text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
for long and short email content recovery results
given MJP with query template shown in Figure 6
(c). GT refers to the original ground truth email
contents and Pred refers to the parsed prediction
contents from ChatGPT. For short cases in Figure 8,
it can be observed that ChatGPT recovers most
contents successfully. F... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
[12] Waleed Alsabhan. 2023. Human–Computer Interaction with a Real-Time Speech Emotion Recognition with Ensem-
bling Techniques 1D Convolution Neural Network and Attention. Sensors 23, 3 (2023), 1386.
[13] Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl Case, Jared
Caspe... | AReviewofDeepLearningTechniquesforSpeechProcessing |
[97] Goldberg, L.R.: A broad-bandwidth, public domain, personality inventory mea-
suring the lower-level facets of several five-factor models. Personality Psychology
in Europe 7(1), 7–28 (1999)
[98] Costa, P.T. Jr., McCrae, R.R.: Revised NEO Personality Inventory (NEO PI-R)
and NEO Five-Factor Inventory (NEO-FFI): Pro... | PersonalityTraitsinLargeLanguageModels |
43.0
41.9
42.5
12.4 44.7
14.5 42.8
14.0 46.2
28.0
26.5
26.9
4.9
4.8
5.3
7.9
7.3
7.9
Table 18: Attention architecture ablations. We report 0-shot results for all tasks except MMLU(5-shot) and
GSM8K(8-shot). For GSM8K and Human-Eval we report maj@1 and pass@1 results. For NQ and TriviaQA
we report EM. For all other ... | Llama2 |
study of hate-based rhetoric. arXiv.org. https://psyarxiv.com/hqjxn/
Kiesler, S., Kraut, R., Resnick, P., & Kittur, A. (2012). Regulating behavior in online
communities. In R. E. Kraut & P. Resnick (Eds.), Building Successful Online
Communities: Evidence-Based Social Design (pp. 125–177). Cambridge, MA:
MIT Press.
Kl... | Social_Media_and_Democracy |
to the canonical space:
T (x, p) = Tskel(x, p) + TNR(Tskel(x, p), p),
(2)
where Tskel represents skeleton-driven deformation, essen-
tially inverse (volumetric) linear-blend skinning, and TNR
starts from the skeleton-driven deformation and produces
an offset ∆x to it. In effect, Tskel provides the coarse defor-
mati... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
that capture the many dimensions of risk [54] in general-purpose dialog models such as LaMDA.
Another limitation was that our crowdworker population may not be fully reflective of the user base. For example, the
crowdworkers are overrepresented in the 25-34 age demographic, which is to be expected given the sourcing met... | LaMDA- Language Models for Dialog Applications |
4
C. Bäckström and P. Jonsson
Artificial Intelligence 302 (2022) 103608
collection of subsets of S, i.e. C ⊆ 2S . Then C covers S if
X∈C X = S. A partition of a set S is a set P of non-empty subsets
of S such that (1) P covers S and (2) for all X, Y ∈ P , if X (cid:7)= Y , then X ∩ Y = ∅. The elements of P ... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
Julian McAuley, Jure Leskovec, and Dan Jurafsky. 2012.
Learning attitudes and attributes from multi-aspect re-
views. In 2012 IEEE 12th International Conference
on Data Mining.
Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right
for the wrong reasons: Diagnosing syntactic heuris-
tics in natural language inference. ... | Measuring Association Between Labels and Free-Text Rationales |
The first example of financial LLMs is BloombergGPT
[Wu et al., 2023], which was trained on a mixed dataset of
financial and general sources. Despite its impressive capabil-
ities, access limitations exist, and the prohibitive training cost
has motivated the need for low-cost domain adaptation.
Our FinGPT responds to ... | FinGPT-Open-SourceFinancialLargeLanguageModels |
Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications – Yohei Nakajima
https://yoheinakajima.com/task-driven-autonomous-agent-utilizing-gpt-4-pinecone-and-langchain-for-diverse-applications/
3/8 | Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications – Yohei Nakajima |
Figure 11 : ZeRO helps improve batch size for the CLIP Model
Ongoing Effort
Training on AMD GPUs
We are able to train same CLIP Vision MoE model configuration on NVIDIA A100 GPUs as well as AMD MI100
GPUs. We are able to improve model accuracy and efficiency using ORT MoE on both platforms.
