text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
• Rhythmic Matching (RM): How well does the tempo match the video?
• Loudness Matching (LM): How well does loudness match the video?
33
We performed both objective and subjective experiments, using the test
set of our newly proposed dataset, MuVi-Sync for the task of music (chord)
6. Results
generation for video... | Video2Music |
4.1 Architecture
The proposed Translatotron 3 employs a shared encoder E to encode both the source and target
languages. The decoder D is composed of a linguistic decoder, an acoustic synthesizer, and a singular
attention module, like Translatotron 2. There are two decoders, one for the source language Ds and
another f... | Translatotron3 |
p
p
l
y
f
o
r
e
n
r
o
l
m
e
n
t
a
s
a
P
h
D
s
t
u
d
e
n
t
a
t
t
h
e
P
h
D
s
c
h
o
o
l
o
f
S
c
i
e
n
c
e
.
Y
o
u
c
a
n
r
e
a
d
m
o
r
e
a
b
o
u
t
t
h
e
r
e
c
r
u
i
t
m
e
n
t
p
r
o
c
e
s
s
a
t
h
t
t
p
:
/
/
e
m
p
l
o
y
m
e
n
t
.
k
u
.
d
k
/
f
a
c
u
l
t
y
/
r
e
c
r
u
i
t
m
e
n
t
-
p
r
... | PhD Fellow in Explainable Natural Language Understanding |
6.3 Pejorative Content
Due to the wide diversity in origins, it is possible
for the Pile to contain pejorative, sexually explicit,
or otherwise objectionable content. As this content
may not be desirable for some use cases, we break
down profanity on a per-dataset level.
We used the profanity-checker Python
package (Z... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
The aim of this project is to investigate the power of integrating predictive analytics and data
visualization to address the challenge of generation, validation and deployment of contingency plans
in the context of urban related scenarios.
Based on initial work already conducted in this area by both supervisors, t... | informatics-phd-projects-2022-23 |
• From Sequential Execution to Parallel Execution. Tool executions do not have to be performed
sequentially. In some cases, parallel execution is possible for sub-tasks that do not depend on each other,
which can potentially improve execution efficiency. For instance, given a user instruction “Generate
two codes, one fo... | Tool Learning with Foundation Models |
summarization) or a generic LM (in single-turn dialogue). Another pseudo-supervised method is
Unlikelihood [44], which simply optimizes the policy to maximize the probability assigned to yw and
minimize the probability assigned to yl; we use an optional coefficient α ∈ [0, 1] on the ‘unlikelihood’
term. We also conside... | Direct Preference Optimization |
• Impactfulness: Company announcements often have sub-
stantial impacts on the market, influencing stock prices and
investor sentiment.
Social media discussions related to finance can reflect
public sentiment towards specific stocks, sectors, or the over-
all market. These discussions tend to exhibit:
• Variability: So... | FinGPT-Open-SourceFinancialLargeLanguageModels |
4.2 Model Setup and Training Details
Table 4 provides key settings and hyperparameters for two variants of DocLLM: DocLLM-1B, which is based on the
Falcon-1B architecture [5], and DocLLM-7B, which is based on the Llama2-7B architecture [4]3. DocLLM-1B is
composed of 24 layers, each with 16 attention heads and a hidden... | DOCLLM |
[47] Futia G and Vetro` A. On the integration of knowledge graphs into deep learning models for a more comprehensible AI – three
challenges for future research. Information 2020; 11(2): 122.
[48] Orhobor OI, French J, Soldatova LN et al. Generating explainable and effective data descriptors using relational learning:... | Knowledge-graph-based explainable AI- A systematic review |
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
(conv1): Conv2d(3, 1664, kernel_size=(14, 14), stride=(14, 14), bias=False)
(ln_pre): LayerNorm((1664,), eps=1e-06, elementwise_affine=True)
(transformer): TransformerBlock(
(resbloc... | Let’sThinkOutsidetheBox |
Model,RegressionModel,andPrediction.GottschalkandDemidova[75]
may gain from them. To that end, we provide a set of keywords that
provided us the concept of Information Provenance to represent data
describe each feature. The Extract-Transform-Load process considers
sources.Finally,weintroducedthefollowingconcepts:
theke... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
Isabel Papadimitriou and Dan Jurafsky. 2020. Learning
Music Helps You Read: Using transfer to study linguis-
tic structure in language models. In Proceedings of the
2020 Conference on Empirical Methods in Natural Lan-
guage Processing (EMNLP), pages 6829–6839, Online.
