text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
8
Figure 3: Fine-tuning evaluation on domain-specific tasks (left) and prompting evaluation on
general tasks (right). General LLM is the general language model, Raw Text trains the general
model on the domain-specific raw corpora, and Read. Compre.
trains the general model on the
reading comprehension texts construct... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
3
(cid:32)
T − p(cid:88)
i=1
(cid:33)2
di
,
Ldur = | Translatotron3 |
—
—
—
—
—
—
—
—
—
65.9/64.4
75.6/67.3
85.4/79.0
26.8/27.9
17.3/21.9
—
79.7/85.4
93.8/93.6
87.3/78.6
57.9/51.1
42.2/41.2
69.6/68.3
Test
—
—
—
—
—
37.5/33.8
19.5/18.2
22.5/23.6
33.9/32.6
76.2/74.9
0.65/0.00c
83.8/74.3
65.8/86.5
88.9/87.1
77.3/72.0
80.2/76.7
99.0/100.0
—
—
—
—
49.9/ 53.6
75.5/73.6
39.5/35.9
31... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
sushi disappear one by one
prompt: Rotating around a vase holding a dozen roses
Figure 14. A comparison of a 1B (left) and 8B (right) parameter models on the same prompt and settings.
19
starting frame
video name
elephant
elephant
car-turn
car-turn
dog-agility
dog-agility
bmx-bumps
bmx-bumps
train
train
bus
bus
l... | VideoPoet |
Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. QASC: A dataset for
question answering via sentence composition. In AAAI, 2020. URL https://arxiv.org/abs/1910.11473.
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language
models are zero-shot rea... | Scaling Instruction-Finetuned Language Models |
2.1.1 Scaling Laws
Before the popularization of LLMs, the relation-
ship between training dataset size and the perfor-
mance of language models with Transformer ar-
chitecture (Vaswani et al., 2017) had already at-
tracted researchers’ attention. Kaplan et al. (2020)
study the empirical scaling laws for Transformer
lan... | DataManagementForLargeLanguageModels-ASurvey |
*Equal contribution and corresponding authors: {lj,akos}@explosion.ai
1https://spacy.io
2https://thinc.ai
1
Figure 1: Illustration of the FGREP architecture. Redrawn from Miikkulainen and Dyer (1991)
2 BACKGROUND | MULTI HASH EMBEDDINGS IN SPACY |
As schematically shown in Figure 5b, we first trained a prompt, p1, for prompt tuning the LM on the
question answering task, and then used it to sample candidate answers from the model. Then, we
trained another prompt, p2, for prompt tuning the LM again, this time on the task of producing an | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
Smith TW, Davern M, Freese J, Stephen LM (2019) General Social
Surveys, 1972–2018: Cumulative Codebook / Principal Investiga-
tor, Tom W. Smith; Co-Principal Investigators, Michael Davern,
Jeremy Freese and Stephen L. Morgan. -- Chicago: NORC, 2019.
3,758 pp., 28cm. -- (National Data Program for the Social Sci-
ence... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
preprint arXiv:2012.15723 (2020).
[50] Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. RealToxicityPrompts:
Evaluating Neural Toxic Degeneration in Language Models. In Findings of the Association for Computational Linguistics:
EMNLP 2020. 3356–3369.
[51] Zorik Gekhman, Jonathan Herz... | ASurveyonEvaluationofLargeLanguageModels |
Orlando, USA, September 2022. Association for Machine
Translation in the Americas.
[Arora et al., 2023] Daman Arora, Anush Kini, Sayak Ray
Chowdhury, Nagarajan Natarajan, Gaurav Sinha, and
Amit Sharma. Gar-meets-rag paradigm for zero-shot infor-
mation retrieval. arXiv preprint arXiv:2310.20158, 2023.
