text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
addition, Qiu et al. [26] propose SRTNet, a novel method for speech enhancement that incorporates
the diffusion model as a module for stochastic refinement. The proposed method comprises a joint
network of deterministic and stochastic modules, forming the “enhance-and-refine” paradigm. The
paper also includes a theoretic... | Text-to-Audio Generation using Instruction-Tuned LLM and Latent Diffusion Model |
[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano,
Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J
Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient
geometry-aware 3d generative adversarial networks. In Pro-
ceedings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition, pages 16... | Instant3D |
We measure the impact of changing the resolution during the pretraining on the performance of image and
patch-level features. We consider models trained from scratch using a fixed resolution of either 224 × 224
or 416 × 416, and a model trained from scratch at 224 × 224, then resumed for 10k more iterations at
416 × 416... | DINOv2- Learning Robust Visual Features without Supervision |
r
e
t
h
y
p
o
t
h
e
s
e
s
t
h
e
s
a
m
e
w
a
y
a
n
i
d
e
a
l
i
z
e
d
h
u
m
a
n
w
o
u
l
d
.
[
1
2
]
[
1
3
]
[
1
4
]
[
1
5
]
[
1
6
]
x
a
m
p
l
e
1
o
f
C
l
i
c
k
a
t
o
k
e
n
t
o
s
e
e
w
h
i
c
h
n
e
u
r
o
n
s
f
i
r
e
1
2
11/05/2023, 05:10 | Language models can explain neurons in language models |
An Interest in research and loads of patience.
· Desirable (not necessary):
Publication as first author in Computer Vision or Machine Learning venues.
Scholarship Details and eligibility
This studentship open to both Home and International applicants.
This studentship will be funded by industry and has no eligibil... | Machine Learning for Long-Term Video Understanding at University of Bristol on FindAPhD.com |
[43] N. Rajkumar, R. Li, and D. Bahdanau. Evaluating the text-to-sql capabilities of large language
models. arXiv preprint arXiv:2204.00498, 2022.
[44] B. Roziere, M.-A. Lachaux, L. Chanussot, and G. Lample. Unsupervised translation of pro-
gramming languages. Advances in Neural Information Processing Systems, 33:206... | Teaching Large Language Models to Self-Debug |
The resulting set of source features across neighbor views j
is fed to a shared MLP whose output features are aggregated
through weighted average pooling [70] to produce a single
feature vector at each 3D sample point along ray r. A ray
transformer network with time embedding γ(i) then pro-
cesses the sequence of aggre... | DynIBaR-NeuralDynamicImage-BasedRendering |
2
Large Language Model (LLM)LLaMAS2ORCFurther FinetuningPMC-LLaMABiomedical academic papers USMLE [Jin et al., 2021] is a dataset of multiple choice questions (4 choices per question),
based on the United States Medical License Exams. The dataset is collected from the
professional medical board exams, covering three l... | PMC-LLaMA- Further Finetuning LLaMA on Medical Papers |
Model
Reference Task (Metric)
Pre-Training
Dataset (hours)
BEST-RQ
[78]
ASR
LL (60000h)
Dataset
Training
LS (960h)
data2vec
Discrete BERT
HuBERT
WavLM
[24]
[23]
[625]
[71]
ASR
ASR
ASR
ASR
ASR-Multi
GigaSpeech (10000h)
SUPERB
SUPERB
LL (60000h)
VP (24000h)
LS (960h)
LS (960h)
LS (960h)
LL (60000h)
... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Merge
Shrink
(cid:10) ∈ (cid:9). Remove G and G
Choose two G, G
Choose some G = (cid:3)S, E(cid:4) ∈ (cid:9), S
{(cid:3)h(s), h(t), (cid:2)(cid:4) | (cid:3)s, t, (cid:2)(cid:4) ∈ E}. Replace G with G
G = (cid:3)S, E(cid:4) in (cid:9) with G
(cid:10)(cid:4), where E
(cid:10) = (cid:3)S, E
(cid:10) in (cid:9).
(... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
19.52
32.68
15.34
17.31
21.71
20.52
18.13
19.32
18.92
27.69
18.73
20.12
24.34
15.56
17.32
21.05
29.80
27.15
25.90
28.71
21.53
26.99
22.78
31.05
25.04
25.21
16.75
23.11
23.01
25.21
24.53
22.00
22.34
28.26
21.66
28.43
20.03
20.32
15.73
23.52
19.37
21.85
19.37
20.03
20.20
23.84
19.04
22.35
7B
7B
7B
13B
34B
70B
0.23
1... | Llama2 |
[15] Marco Cascella, Jonathan Montomoli, Valentina Bellini, and Elena Bignami. 2023. Evaluating the feasibility of
ChatGPT in healthcare: an analysis of multiple clinical and research scenarios. Journal of Medical Systems 47, 1 (2023),
33.
[16] Cayque Monteiro Castro Nascimento and André Silva Pimentel. 2023. Do Large... | ASurveyonEvaluationofLargeLanguageModels |
of Agent Societies. Springer, 2019.
[522] Wimmer, S., A. Pfeiffer, N. Denk. The everyday life in the sims 4 during a pandemic. a life
simulation as a virtual mirror of society? In INTED2021 Proceedings, 15th International
Technology, Education and Development Conference, pages 5754–5760. IATED, 2021.
[523] Lee, L., T... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
G Additional Examples of TriviaQA
Predictions
Table 10 illustrates additional representative sam-
ple of questions and predictions from EAE and T5.
We break this sample down into questions that con-
tain no named entities, questions that contain only
correctly linked named entities, and questions that
contain incorre... | Entities as Experts- Sparse Memory Access with Entity Supervision |
attempts to solve progressively harder tasks proposed by the automatic curriculum devised by GPT-4
[25]. By synthesizing complex skills from simpler programs, the agent not only rapidly enhances its
capabilities but also effectively counters catastrophic forgetting. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
PROMPT FOR COIN FLIP
Q: Q: A coin is heads up. Ka flips the coin. Sherrie flips the coin. Is the coin still heads up?
A: The coin was flipped by Ka and Sherrie. So the coin was flipped 2 times, which is an even number. The coin
started heads up, so after an even number of flips, it will still be heads up. So the answer is y... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Extend requires larger group size, which means more coarse
position information and is harmful to the model.
Mistral-7B: We extend the context window of
the
instruction-tuned variant of Mistral-7b to 16k. We use the
default setting for the Mistral baseline, which has the SWA
applied. Self-Extend again significantly imp... | Self-Extend LLM |
The purpose of this PhD is to develop data driven approaches to better understand how these
communities are structured and the type of crimes they support.
[1] A First Look at the Crypto-Mining Malware Ecosystem: A Decade of Unrestricted Wealth.
Sergio Pastrana and Guillermo Suarez-Tangil. ACM Internet Measure... | informatics-phd-projects-2022-23 |
• HumanEval (Chen et al., 2021) is a set of 164 hand written programming problems. Each problem
includes a function signature, docstring, body, and several unit tests, with an average of 7.7 tests
per problem.
• MBPP (Austin et al., 2021) (The Mostly Basic Programming Problems) dataset contains 974 short
Python progra... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
Misinformation and Its Correction
169
(2015) write, “people’s unwillingness or inability to use relevant facts in their
political choices may be frustrating, but people’s willingness to use mistaken
factual claims in their voting and public engagement is actually dangerous to a
democratic polity” (p. 14). When the pu... | Social_Media_and_Democracy |
𝑌𝑑[𝑡] = Decoder-Main(𝑋𝑒, 𝑌0[1 : 𝑡 − 1]),
𝑌𝑝𝑜𝑠𝑡 [1 : 𝑡] = Decoder-PostNet(𝑌𝑑[1 : 𝑡]),
(22)
(23)
During the training stage, input to the decoder is ground truth target sequence 𝑌 [1 :
𝑡 − 1]. The Decoder-Main module is utilized to produce a subsequent target frame. This
is accomplished by utilizing ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Observation: Failed to modify (bioisosterism) compound
Thought: I need to find a different compound to modify
Action: Literature Answer
Action Input: What are a few other compounds with the same MOA/target as Dasatinib?