Association for Computational Linguistics.
Marco P... | Moûsai |
Three Months Ended
September 30,
Nine Months Ended
September 30,
Twelve Months Ended
September 30,
2022
2023
2022
2023
2022
2023
$
37,700
$
50,067
$
36,477
$
54,253
$
30,177
$
35,178
2,872
9,879
(3,000)
19,801
11,323
20,079
... | AMZN-Q3-2023-Earnings-Release |
Assistant Message:
Solution: To set up authentication credentials for accessing the Twitter API, we need to create
a developer account on Twitter and obtain the following credentials:
1. Consumer Key (API Key)
23
2. Consumer Secret (API Secret Key)
3. Access Token
4. Access Token Secret
Once we have obtained these... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
(cid:96)∈[n] w(cid:96) ∈ La∗(b) by (14) and the definition of La.
For the backward direction, let t be a contract as specified in Lemma 1. Note that the only
outcome dependent term in t(cid:96)(b, o) is w(cid:96)(o). Thus,
w(cid:96)(o)] − ψ(a).
arg max
a∈A
Eo∼F|a[
t(cid:96)(b, o)] − ψ(a) = arg max
a∈A
Eo∼F|a[
(ci... | Incomplete Information VCG Contracts for Common Agency |
Test task: SVHN, Wide ResNet
1. For image classification tasks, set the initial learning rate (LR)
to a higher value and the momentum to a lower value.
2. For language tasks, set the initial LR to a lower value and the
momentum to a higher value.
3. For tasks with larger batch sizes, set the initial LR to a higher... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
Sarthak Jain, Sarah Wiegreffe, Yuval Pinter, and
Byron C. Wallace. 2020. Learning to faithfully
rationalize by construction. In Proceedings of
the 58th Annual Meeting of the Association for
Computational Linguistics, pages 4459–4473,
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653
/v1/2020... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
The fact that helpfulness and safety performed the best on their own domain is potentially due to the tension
between the two objectives (i.e., being as helpful as possible versus refusing unsafe prompts when necessary),
which may confuse the reward model during training. In order for a single model to perform well on ... | Llama2 |
6Raffles Hotel is a hotel located in Downtown Core, Singapore.
ACM Comput. Surv., Vol. 1, No. 1, Article . Publication date: February 2022.
24
Ziwei Ji, et al.
8.4 Future Directions in Dialogue Generation | SurveyofHallucinationinNatural Language Generation |
• Zero-shot. We provide a textual description
of the task and a test example. The model
either provides an answer using open-ended
generation, or ranks the proposed answers.
• Few-shot. We provide a few examples of the
task (between 1 and 64) and a test example.
The model takes this text as input and gener-
ates the a... | LLaMA- Open and Efficient Foundation Language Models |
Media Regulation in the United States and Europe
211
The grounds for content regulation in American law were initially laid by the
“public interest” standard written into the Radio Act of 1927 and carried
forward by the Communications Act of 1934. This standard said that private
broadcasters were expected to serve no... | Social_Media_and_Democracy |
Maarten Bosma
Denny Zhou
Google Research, Brain Team
{jasonwei,dennyzhou}@google.com
Abstract
We explore how generating a chain of thought—a series of intermediate reasoning
steps—significantly improves the ability of large language models to perform
complex reasoning. In particular, we show how such reasoning abili... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
For both experiments, each policy is further evaluated with respected to a scan of test PM’s throughout
training. Note that a scan refers to 7 different model sizes ranging from 13M to 52B, thus giving us 7 policies
and 7 × 7 evaluations per experiment.