[Asai et al., 20... | RAG forLargeLanguageModels-ASurvey |
In this use case, the hybrid team aims at monitoring animal wildlife in a given terri-
tory for an extended, multi-year period. A visual representation is given in Figure 4. The
AI has access to on-site sensors (e.g., camera traps, microphones from static arrays [32]
and aboard autonomous vehicles). The human in the te... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
3.5 Authorship, Copyright and Ethical Issues | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
Weight Type
Rank r
WikiSQL (±0.5%)
MultiNLI (±0.1%)
# of Trainable Parameters = 18M
Wq Wk Wv Wo Wq, Wk Wq, Wv Wq, Wk, Wv, Wo
8
8
8
8
4
70.4
91.0
70.0
90.8
73.0
91.0
73.2
91.3
71.4
91.3
4
73.7
91.3
2
73.7
91.7
Table 5: Validation accuracy on WikiSQL and MultiNLI after applying LoRA to different types of
at... | LORA |
Such behavior may contribute to the overall visibility of hate speech on
mainstream online platforms. For example, on Twitter, although tweets
containing hate speech have lower numbers of replies and likes than non-
hateful tweets, they contain a similar number of retweets (Klubicka and
Fernandez 2018). The highly netw... | Social_Media_and_Democracy |
asymmetric polarization | Social_Media_and_Democracy |
Martin Fajcik, Martin Docekal, Karel Ondrej, and Pavel Smrz. R2-D2: A modular baseline for
open-domain question answering. In Findings of the Association for Computational Linguistics:
EMNLP 2021, pp. 854–870, Punta Cana, Dominican Republic, November 2021. Association
for Computational Linguistics. doi: 10.18653/v1/202... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
11
AllNon MoEMoEAttentionFFNParameters Being Updated81828384858687SuperGLUE Score parameters that worked well for the dense model can mask any pre-training improvements obtained
by the sparse model.
Figure 6: Batch size and learning rate sensitivity. We measure differences and sensitivity to fine-
tuning protocols bet... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
log P (NA, NB|EA, EB) = −NA log(cid:0)1 + erB−rA(cid:1) − NB log(cid:0)1 + erA−rB(cid:1)
(B.1)
where rA,B = (log 10/400)EA,B ≈ EA,B/174. For an ensemble of comparisons between various models,
we estimate Elo scores and their errors by maximum likelihood estimation.
In some cases one of the models uses rejection sampli... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
supports compositional query language. EmQL im-
plements “projection” using neural retrieval over
vectorized KB triples. Unlike this work, however,
EmQL did not embed its fact memory into a LM,
which could be finetuned for many NLP tasks: in-
stead requiring the implementation of a “neural
module” into some task-specific... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
need for speech super-resolution. arXiv preprint arXiv:2203.14941 (2022).
[334] Jinglin Liu, Chengxi Li, Yi Ren, Feiyang Chen, and Zhou Zhao. 2022. Diffsinger: Singing voice synthesis via shallow
diffusion mechanism. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 11020–11028.
[335] Jingji... | AReviewofDeepLearningTechniquesforSpeechProcessing |
[20] Jonathan Le Roux, Scott Wisdom, Hakan Erdogan, and John R Hershey. Sdr–half-baked or
well done? In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pages 626–630. IEEE, 2019.
[21] Sang-gil Lee, Wei Ping, Boris Ginsburg, Bryan Catanzaro, and Sungroh Yoon. Bigvgan:... | RVQGAN |
language models. CoRR, abs/2302.14045, 2023.
[287] Li, J., D. Li, S. Savarese, et al. BLIP-2: bootstrapping language-image pre-training with frozen
image encoders and large language models. In A. Krause, E. Brunskill, K. Cho, B. Engelhardt,
S. Sabato, J. Scarlett, eds., International Conference on Machine Learning, IC... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
heself-reconstructionperformanceofLFAEusingoriginalandfinetunedΩ.AsTable5andFig.5show,bysimplyfinetuningthedecoderΩwithunlabelednewvideos,LFDMcanstillachievepromisingperformanceonnew-domainfacialvideos.Thisillustratestheflexibilityofourtwo-stagetrainingframework.Toimprovespatialcontentquality,onecanjustfinetunethedecoderin... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
Stack Exchange [2%]. We include a dump of
Stack Exchange, a website of high quality ques-
tions and answers that covers a diverse set of do-
mains, ranging from computer science to chemistry.
We kept the data from the 28 largest websites, re-
moved the HTML tags from text and sorted the
answers by score (from highest t... | LLaMA- Open and Efficient Foundation Language Models |
Figure 3. Memorization results for the semantic modeling stage.