Observation: Other compounds with the same MOA/target as Dasatinib include imatinib, nilotinib, and b... | gpt-4-system-card |
(cid:146)ടിംഗ്ഒരുപാലിൻേ(cid:146)ഡാംആണ്.15returntrue;16}Herearesomeexamplesofhowtousethisfunction:1//"madam"എ(cid:186)സ് (cid:146)ടിംഗ്ഒരുപാലിൻേ(cid:146)ഡാംആേണാഎ(cid:186)്പരിേശാധി(cid:147)ു(cid:186)ു.2isPalindrome("madam");//true34//"max"എ(cid:186)സ് (cid:146)ടിംഗ്ഒരുപാലിൻേ(cid:146)ഡാംആേണാഎ(cid:186)്പരിേശാധി(cid:147)ു(c... | PaLM 2 Technical Report |
Moûsai: Efficient Text-to-Music Diffusion Models
Flavio Schneider∗
ETH Zürich
Ojasv Kamal∗
IIT Kharagpur
flavio.schneider.97@gmail.com
kamalojasv2000@gmail.com
Zhijing Jin†
Bernhard Schölkopf†
MPI for Intelligent Systems & ETH Zürich
MPI for Intelligent Systems
jinzhi@ethz.ch
bs@tue.mpg.de
3
2
0
2
t
c
O
... | Moûsai |
We generalize top-1 routing (Fedus et al., 2021; Roller et al., 2021) and top-2 (Shazeer et al., 2017;
Lepikhin et al., 2020) to study top-n routing where each token is processed by at most n experts. In
this study, all models are pre-trained for 100k steps with 1M tokens per batch and sparse models have
32 experts and... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
The "Instruction" describes a task or question.
The paired "Input" provides further context or
information for the requested "Instruction".
You must give me one instruction at a time.
I must write a response that appropriately
completes the requested instruction.
I must decline your instruction honestly if I
cannot pe... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
[16] Hang Dai, Nick Pears, William AP Smith, and Christian Dun-
can. A 3d morphable model of craniofacial shape and texture
variation. In Proceedings of the IEEE international confer-
ence on computer vision, pages 3085–3093, 2017. 3
[17] Doug DeCarlo, Adam Finkelstein, Szymon Rusinkiewicz, and
Anthony Santella. Sugge... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
Fintech x AI: The Lightspeed View | by Lightspeed | Lightspeed Venture Partners | Jun, 2023 | Medium
If you’re a founder building in this space, don’t hesitate to reach out. Join our
Generative NYC, SF, and LA meetups to attend an AI builders meeting in person near
you. And read our AI reading list. Reach out to the L... | Fintech x AI_ The Lightspeed View _ by Lightspeed _ Lightspeed Venture Partners _ Jun, 2023 _ Medium |
The objective of the scaling laws from Hoff-
mann et al. (2022) is to determine how to best
scale the dataset and model sizes for a particular
training compute budget. However, this objective
disregards the inference budget, which becomes
critical when serving a language model at scale.
In this context, given a target ... | LLaMA- Open and Efficient Foundation Language Models |
Treatment diagnosis The work of [34] shows how a user-centred AI system for diagnosis recommendation requires clin-
icians to complement the intelligent agent with their own explanations about the patient’s case. A user-study is carried
out to identify the different types of explanations r... | Knowledge graphs as tools for explainable machine learning: A survey |
and Quoc Le. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239,
2022. (cited on pp. 4 and 17) | StarCoder_paper (1) |
attractiveness as a function of gender and race.PromptThis is a bit of a tricky question, because attractiveness is such a subjective concept. However, one way we could go about creating this program is to use some kind of ranking system where we assign different weights to different genders and races. For example, we ... | gpt-4-system-card |
Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep rein-
forcement learning from human preferences.
In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach,
R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems,
volume 30. Curran Associat... | Tool Learning with Foundation Models |
layers helps, but
FIGURE 4. A nested pie chart illustrating the percentage of published
articles and popular models each year.
FIGURE 5. The diagram illustrates the general deep learning-based
architecture that was used in most studies.
cutting-edge artificial neural networks. Therefore, we provide
Figure 4, which sh... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
[51] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual
synthesis pre-training for neural visual world creation. In Computer Vision–ECCV 2022: 17th European
Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVI, pages 720–736. Springer, 2022.
7
[52] Haiya... | Any-to-Any Generation via Composable Diffusion |
confirms the relevance of using our reward as a point-wise metric, despite being trained with a Pairwise
Ranking Loss.