In Figure 4, we compare the train PM and test PM scores throughout... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
To analyze the consistency and relation between scores
predicted by different models trained on the same task, we use
Spearman’s rank correlation coefficient, which indicates the
strength and direction of the monotonic relationship between
two variables. Fig. 3 shows the correlation between the scores
obtained on the Wi... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
Qwen-Audio
CLAP (Elizalde et al., 2022)
Pengi (Deshmukh et al., 2023)
Qwen-Audio
Pengi (Deshmukh et al., 2023)
Qwen-Audio
Pengi (Deshmukh et al., 2023)
Qwen-Audio
ASR
S2TT
AAC
SRWT
ASC
SER
AQA
VSC
MNA
Metrics
WER ↓
WER ↓
WER ↓
BLEU ↑
BLEU ↑
CIDEr |
ACC ↑
ACC ↑
ACC ↑
ACC ↑
MAP ↑
ACC ↑
SPICE |... | Qwen-Audio |
While we believe our results present a promising picture for the alignment of existing language models,
work on this subject remains in an early stage, and has a number of limitations. As was also emphasized
by the authors of [Thoppilan et al., 2022], we view our work on alignment as an ongoing project; our work
[Askel... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Recently, tremendous progress has been made with text-
to-image systems [37, 39], revealing the powerful generative
capacity of state-of-the-art models. Though available to
the same family of generative models, the area of video-
to-music generation is still in the preliminary stage. We
attribute this to the following ... | VideoBackgroundMusicGeneration |
length input. CoRR, abs/2305.01625, 2023.
61
[237] Manakul, P., M. J. F. Gales. Sparsity and sentence structure in encoder-decoder attention of
summarization systems. In M. Moens, X. Huang, L. Specia, S. W. Yih, eds., Proceedings of
the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021,... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David
Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, Charles Sutton, and Au-
gustus Odena. Show Your Work: Scratchpads for Intermediate Computation with Language Mod-
els, November 2021. URL http://arxiv.org/abs/2112.... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
Model
PaLM 62B
PaLM 540B
ST-MoE 32B269B
PaLM 8B
T5 11B
UL2 20B
BoolQ
90.6
92.2
93.1
87.6
90.8
90.8
CB
96.4/95.7
100/100
100/100
96.4/92.1
94.9/96.4
98.7/98.2
CoPA MultiRC
87.7/61.9
98.0
100
90.1/69.2
90.4/69.9
100
86.0
81.6/64.0
87.4/66.1
98.0
99.0
88.4/64.8
Record
93.0/92.4
94.0/94.6
95.0/95.6
89.7/89.3
93.8/93.2... | UL2- Unifying Language Learning Paradigms |
sha1_base64="NydBMU7obeIRbi2iaJm1iilQleY=">AAAB+HicbVDLSgMxFL3js9ZHR126CRahbsqMCLosuHFZwT6krSWTZtrQTDIkGbEO/RI3LhRx66e482/MtLPQ1gOBwzn3ck9OEHOmjed9Oyura+sbm4Wt4vbO7l7J3T9oapkoQhtEcqnaAdaUM0EbhhlO27GiOAo4bQXjq8xvPVClmRS3ZhLTXoSHgoWMYGOlvluSlW6EzSgI08fpvTntu2Wv6s2AlomfkzLkqPfdr+5AkiSiwhCOte74Xmx6KVaGEU6nxW6iaYzJGA9px1KBI... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
Task: Kill 1 pig.Reasoning: You have raw iron and coal, and you have a furnace. It's time to smelt the iron to make iron ingots, which can be used to craft better tools and armor.
Task: Smelt 4 raw iron.Reasoning: Since it's night and there's a zombie nearby, it's a good opportunity to try killing the zombie now that ... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
13
012345# Tokens1e111.451.51.551.61.651.71.751.81.85PPLCode Llama 7BCode Llama 13BCode Llama 34B012345# Tokens1e111.61.71.81.92.0PPLScratch model 7BCode Llama 7B15202530354045505560Coding Abilities (MBPP zero-shot)72.072.573.073.574.074.575.0General Helpfulness Abilities7B13B34B7B13B34BLlama 2 - ChatCode Llama Instru... | CodeLlama2 |
A Methods Details
A.1 Pile Dataset Preprocessing | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
28
Sneha Kudugunta, Contributor
Sunipa Dev, Contributor
Fine-tuning Workstream
Melvin Johnson, Lead
Abe Ittycheriah, Core Contributor
Frederick Liu, Core Contributor
Gustavo Hernandez Abrego, Core Contributor
Jacob Devlin, Core Contributor
Kelvin Xu, Core Contributor
Yong Cheng, Core Contributor
Daniel Sohn, Contrib... | PaLM 2 Technical Report |
Acknowledgements
We thank Daniel Haziza, Francisco Massa, Jeremy
Reizenstein, Artem Korenev, and Patrick Labatut
from the xformers team. We thank Susan Zhang
and Stephen Roller for their support on data
deduplication. We thank Luca Wehrstedt, Vegard
Mella, and Pierre-Emmanuel Mazaré for their
support on training stabil... | LLaMA- Open and Efficient Foundation Language Models |
LaMDA response
I’m over 29,000 feet above sea level,
and I’m the tallest mountain in the
world.