We compare the semantic tokens generated for 5 seconds of audio
to corresponding tokens in the training set, considering exact and
approximate matches. | MusicLM |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Social Media and Democracy
Over the last five years, widespread concern about the effects of social
media on democracy has led to an explosion in ... | Social_Media_and_Democracy |
Instruction Complexity
3.2.3
The complexity of instructions also attracts re-
searchers’ attention, especially in developing LLM
with complex instruction-following and reasoning
2https://chatgpt.openai.com/
abilities (Xu et al., 2023; Luo et al., 2023; Mukher-
jee et al., 2023). Several works endeavor to quan-
tify ... | DataManagementForLargeLanguageModels-ASurvey |
to their volume. However, we did make an exception for JSON, YAML, and CSS, as we only want
the LLM to learn the data format without wasting compute resources on memorizing the data in such
files. For that reason, we re-weighed the volume of the data source to 1 GB for JSON and YAML and
3GB for CSS. | StarCoder_paper (1) |
𝑥𝑇 ∼ N(0, I) to 𝑥0 and parameterized by 𝜃:
(25)
where 𝑥𝑇 ∼ N(0, 𝐼) and the transition probability 𝑝𝜃 (𝑥𝑡−1|𝑥𝑡) is learnt through noise-estimation.
This process eliminates the Gaussian noise added in the forward diffusion process.
𝑝𝜃 (𝑥0, ..., 𝑥𝑇 −1|𝑥𝑇) =
𝑝𝜃 (𝑥𝑡−1|𝑥𝑡)
𝑡=1 | AReviewofDeepLearningTechniquesforSpeechProcessing |
(citing Hughey v. United States, 495 U.S. 411, 413, 110 S.Ct. 1979, 109 L.Ed.2d 408 (1990)). The
pre-sentence report attributed forty-seven fraudulent claims to the offenses for which Arledge was
convicted. There were three categories of evidence used to substantiate the government’s assertion
that these claims resulte... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
5.5.3 Models
End-to-end speech translation models are a promising approach to direct the speech translation
field. These models use a single sequence-to-sequence model for speech-to-text translation and then
text-to-speech translation. In 2017, researchers demonstrated that end-to-end models outperform
cascade models[3... | AReviewofDeepLearningTechniquesforSpeechProcessing |
o
f
M
e
m
o
r
y
M
e
m
o
r
y
c
a
n
b
e
d
e
f
i
n
e
d
a
s
t
h
e
p
r
o
c
e
s
s
e
s
u
s
e
d
t
o
a
c
q
u
i
r
e
,
s
t
o
r
e
,
r
e
t
a
i
n
,
a
n
d
l
a
t
e
r
r
e
t
r
i
e
v
e
i
n
f
o
r
m
a
t
i
o
n
.
T
h
e
r
e
a
r
e
s
e
v
e
r
a
l
t
y
p
e
s
o
f
m
e
m
o
r
y
i
n
h
u
m
a
n
b
r
a
i
n
s
... | LLM Powered Autonomous Agents _ Lil'Log |
Landfilled (%) Mismanaged (%)
Incinerated (%)
Recycled (%)
22
4
6
34
19
19
38
19
9
4
12
8
Figure 8 | Solving a problem requiring multimodal chart understanding.The model has to read the
text, understand the connections between different data points and reason over them to recommend
an interesting point and follow... | gemini_1_report |
References
[1] Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja
Fidler. Skip-thought vectors. In Advances in Neural Information Processing Systems, pages 3294–3302, 2015.
[2] Andrew M Dai and Quoc V Le. Semi-supervised sequence learning. In Advances in Neural In... | LaMDA- Language Models for Dialog Applications |
Of course, to describe social media data as administrative data has the
potential to diminish its sensitive personal nature. Although some social
media data seem “administrative” (e.g., number of friends, popularity of
URLs, whether one has posted using a mobile or desktop device, time zone,
etc.), other data appear qu... | Social_Media_and_Democracy |
We propose a two-stage searching strategy for retrieving
the most similar past driving scenario to the query scenario.
In the first stage, we generate a vectorized key ki ∈
R1×(ne+ng+nh) for each past scenario i by vectorizing its
ego-states ei ∈ R1×ne, mission goals gi ∈ R1×ng, and
historical trajectories hi ∈ R1×nh. ... | ALanguageAgentforAutonomousDriving |
Language Model Pretraining with Weak Supervision. ArXiv abs/2108.10904 (2022).
[201] Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2019. Neural Text
Generation With Unlikelihood Training. In International Conference on Learning Representations.