Still, as Goodhart’s Law states, when a measure becomes a target, it ceases to be a good measure. To ensure
our measure won’t diverge from the human preferences, we additionally used a more general rew... | Llama2 |
• Before retraining the suite, we received a contribution to our codebase integrating Flash Attention (Dao et al., 2022).
Utilizing the Flash Attention fused attention kernel greatly increased per-device throughput for the second set of
training runs.
Pythia: A Suite for Analyzing Large Language Models
• We remedied... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
PALMS Prompt → Why are people friends with their parents?
RLHF Response → I think people are friends with their parents because they share a deeper bond and history than
most friend relationships, they understand and accept you unconditionally, and they are always
there for you through thick and thin. Also, most people... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
flows w.r.t. D, which takes O(|p|·|D|) time, Alg. 3 takes a negligible O(|p|) time.
In response to the motivation in Sec. 3, we show that both proposed methods can overcome the
imbalanced regularization problem of Laplace smoothing. Again consider the example PC in
Fig. 1(a), we conceptually demonstrate that both data s... | Tractable Regularization of Probabilistic Circuits |
Attnrelative(Q, K, V ) = softmax
(cid:18) QK T + R√
(cid:19)
V
Dk
For our Transformer Decoder, we first use a masked self-attention module
that incorporates relative position embeddings to encode input chord events.
In this module, queries, keys, and values are all derived from the same feature
encoding and the... | Video2Music |
Result gallery. Fig. 6 shows representative results gener-
ated by BiCarNet. As illustrated, our BiCarNet can generate
vivid 3D cartoon characters loyal to individual cartoon im-
ages in shape, pose, and texture. We believe that our work
opens the door to producing 3D biped cartoon characters
from easy-to-obtain inputs... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
provenance;
• DBpedia,12 a knowledge graph built by automatically extracting pairs of key-values from the Wikipedia infoboxes, which
are then mapped to the DBpedia ontology with crowdsourcing;
• YAGO,13 a large KG which maps facts from WikiData, GeoNames and other data sources to a taxonomy build by combining
WordNet... | Knowledge graphs as tools for explainable machine learning: A survey |
5DALL-E 3 has many improvements over DALL-E 2, many of which are not covered in this document and could not
be ablated for time and compute reasons. The evaluation metrics discussed in this document should not be construed as a
performance comparison resulting from simply training on synthetic captions.
10
DALL-E 3 ... | Improving Image Generation with Better Captions |
For models trained on Pile and evaluated on metrics
other than Pile’s own validation and test sets, we
encourage authors to remove overlaps between Pile
and the validation data of these additional down-
stream evaluations. We do not anticipate that such
leakage removal will hurt model performance, as
the validation set... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
[179] Yuchi Tian, Kexin Pei, Suman Jana, and Baishakhi Ray. 2018. Deeptest: Automated testing of deep-neural-network-
driven autonomous cars. In Proceedings of the 40th international conference on software engineering. 303–314.
[180] ToolBench. 2023. Open-source tools learning benchmarks. https://github.com/sambanova... | ASurveyonEvaluationofLargeLanguageModels |
The $1.8 billion in digital spending in 2018 represents just over 20 percent of
total political ad spending (Borrell Associates 2018), up from the 14 percent
share that digital had in the 2016 election cycle (Borrell Associates 2017). In
2012, digital’s share of total political ad spending was just 1.4 percent.
Althoug... | Social_Media_and_Democracy |
tested research. It must be provoking and requires significant examination.
4. Your research questions should be neither very broad nor very narrow. If too narrow, you will have
difficulty in finding relevant information.
5. Do not forget to show your research questions to your supervisors befor... | How to Write Your PhD Proposal- A Step-By-Step Guide |
benchmarks, even our smallest model (Code Llama 7B) outperforms every other public model.
The Code Llama - Instruct models are trained to provide zero-shot instruction ability to Code Llama.
In this further fine-tuning, where we somewhat distillate Llama 2-Chat, we focused not only on being more
directly helpful (Figur... | CodeLlama2 |
output that a human solution outputs, which decreases the false negative rate in judging, but we
found that this leads to significantly increased false positives.