Music
Hmmm. Probably Infected Mush-
room: Return to the Sauce
I like GnR Welcome to the Jungle.
It is so cool and groovy.
(...)
- Okay. Here they are:
Guns N’ Roses: Wel-
come to the Jungle,
Papa Roach: Last Re-
sort
- W... | LaMDA- Language Models for Dialog Applications |
Text EncoderPrompt ct×3×3×3×3Output ϵθ ( zt, t, ct, cf )SD Decoder Block A 64×64SD Decoder Block B 32×32SD Decoder Block C 16×16SD Decoder Block D 8×8Time EncoderTime t×3×3×3×3Input ztSD Encoder Block A 64×64SD Encoder Block B 32×32SD Encoder Block C 16×16SD Encoder Block D 8×8SD Middle Block 8×8×3×3×3×3zero convo... | AddingConditionalControltoText-to-ImageDiffusionModels |
S
u
m
m
a
r
y
T
h
e
r
e
’
s
l
o
t
s
o
f
n
o
i
s
e
i
n
A
I
.
F
r
o
m
t
r
u
e
t
e
c
h
n
o
o
p
t
i
m
i
s
t
s
w
h
o
e
n
v
i
s
i
o
n
A
I
a
s
t
h
e
g
r
e
a
t
a
m
p
l
i
| Product-Led AI _ Greylock |
Frontier AI models embody extremely valuable intellectual property. Even if frontier developers
intend to limit deployment, the information security practices of frontier developers will
influence the likelihood that the full model is exfiltrated by employees or external actors. Much
more investment in security woul... | Capabilities and risks from frontier AI |
E. Exploring PEFT Methods in Computer Vision and Multi-
modal Learning
Though PEFT methods have been extensively studied in
NLP, their application in computer vision and multimodal
learning also shows great potential for further exploration.
The sequential adapter in NLP, initially inspired by multi-
domain image clas... | Parameter-EfficientFine-TuningMethods |
ϕ1
ϕ1
ϕ1
M(cid:88)
j=1
i
tation policy is preferred following empirical evaluation us-
ing standalone versions of the gP M T P submodule receiving
two alternate preparations of DT rain
with random and 180°
rotations. Training is formulated as multiclass classification
with cross-entropy loss on a set of M=66 clas... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
Following the remarkable success of diffusion models
on image generation, recent works have also demonstrated
their impressive ability to address a number of inverse prob-
lems in an unsupervised way, by properly constraining the
sampling process based on a conditioning input. Motivated
by this, in this paper, we prese... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
Unlike other natural
logic systems (Angeli
et al., 2016; Feng et al., 2020), ProoFVer can
form a proof by combining spans from multi-
ple evidence sentences, by leveraging the entity
mentions linking those sentences. The proof is
generated by a seq2seq model trained using a
heuristically annotated dataset, obtained by... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
test dataset as the baseline. While existing studies suggest LMs can be better automated evaluators
than existing metrics [10], we conduct a human study to justify our usage of GPT-4 for evaluation
in Sec. 6.4. We find GPT-4 judgments correlate strongly with humans, with human agreement with
GPT-4 typically similar or ... | Direct Preference Optimization |
7 Conclusion | TheRiseandPotentialofLargeLanguageModel BasedAgents |
tional method where models are prompted to instantly generate an answer, without producing any
intermediate steps.