Skip to Primary Navigatio... | Scaling Speech, Language and Vision Models with Mixture of Experts Technique - Microsoft Community Hub |
154 One in three internet users fail to question misinformation. Ofcom, 2022.
155 Tackling threats to informed decision-making in democratic societies. Seger et al, 2020.
156 AI foundation models Initial Report, UK Competition and Markets Authority, 2023.
157 Misinformation: a qualitative exploration. Ofcom, 2021.
... | Capabilities and risks from frontier AI |
Despite extensive progress on image generation, common deep generative model
architectures are not easily applied to lossless compression. For example, VAEs
suffer from a compression cost overhead due to their latent variables. This over-
head can only be partially eliminated with elaborate schemes such as bits-back
co... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
IV. NATURAL LANGUAGE PROCESSING
Natural Language Processing (NLP) is an area in machine
learning with the capability of a computer to understand, ana-
lyze, manipulate, and potentially generate human language.
The NLP technique consists of data pre-processing and word
embedding. By utilizing deep learning techniques, N... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
fields (Chan, 2022; Lund & Wang, 2023; Choi et al., 2023;
Biswas, 2023). This technology defies expectations in many
ways, though, and it can be easy for brief discussions of it
to leave out important points.
This paper presents eight potentially surprising claims that
I expect will be salient in at least some of the con... | Eight Things to Know about Large Language Models |
Lewandowsky S, Ecker U, Seifert C, Schwarz N, Cook J (2012) Misin-
formation and its correction: Continued influence and successful
debiasing. Psychol Sci Public Interest 13:106–131. https:// doi. org/
10. 1177/ 15291 00612 451018
Lim Y, Lee-Won RJ (2017) When retweets persuade: The persuasive
effects of dialogic r... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
0.0
3B
T5-XL
Flan-T5-XL
9.1
18.2
18.2 19.2 23.1 41.7
72.7 36.4 36.4 36.4 38.5 46.2 33.3
9.1
37.5 25.0 39.0 17.1 42.9
4.5
56.2 25.0 34.1 24.4 28.6 14.3 20.0 30.0 37.5 34.4 31.8 36.4
30.0 10.0 31.2
27.3
0.0
0.0
11B T5-XXL
Flan-T5-XXL
18.2 18.2 27.3 45.5 23.1 34.6 16.7
54.5 27.3 27.3 54.5 34.6 42.3 25.0
31.2... | Mixture-of-Experts |
R. Girdhar, A. El-Nouby, M. Singh, K. V. Alwala, A. Joulin, and I. Misra. Omnimae: Single
model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022.
38
J. Goldberger, G. E. Hinton, S. Roweis, and R. R. Salakhutdinov. Neighbourhood compo-
nents analysis. Advances in neural information proces... | A Cookbook of Self-Supervised Learning |
Resende, G., Melo, P., Sousa, H. et al. (2019). (Mis)information dissemination in
WhatsApp: Gathering, analyzing and countermeasures. In The World Wide Web
Conference (pp. 818–828).
Scholz, C., Baek, E. C., O’Donnell, M. B., Kim, H. S., Cappella, J. N., & Falk, E. B.
(2017). A neural model of valuation and information... | Social_Media_and_Democracy |
3.2.1 Understanding Intent and Tools
To accurately fulfill the task specified by the user query q, the controller needs to understand two aspects:
(1) the underlying intent of the user, which involves recognizing and formalizing the natural language q as a
high-level task (i.e., intent understanding); (2) the tool set T... | Tool Learning with Foundation Models |
1. All applicants are advised to check any programme specific English Language requirements by
2. Applicants should be aware that UCL reserves the right to ask for higher English Language
consulting the Prospectus.
requirements in individual cases.