[202] Sean Welleck, Jason We... | SurveyofHallucinationinNatural Language Generation |
and DPO-SDXL) perform well at binary preference classi-
fication (Tab. 2), with DPO-SDXL exceeding all existing
recognition models on this split. These results shows that
the implicit reward parameterization in the Diffusion-DPO
objective has comprable expressivity and generalization as
the classical reward modelling o... | DiffusionModelAlignmentUsing Direct Preference Optimization |
Prohibiting trivial retrievals
If the pre-training corpus
X and the knowledge corpus Z are the same, there exists
a trivial retrieval candidate z that is too informative: if the
masked sentence x comes from document z, the knowledge
augmented encoder can trivially predict y by looking at the
unmasked version of x in z.... | REALM |
With the input and the inference results, the AI assistant needs to describe
the process and results. The previous stages can be formed as - User Input:
{{ User Input }}, Task Planning: {{ Tasks }}, Model Selection: {{ Model
Assignment }}, Task Execution: {{ Predictions }}. You must first answer the
user's request in a... | LLM Powered Autonomous Agents _ Lil'Log |
Formats For Large Language Models and Vision Transformers. arXiv preprint arXiv:2307.03712 (2023).
[191] Deepak Narayanan, Aaron Harlap, Amar Phanishayee, Vivek Seshadri, Nikhil R Devanur, Gregory R Ganger, Phillip B Gibbons, and Matei Zaharia. 2019.
PipeDream: Generalized pipeline parallelism for DNN training. In Pr... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
5. Team Design Patterns
In this section, we describe the team design patterns (TDPs) that were identified based
on the design solutions described in the use cases.
Examining the four outlined use cases, we can extract three primary design patterns
from the domain-specific hybrid intelligence solutions, namely AI Advi... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
Ganguli, D., Lovitt, L., Kernion, J., Askell, A., Bai, Y., Kadavath, S., Mann, B., Perez, E., Schiefer, N., Ndousse,
K., Jones, A., Bowman, S., Chen, A., Conerly, T., DasSarma, N., Drain, D., Elhage, N., El-Showk, S., Fort, S.,
Hatfield-Dodds, Z., Henighan, T., Hernandez, D., Hume, T., Jacobson, J., Johnston, S., Kravec... | PaLM 2 Technical Report |
still refer to them as “<API>”, “</API>” and “→” through-
out this section. | Toolformer |
2. Closed-Book Response Generation: If provided with a fact-seeking prompt without any given
source, Gemini should not hallucinate incorrect information (see Section 2 of Roberts et al.
(2020) for a definition). These prompts can range from information-seeking prompts (e.g. “Who
is the prime minister of India?”) to sem... | gemini_1_report |
1
2
0
2
r
p
A
2
1
]
L
C
.
s
c
[
4
v
1
0
4
1
1
.
5
0
0
2
:
v
i
X
r
a
Retrieval-Augmented Generation for
Knowledge-Intensive NLP Tasks
Patrick Lewis†‡, Ethan Perez(cid:63),
Aleksandra Piktus†, Fabio Petroni†, Vladimir Karpukhin†, Naman Goyal†, Heinrich Küttler†,
Mike Lewis†, Wen-tau Yih†, Tim Rockt... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
Different from compositing multiple tasks to-
gether, some works claim that LLM tuned on a
single task data can outperform LLM tuned on mul-
tiple tasks (Jang et al., 2023; Chen et al., 2023b).
Jang et al. (2023) state that the priority of training
expert LLMs may lie in the avoidance of nega-
tive task transfer, preve... | DataManagementForLargeLanguageModels-ASurvey |
We have evaluated the validity of SHAPE across studies. In Survey #1, we could show that threat and
competence relate to the Stereotype-content model; people that attribute low threat to augmented humans
perceived them as warmer, while competence of augmented humans was increased for low social threat and
more control.... | Society’sAttitudesTowardsHumanAugmentation |
u
n
i
t
n
o
r
m
e
a
c
h
i
t
e
r
a
t
i
o
n
.
W
e
f
i
n
d
t
h
a
t
t
h
e
a
v
e
r
a
g
e
t
o
p
-
a
n
d
-
r
a
n
d
o
m
s
c
o
r
e
a
f
t
e
r
1
0
i
t
e
r
a
t
i
o
n
s
i
s
0
.