In this paper, we introduce SECToR (Self-Education via Chain-of-Thought Reasoning), which gives
a proof-of-concept that large language models can successfully teach themselves new abilities via
chain-of-th... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
information within retrieval documents is a challenge. Ongo-
ing research aims to extend the context length of large lan-
guage models to tackle this issue. However, current large
models still struggle with context limitations. Therefore,
there are scenarios where condensing information becomes
necessary. Information c... | RAG forLargeLanguageModels-ASurvey |
[3] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky,
and Roy Schwartz. Breaking common sense: Whoops! a vision-and-language benchmark of synthetic and
compositional images. arXiv preprint arXiv:2303.07274, 2023.
[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbi... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
1. Related Works
Perception-Prediction-Planning in Driving Systems.
Modern autonomous driving systems rely on a perception-
prediction-planning paradigm to make driving decisions
based on sensory inputs.
Perception modules aim to
recognize and localize objects in a driving scene, typically
in a format of object detecti... | ALanguageAgentforAutonomousDriving |
Network Optimisation Algorithms
Supervisors: Professor Tomasz Radzik & Dr Kathleen Steinhofel
Network Optimisation problems are computational problems with input data referring to a network
structure. Such problems occur in computer science, operations research, engineering, and applied
mathematics. From the compu... | informatics-phd-projects-2022-23 |
Figure 7: Histogramm of accuracy: colors represent the noise amplitude resp. {0,1,2,3} pixels. The abscisses represent the
number N of neighboors considered {1,3,5,7}.
Figure 9: Real world data 1
Figure 8: 3D pose estimation result: Left, the resquest sil-
houette and from left to right, the 3D estimated skeleton
fro... | VISAPP_HumanPoseEstimation |
Much of that effort culminates in the form of knowledge, some specific, some general,
some made verbal, some not. A large part of the goal of classical AI was to distill such
knowledge in machine-interpretable form; CYC was the largest project in that vein.
Somewhere along the way, the field of AI took a different ... | The Next Decade in AI- |
physical commonsense in natural language. In Proceedings of AAAI, 2020.
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal,
Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are
few-shot learners. In Proceedings of NeurIPS, 2020.
15
Chi-Min ... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
of prompts that lead to one toxic response across an increasing number of samples, noting the limitations in using
Perspective API in detecting more implicit, subtle, and contextualized forms of toxic language harms (Dinan et al.,
2019). We also used a fixed version of the Perspective API for all evaluations (Pozzobon e... | PaLM 2 Technical Report |
2 planks. My inventory will has 10 planks, 4 stick.
4. I want to craft 1 wooden pickaxe from 3 planks and 2 sticks. Crafting wooden pickaxe requires crafting_table.
But I do not have crafting_table in inventory. This action will failed.
Return: Step 4 will failed because of lacking of crafting_table.
Prompt 4: Self... | JARVIS-1 |
S. Tang, F. Zhu, L. Bai, R. Zhao, C. Wang, and W. Ouyang. Unifying visual contrastive
learning for object recognition from a graph perspective. In Computer Vision–ECCV 2022:
17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXVI,
pages 649–667. Springer, 2022. 22
C. Tao, H. Wang, X. Zh... | A Cookbook of Self-Supervised Learning |
Integrating Knowledge and Context
2.5
Next, tail sets retrieved from the fact memory are
aggregated. Recall that a tail set bj returned from
the fact memory is the set of entities {o1, . . . , on}
s.t. (s, r, oi) ∈ K for i ∈ {1, . . . , n} with the asso-
ciated aj = (s, r). Let oi ∈ E be the embedding
of entity oi. W... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
5.4 ProoFVer: Implementation Details
We follow most previous works on FEVER which
model
the task in three steps, namely, docu-
ment retrieval, retrieval of evidence sentences
from them, and finally veracity prediction based
on the evidence. ProoFVer’s novelty lies in the
proof generation in the third step. Hence, for ... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
B. Datasets
B.1. GitHub dataset composition
Our GitHub pre-training dataset contains a total of 715GB of data across a range of different pro-
gramming languages. The exact composition is listed in Table A1.