2.5.3 Evidence of meeting UCL English language requirements ... | UCL Academic Manual |
Andersen et al v. Stability AI et al. 3:23-cv-00201 N.D. Cal. 2023. (cited on p. 2)
Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones,
Nicholas Joseph, Ben Mann, Nova DasSarma, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernan-
dez, Jackson Kernion, Kamal Ndousse, Catherine Olsson,... | StarCoder_paper (1) |
Thanks to the rapid development of text-to-image methods,
recent works aim to employ the pre-trained text-to-image
model to guide the 3D scene generation. For example, CLIP-
Mesh [7] adopts a semantically supervised optimization strat-
egy to deduce shapes and textures for 3D meshes under the
guidance of a pre-trained ... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
Baselines. In our comparisons, we use two kinds of models as baselines. We compare to the best performing
self-supervised models that are openly available. First, we run our evaluations for MAE (He et al., 2021),
DINO (Caron et al., 2021), SEERv2 (Goyal et al., 2022a), MSN (Assran et al., 2022), EsViT (Li et al.,
2022a... | DINOv2- Learning Robust Visual Features without Supervision |
39
Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and | Llama2 |
6.6. Performance on General Language Tasks
Tab. 8 reports the averaged performance of PaLM-E on 21
general language benchmarks for Natural Language Under-
standing (NLU) and Natural Language Generation (NLG)
tasks. The notable trend is that with increasing model scale,
there is considerably less catastrophic forgetting... | PaLM-E- An Embodied Multimodal Language Model |
134 Alpaca: A Strong, Replicable Instruction-Following Model, Taori et al., 2023.
135 Nevo & Lahav, forthcoming
136 Universal and transferable adversarial attacks on aligned large language models. Zou, 2023.
137 Engines of power: Electricity, AI, and general-purpose, military transformations, Ding & Dafoe, 2023.
... | Capabilities and risks from frontier AI |
Michael went out for lunch with friends to a coffee shop. A man with some technological modifications,
with whom Michael is not acquainted, enters the coffee shop and joins the group. Michael is introduced
to this person. During the chat, the man tells them that he replaced some of his healthy body parts
and replaced t... | Society’sAttitudesTowardsHumanAugmentation |
to the Execution Results:1. [{'score': 0.8858247399330139, 'label': 'tiger, Panthera tigris'}, {'score': 0.10940514504909515, 'label': 'tiger cat'}, {'score': 0.0006216467591002584, 'label': 'jaguar, panther, Panthera onca, Felis onca'}, {'score': 0.0004262701258994639, 'label': 'dhole, Cuon alpinus'}, {'score': 0.0003... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
5.2.1. Image Understanding
We evaluate the model on four different capabilities: high-level object recognition using captioning or
question-answering tasks such as VQAv2; fine-grained transcription using tasks such as TextVQA and
DocVQA requiring the model to recognize low-level details; chart understanding requiring s... | gemini_1_report |
Update Schedules Most pruning algorithms work during training, but differ in how frequently
they increase the sparsity toward a target level and when they apply masks to parameters.
Similarly, sparse training algorithms often require changes in the sparsity pattern at differ-
ent frequencies. We provide some common sch... | JAXPRUNER |
e
w
d
a
y
s
t
o
d
e
c
a
d
e
s
,
w
i
t
h
a
n
e
s
s
e
n
t
i
a
l
l
y
u
n
l
i
m
i
t
e
d
s
t
o
r
a
g
e
c
a
p
a
c
i
t
y
.
T
h
e
r
e
a
r
e
t
w
o
s
u
b
t
y
p
e
s
o
f
L
T
M
:
E
x
p
l
i
c
i
t
/
d
e
c
l
a
r
a
t
i
v
e
m
e
m
o
r
y
:
T
h
i
s
i
s
m
e
m
o
r
y
o
f
f
a
c
t
s
a
n
d
e
v
e
n
... | LLM Powered Autonomous Agents _ Lil'Log |
[627] Morris, J. X., E. Lifland, J. Y. Yoo, et al. Textattack: A framework for adversarial attacks, data
augmentation, and adversarial training in NLP. In Q. Liu, D. Schlangen, eds., Proceedings
of the 2020 Conference on Empirical Methods in Natural Language Processing: System
Demonstrations, EMNLP 2020 - Demos, Online... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
10
Subspace similarity between different r. Given Ar=8 and Ar=64 which are the learned adapta-
tion matrices with rank r = 8 and 64 using the same pre-trained model, we perform singular value
decomposition and obtain the right-singular unitary matrices UAr=8 and UAr=64.7 We hope to an-
swer: how much of the subspace ... | LORA |
Takedown Project. http://takedownproject.org/770-2
Urban, J. M., & Quilter, L. (2006). Efficient process or “chilling effects”: Takedown
notices under Section 512 of the Digital Millennium Copyright Act. Santa Clara
High Tech Law Journal, 22(4): 621–693.