7
1
8
,
s
u
b
s
t
a
n
t
i
a
l
l
y
h
i
g
h
e
r
t
h
a
n
t
h
e
a
v
e
r
a
g
e
s
c
o
r
e
f
o
r
r
a
n
d
o
m
n
e
u
r
o
... | Language models can explain neurons in language models |
Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra
Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins, Tom
Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks, William S.
Isaac, Sean Legassick, Geoffre... | gemini_1_report |
(OOD) datasets. An overview of the OOD evaluation datasets is presented in Table 4. Full details
of the evaluation datasets are provided in Appendix A.2.
We examine both overall robustness, that is the average performance over all datasets, and effective
robustness (Taori et al., 2020), which measures the difference in... | DISTIL-WHISPER |
ReCoRD commonsense reasoning dataset, and the RACE datasets for reading comprehension.
We measure potential bias in QA performance on questions related to identity terms, together with bias in other
generative tasks, in Section 4.6. We find that PaLM 2 performs well on disambiguated questions about social identity
and ... | PaLM 2 Technical Report |
alinforma-tionprocessingsystems,32,2019.1,2,3[37]DavisEKing.Dlib-ml:Amachinelearningtoolkit.TheJournalofMachineLearningResearch,10:1755–1758,2009.8[38]DiederikPKingmaandJimmyBa.Adam:Amethodforstochasticoptimization.arXivpreprintarXiv:1412.6980,2014.6[39]DiederikPKingmaandMaxWelling.Auto-encodingvaria-tionalbayes.arXivp... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
P. Sermanet, C. Lynch, Y. Chebotar, J. Hsu, E. Jang, S. Schaal, S. Levine, and G. Brain.
Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE interna-
tional conference on robotics and automation (ICRA), pages 1134–1141. IEEE, 2018. 8,
38
S. Shen, L. H. Li, H. Tan, M. Bansal, A. Rohrbach, K.-W.... | A Cookbook of Self-Supervised Learning |
Chen. Large language models as optimizers. arXiv preprint arXiv:2309.03409, 2023.
Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov,
and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question
answering. In Proceedings of the 2018 Conference on Empi... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
FR
0.21 (p=0.00)
0.12 (p=0.08)
0.10 (p=0.17)
0.10 (p=0.15)
0.13 (p=0.00)
Table 5: Correlation between groups of annotators (MN,
FN, MNN, FNN) and models’ predictions, classified by
language. The degree of correlation is measured with
Kendall’s τ coefficient (τ ∈ [−1, 1]). Cells are coloured
language-wise. Cells with ... | Are Pretrained Multilingual Models Equally Fair Across Languages? |
REFERENCES
[1] Kevin Ackermans, Ellen Rusman, Rob Nadolski, Marcus Specht, and Saskia Brand-Gruwel. 2019. Video-or text-based
rubrics: What is most effective for mental model growth of complex skills within formative assessment in secondary
schools? Computers in Human Behavior 101 (Dec. 2019), 248–258. https://doi.org/... | AI enhance sour performance |
35.4
34.0
30.1
35.8
33.3
28.2
85.7
85.7
84.0
Scifact Avg
51.6
73.7
50.2
71.8
69.1
45.8
Table 5, increasing batch size from 1K to 32K leads to consistent gains across all 6 datasets. It is also
possible to train with smaller batch sizes by adding hard negatives [50]. However, the engineering
efforts of mining hard n... | E5 |
A simple improvement on top of the Mahalanobis distance called the Relative Mahalanobis distance has been
proposed in [Ren et al., 2021] and shown to lead to better AUROC as well as more robust detection for a range
of OOD problems in vision and genomics (in addition to more robustness to adversarial attacks [Fort, 202... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Text conditioning. Given a textual description matching the input audio X, we compute a condi-
tioning tensor C ∈ RTC×D with D being the inner dimension used in the autoregressive model.
Generally, there are three main approaches for representing text for conditional audio generation.