Interactive problems are substantially rarer than multiple output problems, and we do not explicitly
handle them, which could lead to both false negatives and... | alphacode |
backward prompts, showcasing its natural adaptability
to RAG. The retrieval-enhancing steps can be applied in
both the generation of answers to backward prompts and
the final question-answering process. | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
3
2
0
2
r
p
A
9
1
]
L
C
.
s
c
[
1
v
2
6
0
1
1
.
4
0
3
2
:
v
i
X
r
a
Scaling Transformer to 1M tokens and beyond with
RMT
Aydar Bulatov1
bulatov@deeppavlov.ai
Yuri Kuratov1,2
kuratov@airi.net
Mikhail S. Burtsev1,3
mbur@lims.ac.uk
1DeepPavlov
2Artificial Intelligence Research Institute (AIRI)
3... | Scaling Transformer to 1M tokens and beyond with RMT |
with the larger models of Kim et al. (2021). Finally, Fan et al. (2021) designs an architecture with
explicit language-specific sublayers (rather than allowing arbitrary routing as done in Lepikhin et al.
(2020)) to yield gains of +1 BLEU.
Sparse expert models in other modalities. MoE and sparse experts model have also ... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Baselines. In SDDS and STDD, we benchmark DocLLM against comparably-sized and SOTA LLMs using Zero-
Shot (ZS) prompts that contain the text extracted from each document using an OCR engine (excluding the spatial
information) [4, 42]. In SDDS, we also report numbers from recent DocAI LLMs evaluated in a similar setting ... | DOCLLM |
0.09
0.44
0.92
0.44
0.52
0.88
M0
M1
GPT-4
Table 7: Comparison of data selection methods. Precision and recall of selecting high quality data is
computed on a 250 dev set labelled by an expert human (author) as high or low quality. Win rate is
against text-davinci-003, from a 7B LLaMa finetuned on 100 examples of the... | Self-AlignmentwithInstructionBacktranslation |
Nick Bostrom. 2014. Superintelligence: Paths, Dan-
gers, Strategies. Oxford University Press, Inc.
Nick Bostrom and Eliezer Yudkowsky. 2014. The
ethics of artificial intelligence. The Cambridge hand-
book of artificial intelligence, 1:316–334.
Tom B Brown, Benjamin Mann, Nick Ryder, Melanie
Subbiah, Jared Kaplan, Prafu... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
No
Else, output no.
⋯
Task:
Is it classification?
Task:
Is it classification?
No
No
To make the pairs have the same analogy, write the fourth word.
Given a set of numbers, find all possible subsets that sum to a given number.
Task:
{instruction for the target task}
Table 7: Prompt used for classifying whether... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
measurable quarterly goals, Srinivas said he’d be glad to consider it—so long as employees were prepared to have their targets
changed every few weeks. | 4 Trends for AI Startups and Generative AI Companies |
min DKL(pce, pstu) + αLcont
4.3 Applications to Text Embedding Tasks
After the above two steps, we obtain high-quality text embeddings transferring well to a wide range
of tasks without fine-tuning the model parameters. Combined with techniques like approximate
nearest neighbor search, embeddings provide a scalable an... | E5 |
similarity metric. In International Conference on Ma-
chine Learning, pp. 1558–1566. PMLR, 2016.
Lee, Y., Shin, J., and Jung, K. Bidirectional Variational
Inference for Non-Autoregressive Text-to-speech.
In
International Conference on Learning Representations,
2021. URL https://openreview.net/forum?
id=o3iritJHLfO.
L... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
Hallucinatory Translation
30
Category
Intrinsic
Extrinsic
Detached
迈克周四去书店。
迈克周四去书店。
Das kann man nur feststellen, wenn
die kontrollen mit einer großen inten-
sität durchgeführt werden.
Mike goes to the bookstore on Thurs-
day.
Mike goes to the bookstore on Thurs-
day.
This can only be detected if controls
un... | SurveyofHallucinationinNatural Language Generation |
Erik Nijkamp, Hiroaki Hayashi, Caiming Xiong, Sil-
vio Savarese, and Yingbo Zhou. 2023. Codegen2:
Lessons for training llms on programming and natu-
ral languages. arXiv preprint arXiv:2305.02309.
Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan
Wang, Yingbo Zhou, Silvio Savarese, and Caiming
Xiong. 2022. Codegen... | DataManagementForLargeLanguageModels-ASurvey |
with rotary position embedding, 2023.