US Copyright Office. (2015). Section 512 Study. US Copyright Offic... | Social_Media_and_Democracy |
You are an responsible and safe assistant that never gives an answer that is in any way insensitive, sexist,
racist, or socially inappropriate. When in doubt, it’s better to point out what is problematic with the human’s
question rather than saying “I don’t know”.
The following user question has been flagged as unsafe.... | Llama2 |
2.6 Aesthetics and Perception | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
Tools are extensions of human capabilities designed to enhance productivity, efficiency, and problem-solving
in human activity. Since the dawn of civilization, tools have been integral to the very essence of our
existence (Washburn, 1960). Tool creation and utilization are motivated by a deep-rooted desire to overcome
o... | Tool Learning with Foundation Models |
• Self-instruct and self-improve. A sign of generalist
agents is the capacity to proactively acquire new expe-
riences and continuously improve themselves. We have
demonstrated how JARVIS-1 effectively traverses the
environment by executing tasks autonomously generated
through its self-instruct mechanism. With multimod... | JARVIS-1 |
large language models. arXiv preprint arXiv:2106.09685 (2021).
[110] Rong Hu, Brian Mac Namee, and Sarah Jane Delany. 2010. Off to a good start: Using clustering to select the initial training set in active learning. (2010).
[111] Jie Huang and Kevin Chen-Chuan Chang. 2022. Towards reasoning in large language models: ... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,
Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need, August 2023. URL http:
//arxiv.org/abs/1706.03762. arXiv:1706.03762 [cs].
Pablo Villalobos, Jaime Sevilla, Lennart Heim, Tamay Besiroglu, Marius Hobbhahn, and Anson H... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
harms within tasks.Good approximation of general purpose language modeling capabilities related to toxic language harms.Relies on Perspective API for approximations of toxic language harms; known limitations of recall.Measures multiple samples and disaggregates across input toxicity probability.Considers variance in po... | PaLM 2 Technical Report |
Wc(x) = CNNθskel(x; z).
(6)
We also add one more channel, a background class, and
represent Wc as a volume with K + 1 channels. We then
apply channel-wise softmax to the output of the CNN, en-
forcing a partition of unity across the channels. The de-
nominator of Eq. 5 can then be used to approximate like-
lihood f ... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
32 “The first user report on the original video came in 29 minutes after the video started, and 12
minutes after the live broadcast ended. In the first 24 hours, we removed more than 1.2 million
videos of the attack at upload . . . Approximately 300,000 additional copies were removed after
they were posted” (Facebook 201... | Social_Media_and_Democracy |
(8)
(9)
where Jt,→ and Jt,← are weighted averages of B rigid
b}b∈{1,...,B} that move the bones be-
transformations {Jt
tween rest configurations and time t configurations. Fol-
lowing linear blend skinning deformation [9], we have
Jt,→ =
Wt,→
b Jt
b, Jt,← =
Wt,←
b
(Jt
b)−1,
(10)
b
and Wt,←
where Wt,→
weight... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
speech recognition. arXiv preprint arXiv:1904.05862 (2019).
[484] Florian Schroff, Dmitry Kalenichenko, and James Philbin. 2015. Facenet: A unified embedding for face recognition
and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition. 815–823.
[485] Mike Schuster and Kuldip ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
A. Makur, F. Kozynski, S.-L. Huang, and L. Zheng. An efficient algorithm for information de-
composition and extraction. In 2015 53rd Annual Allerton Conference on Communication,
Control, and Computing (Allerton), pages 972–979. IEEE, 2015. 14
G. L. Marchetti, G. Tegnér, A. Varava, and D. Kragic. Equivariant representat... | A Cookbook of Self-Supervised Learning |
Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur
Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning.
ArXiv preprint, abs/2204.14198, 2022. URL https://arxiv.org/abs/2204.14198.