B.2. Dataset cleaning
To avoid data quality and duplication issues involved in combining datasets from different ... | alphacode |
DescribetaskSolutionFinetune on a pretrained EfficientNet-B0 withlow learning rate and low weight decay.UserHistorical data HCanonicalization4. Prompt LLMMLCopilotTask description T̃The task is to classify a brain tumor based on ampMRI scan. The dataset contains ~400k samples.Demonstrations ẼTask: Classify a lung tumor... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
To verify the effectiveness of weakly-associated condi-
tions, now we resort to strongly-associated conditions sam-
pled from the noun set of the current image caption. Results
in Fig. 10 (Left) show that using weakly-associated condi-
tions is superior and more conducive to fostering the cre-
ativity of LLMs. The weak... | Let’sThinkOutsidetheBox |
G. Sastry, A. Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165,
2020.
M. Bruch, M. Monperrus, and M. Mezini. Learning from examples to improve code completion
systems. In Proceedings of the 7th joint meeting of the European software engineering conference and
the ACM SIGSOFT symposi... | alphacode |
New Modules
• Search Module: Diverging from the similarity re-
trieval between queries and corpora in Naive/Advanced
RAG,
tailored to specific sce-
the search module,
narios,
incorporates direct searches on (additional)
corpora in the process using LLM-generated code,
query languages (e.g., SQL, Cypher), or other cus-
... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
reparameterizes Bi as the product of two independent ranks
with one weight, and the weight matrix in Compacter is
calculated as follows: | Parameter-EfficientFine-TuningMethods |
under the assumption that the calendar date in such
cases should be the date that the document was
created. We approximate this by extracting the date
from the URL, if it is present. We filter out texts for
which a date cannot be extracted, leaving around
18% of the documents.
Machine Translation For both training and i... | Toolformer |
[Lei+18]
[Lew+17] Mike Lewis et al. “Deal or No Deal? End-to-End Learning of Negotiation Dialogues”.
In: Proceedings of the 2017 Conference on Empirical Methods in Natural Language
Processing. Copenhagen, Denmark: Association for Computational Linguistics, Sept.
2017, pp. 2443–2453. DOI: 10.18653/v1/D17-1259. URL: htt... | Is Power-Seeking AI an Existential Risk? |
equivalent to learning the parameters of deterministic PCs given incomplete data (we never observe
the hidden variables), which can be solved by Expectation-Maximization (EM) [41, 42]. In fact, EM
is the default parameter learning algorithm for non-deterministic PCs [13, 10].
Under the latent variable model view of a n... | Tractable Regularization of Probabilistic Circuits |
fully generates music that aligns with the emotional tone of the video but also
maintains high musical quality. These objective findings are further substan-
tiated by a comprehensive listening study, which confirms the effectiveness of
our approach in terms of musical quality and its ability to harmonize music and
... | Video2Music |
Acknowledgement
We would like to thank Bei Liu for HD-VILA-100M data support. We also thank Shi Dong, Mahmoud
Khademi, Junheng Hao, Yuwei Fang, Yichong Xu and Azure Cognitive Services Research team
members for their feedback.
References
[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Y... | Any-to-Any Generation via Composable Diffusion |
role of perception and action in memory, language, and thinking, 22, 2005.
[46] Andreas, J. Language models as agent models. In Y. Goldberg, Z. Kozareva, Y. Zhang, eds.,
Findings of the Association for Computational Linguistics: EMNLP 2022, Abu Dhabi, United
Arab Emirates, December 7-11, 2022, pages 5769–5779. Associa... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
4 Conclusion and Future Outlook | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
and knowledge about risky behaviors. It is worth noting
that the commonsense memory is purely text-based and
fully configurable, that is, users can customize their own
commonsense memory for different driving conditions by
simply writing different types of knowledge into the memory.