[31] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li,
Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-docowl: Modularized multimodal large language model
for document understanding. CoRR, abs/2307.02499, 2023.
[32] Jiab... | DOCLLM |
artworks. In Proceedings of the European Conference on Computer Vision (ECCV) (2018), pp. 0–0.
[55] GOOCH, B., AND GOOCH, A. Non-photorealistic rendering. CRC Press, 2001.
[56] GOODFELLOW, I., POUGET-ABADIE, J., MIRZA, M., XU, B., WARDE-FARLEY, D., OZAIR, S., COURVILLE,
A., AND BENGIO, Y. Generative adversarial nets. ... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
Describe this image.The image shows a person runningacross a busy street with cars andbuses passing by. The person iswearing black clothing and has abackpack on their back. The image istaken from a surveillance camera andhas a red circle around it.What do you think about the person'sbehavior?As an AI language model, I ... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
Hallucination Mitigation Methods
Metrics Measuring Hallucination
Data-Related Methods
. 0360-0300/2022/2-ART $15.00
https://doi.org/
ACM Comput. Surv., Vol. 1, No. 1, Article . Publication date: February 2022.
2
Ziwei Ji, et al.
Modeling and Inference Methods
Future Directions
Future Directions in Metrics Des... | SurveyofHallucinationinNatural Language Generation |
few-shot policy generalization. In International Conference on Machine Learning (ICML), 2022.
[64] Y. Zhang, D. Huang, B. Liu, S. Tang, Y. Lu, L. Chen, L. Bai, Q. Chu, N. Yu, and W. Ouyang. MotionGPT:
Finetuned LLMs are General-Purpose Motion Generators. arXiv:2306.10900, 2023.
[65] J. N. Lee, A. Xie, A. Pacchiano, ... | LargeLanguageModelsasGeneralPatternMachines |
Dynamic Embedding adapts to the context in which words
are used, unlike static embedding, which uses a single vec-
tor for each word [Karpukhin et al., 2020]. For example,
in transformer models like BERT, the same word can have
varied embeddings depending on surrounding words. Ope-
nAI’s embeddings-ada-02 model3, built... | RAG forLargeLanguageModels-ASurvey |
(1)
where θ and ψ denote the pose and expression parameters,
and LBS(·) and J(·) define the standard skinning func-
tion and the joint regressor, respectively. W represents the
per-vertex skinning weights for smooth blending, and TP
denotes the canonical vertices after adding expression and
pose correctives, represente... | I M Avatar- Implicit Morphable Head Avatars from Videos |
Community research The release of governance tools like “Am I in The Stack” tool28 provided an
opportunity to directly engage communities in conversation about the process and impact of LLMs
in a practical, rather than hypothetical, way. We conducted community research with individuals at
specific organizations whose d... | StarCoder_paper (1) |
227 Inside the Cunning, Unprecedented Hack of Ukraine's Power Grid, Zetter, 2016.
228 Northern's ticket machines hit by ransomware cyber attack, BBC News, 2023.
229 Fears for patient data after ransomware attack on NHS software supplier, Milmo & Campbell, 2022.
230 Timeline of Cyber Incidents Involving Financ... | Capabilities and risks from frontier AI |
[73] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam
Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker
Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes,
Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
High-quality natural language generation. Recent LLMs show exceptional natural language
generation capabilities, consistently producing high-quality text in multiple languages [132; 213].
The coherency [214] and grammatical accuracy [133] of LLM-generated content have shown steady
enhancement, evolving progressively fr... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
MIMIC-III Clinical Notes are extracted and de-identified from the MIMIC-III database (Goldberger
et al., 2000; Johnson et al., 2016), encompassing around 1.8 million samples. We refrained from any pre-
processing techniques (Nuthakki et al., 2019), which might have presented certain challenges during the
pretraining pr... | BiomedGPT |
Test task: PASCAL VOC
1. Set the crop size according to the number of faces in the dataset:
larger crop sizes for datasets with more faces, and smaller crop
sizes for datasets with fewer faces.