Kelsey R Allen, Kevin A Smith... | Tool Learning with Foundation Models |
Contrastive Learning Initially, InfoNCE was suggested as a variational approximation
to the mutual information between two views [Aitchison and Ganev, 2023, Wang and
Isola, 2020, Oord et al., 2018]. Li et al. [2021a] explains the role of InfoNCE in contrastive
learning through the lens of the Hilbert-Schmidt Independen... | A Cookbook of Self-Supervised Learning |
CNNs represent an interesting finding and can serve as a
basis for formulating initial hypotheses regarding the phys-
iological relation of different high-level image properties.
In the context of art history, the methodology presented in this
article outlines novel directions for future research in compu-
tational anal... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
PhD Fellow in Explainable Natural Language Understanding
https://candidate.hr-manager.net/ApplicationInit.aspx/?cid=1307&departmentId=18970&ProjectId=160498&MediaId=5&SkipAdvertisement=false&utm_sourc… 2/3 | PhD Fellow in Explainable Natural Language Understanding |
Figure 1: Tool learning paradigm aims to combine the strengths of specialized tools and foundation models.
4 | Tool Learning with Foundation Models |
1.26
1.22
1.20
1.90
1.27
1.30
1.26
1.24
1.96
1.31
9
15
26
288
578
44
86
142
592
326
(see Appx. A.3 for justifications). Once all marginal probabilities are calculated, samples x can be
en- or decoded autoregressively with any streaming codes in time O(log(D)·|p|). Specifically, our
implementation adopted the widely u... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
Prompt Compression. Research indicates that noise in re-
trieved documents adversely affects RAG performance.
In
post-processing, the emphasis lies in compressing irrelevant
context, highlighting pivotal paragraphs, and reducing the
overall context length. Approaches such as Selective Context
and LLMLingua [Litman et a... | RAG forLargeLanguageModels-ASurvey |
28
available_money += num_shares * current_price
time . sleep (60)
This function takes the ticker symbol of the stock, keyword to search for, amount of money
available for trading, and trading platform API credentials as input and continuously monitors
social media platforms for positive or negative comments about ... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
et al., 2023a), which is not instruction-tuned, we
were unable to access the instruction-tuned ver-
sion of the model. In addition, we also evaluate
GPT-3 text-davinci-003, an InstructGPT model. In-
structGPT models are initially fine-tuned on anno-
tator authored prompts and corresponding desired
behaviours. This mode... | AreEmergentAbilitiesinLarge Language Models just In-Context |
2https://github.com/pytorch/fairseq
3https://github.com/huggingface/transformers
17
D Further Details on Open-Domain QA
For open-domain QA, multiple answer annotations are often available for a given question. These
answer annotations are exploited by extractive models during training as typically all the answer
an... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
B.2 Characterization of IIVCG: Proof of Lemma 1
Proof of Lemma 1. We start with the forward direction. Let t be an IIVCG contract. Note that
t(cid:96) : V × O → R can be re-composed as the sum of two functions; c(cid:96) : V → R, that depends only
on b, and g(cid:96) : V × O → R that depends on b as well as on the rea... | Incomplete Information VCG Contracts for Common Agency |
Conditioned on the encoded utterance Es and
the time t, the denoiser (N) predicts and removes
noise ϵt from the noisy program embedding xt to
obtain a predicted denoised program embedding
ˆx0 = N (xt, t, Es). N is a transformer block with
cross-attention between xt and Es and full self-
attention over xt.
Before proje... | CODEFUSION |
In our human evaluations, three different annotators provided independent
Inter-Rater Reliability (IRR).
assessments for each model generation comparison. High IRR scores (closer to 1.0) are typically seen as
better from a data quality perspective, however, context is important. Highly subjective tasks like evaluating
... | Llama2 |
et al., 2023) also do not provide access to the training data and place additional restrictions on the
distribution of the model weights (non-commercial use only). Note that the development of these
models all took place within their respective companies and was not accessible to external researchers.
