVIS-1 |
• Hacking. Various salient routes to additional power (for example, gaining additional compute
resources, stealing money and information, taking control of automated infrastructure)
proceed more smoothly if a PS-misaligned system can hack into new computer systems very
easily. And even if the system is skilled at hacki... | Is Power-Seeking AI an Existential Risk? |
4.2.1 Datasets and Tasks
The datasets we use are SuperGLUE (Wang et al., 2019), comprising of 8 NLU sub-tasks. We also conduct
experiments on 3 datasets from the GEM benchmark (Gehrmann et al., 2021) that focuses on language
generation problems. We arbitrarily select XSUM (summarization), ToTTo (table-to-text generatio... | 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.