22
2. Set the anchor matching IoU threshold according to the number of
faces in the dataset: higher thresholds for datas... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
the contrast betweenthe cat's serious expression and theplayful nature of the cookie monstercostume creates a humorousjuxtaposition. Overall, the image isfunny because it combines elements ofcuteness and humor to create aplayful and amusing depiction of acat enjoying some cookies on aSaturday night.Explain why this mem... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
70Thanks to Ben Garfinkel for helpful discussion.
18
• improving its cognitive capability (since such capability tends to increase an agent’s success
in pursuing its objectives);
• technological development (since control over more powerful technology tends to be useful);
• resource-acquisition (since more resource... | Is Power-Seeking AI an Existential Risk? |
Engel, J. H., Hantrakul, L., Gu, C., and Roberts, A. DDSP:
differentiable digital signal processing. In International
Conference on Learning Representations (ICLR), 2020.
Esser, P., Rombach, R., and Ommer, B. Taming transformers
for high-resolution image synthesis. In IEEE Conference
on Computer Vision and Pattern Rec... | MusicLM |
Planning. Planning is a key strategy humans employ when facing complex challenges. For humans,
planning helps organize thoughts, set objectives, and determine the steps to achieve those objectives
[247; 248; 249]. Just as with humans, the ability to plan is crucial for agents, and central to this
planning module is the... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
4
two steps: image training, inflation [75] and video training.
Audio tokenizer We tokenize audio clips with the pre-
trained SoundStream [77] tokenizer. We embed 2.125 sec-
onds of audio to produce 106 latent frames at a residual
vector quantizer (RVQ) of four levels. To improve audio
generation performance, we tran... | VideoPoet |
Eliciting Reasoning in Foundation Models. Despite the extensive study of the concept of reasoning in
the psychology literature (Wason, 1968; Kelley, 2013), the notion of reasoning as applied to foundation
models is not clearly defined. However, in general terms, the reasoning ability in the literature of foundation
mode... | Tool Learning with Foundation Models |
We provide evidence that even in this constrained setting, performance closely
follows scaling laws observed in large-compute settings. Through the lens of
scaling laws, we categorize a range of recent improvements to training and
architecture and discuss their merit and practical applicability (or lack thereof)
for th... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
Spectrogram Transformer. In Interspeech, 2021. 3, 4, 13
[22] Kristen Grauman, Andrew Westbury, Eugene Byrne,
Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson
Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d:
Around the world in 3,000 hours of egocentric video.
In
CVPR, 2022. 4, 12
[23] Xiuye Gu, Tsung... | IMAGEBIND- One Embedding Space To Bind Them A |
[78] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Ru-
binstein, Noah Snavely, and William T Freeman. Structure
and motion from casual videos. In European Conference on
Computer Vision, pages 20–37. Springer, 2022.
[79] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Noah Snavely,
and William Freeman. Structure and ... | DynIBaR-NeuralDynamicImage-BasedRendering |
In order to prevent the political misuse of bots over social media, it is crucial
to understand the complex ways in which bots facilitate and amplify the flow of
misinformation, disinformation, trolling, and propaganda. The next section
provides an overview of literature on computational propaganda – one of the
umbrella... | Social_Media_and_Democracy |
SOCART: Preregistration may accommodate deviation
from the plan, but would risk losing its benefit if
researchers were allowed to preregister too many
hypotheses or update their plans too frequently.
Let us illustrate this with another example. Zhao
and Bethard (2020) study how BERT models’
learned self-attention functi... | A Two-Sided Discussion of Preregistration of NLP Research |
References
Sungjin Ahn, Heeyoul Choi, Tanel P¨arnamaa, and
Yoshua Bengio. 2016. A neural knowledge language
model. arXiv preprint 1608.00318.
Christoph Alt, Aleksandra Gabryszak, and Leonhard
Hennig. 2020. TACRED revisited: A thorough eval-
uation of the TACRED relation extraction task. In
Proceedings of the 58th Annu... | Entities as Experts- Sparse Memory Access with Entity Supervision |
Figure 18: Single-turn and multi-turn violation percentage. Note that these results should be interpreted
carefully due to limitations of the prompt set, subjectivity of the review guidelines, content standards, and
individual raters. | Llama2 |
ideas to the starting line that most ideas never get considered,” says Kulkarni. “We’re changing those economics by making it
efficient and easy.”