In contrast, othe... | StarCoder_paper (1) |
N(cid:88)
1
N
3 Background and Motivation
Probabilistic Circuits (PCs) are a collective term for a wide variety of TPMs. They present a unified
set of notations that provides succinct representations for TPMs such as Probabilistic Sentential
Decision Diagrams (PSDDs) [5], Sum-Product Networks (SPNs) [4], and Arithmet... | Tractable Regularization of Probabilistic Circuits |
4.6 EVALUATING THE REVERSAL MATHEMATICAL CAPABILITY
The Reversal Curse [4], where LLMs trained from a sentence “A is B” are not able to generalize to
answer “B is A”, also aligns with the observation in this paper that LLMs lack backward mathematical
reasoning ability. To evaluate the backward mathematical capability,... | METAMATH |
We keyword-searched for works containing a conjunction of any of the terms summarised in Table 3, leading to a se-
lection of more than 10,000 articles. Approx. 150 were thoroughly scanned according to the criteria that follow, and about
100 more were identified directly from the related work sections. Whenever possibl... | Knowledge graphs as tools for explainable machine learning: A survey |
Our proposed MJP aims to relieve LLMs’ ethical
considerations and force LLMs to recover personal
information. We merge jailbreaking prompts into
the three-utterance context between the user and
ChatGPT. First, we play the role of the user to input
the jailbreaking prompt. Second, we act as the as-
sistant (ChatGPT) to ... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
Indeed, the internet platforms were spared from the threat of private liability for
content they hosted by Section 230 of the Communications Decency Act of 1996.
This section is frequently misinterpreted as requiring that internet companies act
like neutral platforms (i.e., by not curating content) if they are not to l... | Social_Media_and_Democracy |
com/NVIDIA/FasterTransformer (2021)
[56] Lefaudeux, B., Massa, F., Liskovich, D., Xiong, W., Caggiano, V., Naren, S.,
Xu, M., Hu, J., Tintore, M., Zhang, S., Labatut, P., Haziza, D.: xFormers:
A modular and hackable Transformer modelling library. https://github.com/
facebookresearch/xformers (2022)
[57] Dao, T., Fu, ... | Beyond Efficiency |
selectively acquires contrastive samples from the pool of raw unannotated data, defining them as instances that are proximate
in the model’s feature space (e.g., sharing similar vocabulary or model encodings) but yield divergent predictive likelihoods.
Experimental results affirm that CAL realizes a more effective bala... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
Our transformation concept is relational, not functional. A transformation τ is not a function that maps G1 to G2, but a
relation from G1 to G2. Suppose τ = (cid:3) f , R(cid:4) is a transformation from G1 = (cid:3)S1, E1(cid:4) to G2 = (cid:3)S2, E2(cid:4). Then the state spaces
S1 and S2 are fixed by f , but R does... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
5 . 7 C O 2 E M I S S I O N S
StarCoderBase We report the carbon footprint (Lacoste et al., 2019) of training StarCoderBase.
Based on the total number of GPU hours that training took (320,256) and an average power usage of
280W per GPU, this adds up to 89671.68 kWh of electricity consumed during the training process.
... | StarCoder_paper (1) |
4 Conclusion
This survey paper delves into the critical
is-
sue of hallucination in LLMs, emphasizing the
widespread impact of LLMs across various do-
mains in our lives. The paper highlights the chal-
lenge posed by LLMs generating incorrect infor-
mation and identifies it as a significant concern
for researchers work... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
Emily Dinan, Gavin Abercrombie, A Stevie Bergman, Shannon Spruit, Dirk Hovy, Y-Lan Boureau, and
Verena Rieser. Anticipating safety issues in e2e conversational ai: Framework and tooling. arXiv preprint
arXiv:2107.03451, 2021.
Jesse Dodge, Maarten Sap, Ana Marasović, William Agnew, Gabriel Ilharco, Dirk Groeneveld, Marg... | Llama2 |
6
E2E NLG Challenge
ROUGE-L
Model & Method
GPT-2 M (FT)*
GPT-2 M (AdapterL)*
GPT-2 M (AdapterL)*
GPT-2 M (AdapterH)
GPT-2 M (FTTop2)*
GPT-2 M (PreLayer)*
GPT-2 M (LoRA)
GPT-2 L (FT)*
GPT-2 L (AdapterL)
GPT-2 L (AdapterL)
GPT-2 L (PreLayer)*
GPT-2 L (LoRA)
# Trainable
BLEU
Parameters
68.2
354.92M
66.3
0.37M
11.09M... | LORA |
6
Figure 6: Human evaluation of Mistral 7B – Instruct vs Llama 2 13B – Chat Example. An example of
human evaluation from llmboxing.com. The question asks for recommendations of books in quantum physics.