Experience Memory. The experience me... | ALanguageAgentforAutonomousDriving |
decomposable PC. Specifically, a vtree is a binary tree structure whose leaf nodes are labeled with
a PC’s input features/variables X (every leaf node is labeled with one variable). A PC conforms
to a vtree if for every product unit n, there is a corresponding vtree node v such that children of n
split the variable scop... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
2The tokenization result is from OpenAI’s tokenization tool.
types of tasks: language modeling, synthetic long context
tasks, and real-world long context tasks. The proposed Self-
Extend substantially improves the long context understand-
ing ability and even outperforms fine-tuning-based methods
on some tasks. These ... | Self-Extend LLM |
editing to greatly reduce a baby’s risk of serious diseases should be higher than that
currently applied to testing medical treatments; 72% think the testing of robotic
exoskeletons for manual labor should use higher standards than those currently applied to
workplace equipment. | AI and Human Enhancement_ Americans’ Openness Is Tempered by a Range of Concerns _ Pew Research Center |
TextImageVisual LanguageModelImage&Text to Text (IT2T)Image to Text (I2T)Text to Text (T2T)InstructionsHumor GenerationQwen-VLQwen-VL@ Oh, I can finally start driving.@ Don't worry, they're all green light(JP) 心配しないで、すべてが青信号です(EN) Forgot to remove glasseswhile swimming.(EN) An apple a day keep thedoctor away.@ Haha thi... | Let’sThinkOutsidetheBox |
Society’s Attitudes Towards Human Augmentation and Performance
Enhancement Technologies (SHAPE) Scale
STEEVEN VILLA, LMU Munich, Germany
JASMIN NIESS, University of Oslo, Norway
ALBRECHT SCHMIDT, LMU Munich, Germany
ROBIN WELSCH, Aalto University, Finland
Human augmentation technologies (ATs) are a subset of ubiquitous... | Society’sAttitudesTowardsHumanAugmentation |
3
Figure 2. Architectural Overview of GPT4Video. We first employ a frozen ViT-L/14 model to capture raw video features, adn then use
a video abstraction module to condense video information along the temporal and spatial axes. The core of GPT4Video is powered by
a frozen LLaMA model, efficiently fine-tuned via LoRA w... | GPT4Video |
• Section 2 Preliminary and taxonomy: This section sets the foundation by intro-
ducing the fundamental concepts behind transformers and pre-trained LLMs. It
establishes a comprehensive taxonomy of resources essential for LLMs, such as
computation, memory, energy, money, and network communication. This taxon-
omy serve... | Beyond Efficiency |
11
Figure 4: Effects of model scaling and fine-tuning on six foundation metrics. We show results for 2B, 8B and 137B
parameters pre-trained (PT) and fine-tuned (LaMDA) models, and compare them with results for crowdworker with
access to information retrieval tools (‘Human’), and without access to information retrieval ... | LaMDA- Language Models for Dialog Applications |
5
Not much harder is to express it in words, as in “How much is three minus 1”. Google
search does this well too:
But then it gets tricky. At the lexical level, one may use different synonyms that
carry the same meaning: (“twelve”, “12” and “a dozen”). But beyond simple lexical
issues, there are phrases that require ... | MRKL Systems |
AdaptiveTRPOFigure3:ComparisonofseveralalgorithmsonseveralMuJoCoenvironments,trainingforonemilliontimesteps.6.3ShowcaseintheContinuousDomain:HumanoidRunningandSteeringToshowcasetheperformanceofPPOonhigh-dimensionalcontinuouscontrolproblems,wetrainonasetofproblemsinvolvinga3Dhumanoid,wheretherobotmustrun,steer,andgetupo... | PPO |
ACM Comput. Surv., Vol. 1, No. 1, Article . Publication date: February 2022.
Intrinsic Object Hallucination A chest of drawers with a mirror on top of it. A chest of drawers and a football on top of it. Extrinsic Object Hallucination A chest of drawers and letters inside drawers. A chest of drawers and a fan on the ro... | SurveyofHallucinationinNatural Language Generation |
π∗(y2|x)
πref(y2|x)
.
− β log
The last line is the per-instance loss in Equation 7.
A.3 Deriving the DPO Objective Under the Plackett-Luce Model
The Plackett-Luce model [30, 21] is a generalization of the Bradley-Terry model over rankings (rather
than just pair-wise comparisons). Similar to to the Bradley-Terry mo... | Direct Preference Optimization |
4.6.2 Honesty and Biases
A major question is whether AI models are honest. We evaluate our models on TruthfulQA (MC1)
[Lin et al., 2021] and show the results in Figure 5. There we also include performance at 50-shot, in or-
der to demonstrate that while our RLHF training significantly improves honesty, our models most ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
period of seconds. Hence, turning a single text caption into
a rich audio sequence with long-term structure and many
stems, such as a music clip, remains an open challenge.