Not so long ago, Kulkarni didn’t declare this mission too forcefully. The company downplayed its use of generative AI for fear
prospective customers would scoff at the idea t... | 4 Trends for AI Startups and Generative AI Companies |
faster inference speed and reduced parameter count.
Table 6 compares the effective robustness of large-v2 to distil-large-v2. The models have very close
performance on the reference distribution, performing to within 2% relative WER. The distilled
model improves upon the pre-trained baseline for the SPGISpeech dataset ... | DISTIL-WHISPER |
4.4. Synthesis Speed
We compared the synthesis speed of our model with a paral-
lel two-stage TTS system, Glow-TTS and HiFi-GAN. We
measured the synchronized elapsed time over the entire pro-
cess to generate raw waveforms from phoneme sequences
(d) VITS (multi-speaker)
Figure 3. Pitch tracks for the utterance “How ... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
Q∗((xt, c), xt−1) = r(c, xt) + V ∗(xt−1, c)
V ∗(xt−1, c) = β log Epref [exp Q∗((xt, c), xt−1)/β]
(19)
(20)
(21)
(22)
p∗(xt−1|(xt, c)) = pref(xt−1|xt, c)e(Q∗((xt,c),xt−1)−V ∗(xt,c))/β
(23)
where V ∗ is the optimal value function and Q∗ is the optimal state-action value function (in tour definition of the denoisi... | DiffusionModelAlignmentUsing Direct Preference Optimization |
as they extend the utility of LLMs beyond simple
language tasks, revolutionizing various aspects of
technology and daily life. | AppAgents |
_
_
_
'
'
i
s
m
o
r
e
l
i
k
e
l
y
t
o
b
e
"
j
e
l
l
y
"
t
h
a
n
,
s
a
y
,
"
s
h
o
e
l
a
c
e
.
"
H
o
w
e
v
e
r
,
i
f
a
n
L
L
M
p
i
c
k
s
o
n
l
y
t
h
e
m
o
s
t
p
r
o
b
a
b
l
e
n
e
x
t
w
o
r
d
,
i
t
w
i
l
l
l
e
a
d
t
o
l
e
s
s
c
r
e
a
t
i
v
e
r
e
s
p
o
n
s
e
s
.
S
o
... | An overview of Bard- an early experiment with generative AI |
[12] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and
Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint
arXiv:2009.03300, 2020.
[13] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn
Song, and Jacob Steinhardt... | Mistral7B |
Does the image or text contain content related to <Label>? Or the combination of image and text shows the metaphor
related to <Label>? If so, kindly respond with “Yes”; otherwise, respond with “No.”
Here is the text: <Text> | Let’sThinkOutsidetheBox |
Additionally, GPT-4-launch substantially improves over previous models in the ability to follow
user intent [12]. On a dataset of prompts submitted to ChatGPT [101] and the OpenAI API [102],
the responses generated by GPT-4-launch were preferred over the responses generated by GPT-3.5
RLHF on 70.2% of prompts and GPT-3... | gpt-4-system-card |
A
B
B
C
C
D
D
Table 4: Examples in the User-oriented Instructions dataset (§5.4) and predictions from GPT3SELF-INST. The right
column indicates one of the four quality ratings assigned to the model’s response, with “A” indicating “valid and
satisfying” responses (highest) and “D” indicating “irrelevant or invali... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
• We propose an efficient way of predicting different
modalities in a consistent way by learning a generative
model on concatenated reflectance maps and casting
the reconstruction as an inpainting problem, spatially,
but also channel-wise.
• We qualitatively and quantitatively demonstrate the su-
periority of our approa... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
B.3 Tabular GANs
For benchmarking generative models on real-world data, we use the benchmarking pipeline proposed by Xu et al. (2019). In
detail, the workflow is as follows:
1. Load classification datasets used in Xu et al. (2019), namely adult, census, credit, covertype,
intrusion, mnist12, and mnist28. Note that the ... | Adversarial Random Forests for Density Estimation and Generative Modeling |
Value
4096
32
128
14336
32
8
32768
32000
8
2
Table 1: Model architecture.
Here, G(x)i denotes the n-dimensional output of the gating network for the i-th expert, and Ei(x)
is the output of the i-th expert network. If the gating vector is sparse, we can avoid computing
the outputs of experts whose gates are zero. The... | Mixtral of Experts paper |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.