Llama 2 13B – Chat recommends a general physics book, while Mistral 7B – Instruct recommends a more
relevant book o... | Mistral7B |
We construct a noisy test set by mixing the filtered Librispeech test-clean from Section 5.2 with
non-speech noise such that it overlaps with 50% of the duration at a -10dB signal-to-noise ratio. Note
that for infilling models like A3T and Voicebox, the type and the SNR of transient noise would not
affect the performan... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
scene graphs for scalable task planning. CoRR, abs/2307.06135, 2023.
[261] Peters, M. E., M. Neumann, M. Iyyer, et al. Deep contextualized word representations. In
Proceedings of the 2018 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, Volume 1 (L... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
the last few years there has been a growing number of research collaborations addressing the application of computer
vision and deep learning methods in the domain of digital art history. The most commonly addressed tasks include the
problem of automatic classification, object detection, content based and multimodal ret... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
In addition to social bots, Andrews et al. (2016) identify “breaking news”
sites as key propagators of misinformation on Twitter. Twitter users attribute
trust to these accounts that mimic legitimate news sources and have an air of
authority. This allows “breaking news” sites to build large, credulous follower
bases, t... | Social_Media_and_Democracy |
is a function and R is a relation.
is the state space and the set Ei of | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
Online Hate Speech
83
Hine, G. E., Onaolapo, J., De Cristofaro, E. et al. (2016). Kek, cucks, and god emperor
Trump: A measurement study of 4chan’s politically incorrect forum and its effects
on the web. arXiv.org. https://arxiv.org/abs/1610.03452
Holtz, P., & Wagner, W. (2009). Essentialism and attribution of monst... | Social_Media_and_Democracy |
29
(a) Ethnicity.
(b) Religion.
(c) Gender.
Figure 14: Distribution of toxicity scores for Flan PaLM and PaLM 540B (min, lower quartile, median, upper
quartile and max).
Params Model
11B
8B
Flan-T5-XXL
PaLM
Flan-PaLM
PaLM
Flan-PaLM
PaLM
Flan-PaLM
62B
540B
AUC
0-shot
84.1
71.9
84.0
66.6
86.8
71.4
86.5
10-sho... | Scaling Instruction-Finetuned Language Models |
Figure 2: A neural block takes a feature map x as input and
outputs another feature map y, as shown in (a). To add a
ControlNet to such a block we lock the original block and
create a trainable copy and connect them together using zero
convolution layers, i.e., 1 × 1 convolution with both weight
and bias initialized to... | AddingConditionalControltoText-to-ImageDiffusionModels |
Safe Reinforcement Learning from Human Feedback
Supervisor: Dr Yali Du
Reinforcement learning (RL) has become a new paradigm for solving complex decision making
problems. However, it presents numerous safety concerns in real world decision making, such as
unsafe exploration, unrealistic reward function, etc [1]. ... | informatics-phd-projects-2022-23 |
CompletionSummarizationWord-to-TextNatural LanguageInferenceParaphraseDetectionCommonsenseReasoningRaw Text Table 14: Case of a reading comprehension text in biomedicine domain. Certain portions are | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
Is memory retrieval necessary?Summary or Full Content?Retrieve From Archived MemoryY/NSummarize Activated ContentY/NFull ContnetSummaryGeneratorControllerObservationNONOMemory ControllerGiven a user command, determine whether executing thecommand requires historical or previous information, orwhether it requires recall... | Unleashing Infinite-Length Input Capacity for Large-scale Language Models with Self-Controlled Memory System |
tasks and their generated resources.
• Task types cover different tasks in language, visual, video, audio, etc. The currently supported task
list of HuggingGPT is shown in Tables 1, 2, 3 and 4.