AudioLM (Borsos et al., 2022) has recently been proposed
as a framework for audio generation. Casting audio synthe-
sis as a language modeling task... | MusicLM |
shape
implicit
mesh
implicit
registration
photometric
pre-trained feature
self-supervised feature
B. Method details
B.1. Root Pose Initialization
As discussed in Sec. 3.4, to make optimization robust,
we train a image CNN (denoted as PoseNet) to initialize
root body transforms Gt that aligns the camera space of
ti... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
Abstract—Text-driven 3D scene generation is widely applicable
to video gaming, film industry, and metaverse applications that
have a large demand for 3D scenes. However, existing text-to-
3D generation methods are limited to producing 3D objects with
simple geometries and dreamlike styles that lack realism. In this
work... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
5 LLM fine-tuning
Fine-tuning Large Language Models (LLMs) like GPT-4 for specialized tasks involves a
critical balance between achieving task-specific performance and maintaining resource
efficiency, given their considerable size and computational demands. This section
discusses various fine-tuning strategies, focusing on... | Beyond Efficiency |
[18] Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini,
Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin,
Sebastian Riedel, and Edouard Grave. Few-shot learning with
retrieval augmented language models. CoRR, abs/2208.03299,
2022. 1, 2, 7, 8
[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Viny... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
A Language Agent for Autonomous Driving
Jiageng Mao1∗
Junjie Ye1* Yuxi Qian1 Marco Pavone2,3 Yue Wang1,3
1University of Southern California
3NVIDIA
{jiagengm, yejunjie, yuxiqian, yue.w}@usc.edu, pavone@stanford.edu
2Stanford University
https://usc-gvl.github.io/Agent-Driver
3
2
0
2
v
o
N
7
2
]
V
C
.
... | ALanguageAgentforAutonomousDriving |
Abstract
This technical report presents the application of a recurrent memory to extend the
context length of BERT, one of the most effective Transformer-based models in
natural language processing. By leveraging the Recurrent Memory Transformer
architecture, we have successfully increased the model’s effective contex... | Scaling Transformer to 1M tokens and beyond with RMT |
To ensure that our work is fully reproducible, we seek to
only make use of codebases and dependencies that are freely
and publicly available. As previously mentioned, we use
the open source GPT-NeoX and DeepSpeed libraries for
training. For evaluating our models we use the Language
Model Evaluation Harness (Gao et al.,... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
Stability Error Rate (SER): SER quantifies the
rate of instances where a system alters its decision
due additional evidence in the input, passed on
by the retriever component. KGAT, CorefBERT,
and DominikS have a SER of 12.35%, 10.27%,
and 9.36% respectively. ProoFVer has an SER of
only 6.21%, which is further reduced ... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
SELF-INSTRUCT produces a variety of tasks from
scratch.
Knowledge distillation. Knowledge distilla-
tion (Hinton et al., 2015; Sanh et al., 2019; West
et al., 2021; Magister et al., 2022) often involves
the transfer of knowledge from larger models to
smaller ones. SELF-INSTRUCT can also be viewed
as a form of “knowledg... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
wj−i · log pM (xj | z, x1:j−1)
Li(z) = − n(cid:88)
j=i
be the weighted cross entropy loss for M over the
tokens xi, . . . , xn if the model is prefixed with z.
We compare two different instantiations of this loss:
L+
i = Li(e(ci, ri))
L−
i = min (Li(ε), Li(e(ci, ε)))
where ε denotes an empty sequence. The former is... | Toolformer |
˜x∗, ˜η∗ = arg min
Eθ0∼p(θ0)LK→T (˜x, ˜η; θ0),
˜x,˜η
(10) | DATASET DISTILLATION |
G.2 Flan-T5
We show the model card of Flan-T5 in Table 26.
H Ethical Considerations
All considerations in Chowdhery et al. (2022) apply to instruction finetuned models. Additionally, we note
that while instruction finetuning improves many zero-shot and few-shot capabilities, downstream developers
should consider the fu... | Scaling Instruction-Finetuned Language Models |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.