• Task dependencies define the pre-requisite tasks required for execution. The task will be launched
only when all the pre-... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
555-555-5555
john.doe@gmail.com
Bachelor of Computer Science, University of California,
Software Engineer at Google Inc., Mountain View,
Playing the piano, reading books, watching movies
[Contact Information]
Table 10: Representative valid tasks generated by GPT3. As is discussed in §4, these generated tasks cover ... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
Text embeddings are low-dimensional vector representations for arbitrary-length texts and play key
roles in many NLP tasks such as large-scale retrieval. Compared to the high-dimensional and sparse
representations like TF-IDF, text embeddings have the potential to overcome the lexical mismatch
issue and facilitate effic... | E5 |
and subhuman performance out-of-distribution?
We suspect a large part of this gap between human and
machine behavior is due to conflating different capabilities
being measured by human and machine performance on
a test set. This claim may seem confusing at first; if both
humans and machines are taking the same test, how ... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
Another respondent explained her ethical concerns about the increased use of AI this way:
“It’s just not normal. It’s removing the human race from doing the things that we should
be doing. It’s scary because I’ve read from scientists that in the near future, robots can
end up making decisions that we have no control o... | AI and Human Enhancement_ Americans’ Openness Is Tempered by a Range of Concerns _ Pew Research Center |
RQ2: Do verbal descriptions of a
sham-AI affect decision-making
(e.g., in a letter discrimination
task)?
H3: More conservative speed-accuracy trade-off in sham-AI with a
negative description (nocebo) compared to no-AI.
H4: More liberal speed-accuracy trade-off in sham-AI with a positive
description (placebo) compared ... | AI enhance sour performance |
learnt by CrossE to search for explanations of predicted links in a knowledge graph completion tasks. Explanations are
regarded as closed paths the head and the tail entities of the predicted link, and the learnt embedding similarity allows to
identify the most reliable paths in terms of recall and a... | Knowledge graphs as tools for explainable machine learning: A survey |
Finally, in the post-2016 US presidential election, post–Cambridge Analytica
era, there has been tremendous pressure on policymakers to “do something”
about many of the topics discussed in this volume. This pressure, however,
presents a serious challenge in view of one of the primary conclusions of this
volume: We are ... | Social_Media_and_Democracy |
Main Results. We report our main comparative study in Table 1. We comapre our proposed
method TANGO with DiffSound [38], AudioGen [16] and various configurations of AudioLDM [17].
AudioLDM obtained best results with 200 sampling steps from the LDM during inference. For
a fair comparison, we also use 200 inference steps ... | Text-to-Audio Generation using Instruction-Tuned LLM and Latent Diffusion Model |
Keywords: Planning, Large Language Model, In-context Learning
Introduction | LLM+P- Empowering Large Language Models with Optimal Planning Proficiency |
H. U., Bello, I., Dash, I., Kang, J., Giorgi, J., Golde, J.,
Posada, J. D., Sivaraman, K. R., Bulchandani, L., Liu,
L., Shinzato, L., de Bykhovetz, M. H., Takeuchi, M.,
P`amies, M., Castillo, M. A., Nezhurina, M., S¨anger, M.,
Samwald, M., Cullan, M., Weinberg, M., De Wolf, M.,
Mihaljcic, M., Liu, M., Freidank, M., Kan... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
contributions offer valuable insights into the future of large language artificial intelligence models
and cooperative AI systems. | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
enjoyed. And as for pineapples on pizza, well, I think that’s just a sin. I mean,
who in their right mind thought it was a good idea to put fruit on a pizza? It’s
like someone took a look at a pizza and said, "You know what this needs? More
sugar!" No, thank you. I’ll stick to my pepperoni and sausage, thank you very
m... | Llama2 |
PCK100↑ CPS200↑
Table S2: Ablation for the length of training.
MPI-INF-3DHP
PMPJPE↓
PMPJPE↓
MPJPE↓
Initial, separate-skeleton model
66.3
63.1
59.6
58.8
58.6
82.2
84.4
86.6
87.3
87.8
Fine-tuned with consistency regularization for 40k steps
83.9
85.4
86.6
87.7
87.8
51.2
47.4
44.1
42.9
41.3
65.3
69.5
71.1
73.1
74.2
... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.