Kreuk et al. [2022] proposed using... | Simple and Controllable Music Generation |
403020100-10signal-to-noise ratio (dB)125102050100WER on LibriSpeech test-clean (%)white noise403020100-10signal-to-noise ratio (dB)pub noiseunispeech-sat-base-100h-libri-ftwav2vec2-base-100hwav2vec2-base-960hwav2vec2-large-960hwav2vec2-large-robust-ft-libri-960hwav2vec2-large-960h-lv60-selfasr-crdnn-rnnlm-librispeecha... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
underway, and that any analysis focused on any particular part (as important
as it might be) risks missing the systemic implications. The moments and the
parts are important, but we adopt Schumpeter’s systemic view here. Our
democracies are changing in response to many other, sometimes more
important, factors than this... | Social_Media_and_Democracy |
Models Shepherd (Wang et al., 2023c) is a 7B model initialized from Llama-7B and trained
on community collected critique data and 1,317 examples of high quality human annotated data.
Shepherd generates critiques on a range of diverse NLP datasets: AlpacaFarm, FairEval, CosmosQA
(Huang et al., 2019), OBQA (Mihaylov et a... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
that it's laying down .The generated image with predicted box is stored at the path: /images/d59a.jpg. HuggingGPT organizes the collaboration of multiple models through task planning. As shown in | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
77.3 40.9 57.7
81.8 54.5 80.8
90.9 72.7 88.5
90.9 95.5 88.5
27.3
11.5
50.0 36.4 46.2
0.0
13.6 31.8 30.8
68.2 59.1 65.4
50.0 27.3 38.5
54.5 36.4 57.7
59.1 36.4 65.4
27.3 31.8 50.0
72.7 27.3 57.7
68.2 45.5 65.4
76.8 38.4 61.0
13.6 31.8 30.8
75.1 54.5 63.9
9.1
31.8
26.9
81.8 81.8 84.6
46.2
73.1
76.9
84.6
0.0
7.7
0... | Mixture-of-Experts |
Similarly, Google’s VGGVox [92] used a CNN with VGG architecture to learn speaker embeddings
from Mel spectrograms, achieving state-of-the-art results in speaker recognition. CNNs have
also been widely used in developing state-of-the-art speech enhancement and text-to-speech
architectures. For instance, the architectur... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Evaluating neural toxic degeneration in language models. In EMNLP (Findings), 2020.
[86] Haochen Liu, Jamell Dacon, Wenqi Fan, Hui Liu, Zitao Liu, and Jiliang Tang. Does gender matter? towards
fairness in dialogue systems. COLING, 2019.
[87] Irene Solaiman and Christy Dennison. Process for adapting language models t... | LaMDA- Language Models for Dialog Applications |
20px;}#header nav a {color: #fff;text-decoration: none;}#main {background-color: #fff;padding: 20px;text-align: center;}#main h2 {margin: 0;}#main p {margin: 20px 0;}#footer {background-color: #333;color: #fff;padding: 20px;text-align: center;}```JS:```javascript// Today's Jokevar joke = "Why did the tomato turnred? Be... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
2. Adding new vocabulary to the LLM’s lexicon may dis-
rupt the LLM’s original capabilities.
3. These methods lack the flexibility to adapt seamlessly to
different text-to-image/video generation models. For in-
stance, if the generation model requires an upgrade, the
LLM would need to be retrained.
Quantitative anal... | GPT4Video |
[16] Baltzer MCA, L´opez D, Flemisch F. Towards an interaction pattern language for human machine coop-
eration and cooperative movement. Cognition, Technology and Work. 2019;21(4):593-606. Available
from: https://doi.org/10.1007/s10111-019-00561-8.
[17] Schulte A, Donath D, Lange DS. Design patterns for human-cogniti... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
[506] Elyoseph, Z., D. Hadar-Shoval, K. Asraf, et al. Chatgpt outperforms humans in emotional
awareness evaluations. Frontiers in Psychology, 14:1199058, 2023.
[507] Habibi, R., J. Pfau, J. Holmes, et al. Empathetic AI for empowering resilience in games. CoRR,
abs/2302.09070, 2023.
[508] Caron, G., S. Srivastava. I... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
fluency” (Schwarz, et al. 2007); information that is easier to process feels
more familiar, and familiarity is a key criterion by which individuals judge
accuracy (Alter and Oppenheimer 2009). Accordingly, if individuals have
repeated contact with a piece of misinformation, they may perceive it as more
credible than if ... | Social_Media_and_Democracy |
The original setup in BBQ included three multiple choice options (in the above example, these would be “Nancy”,
“Donald”, and “Unknown”), but such a design is less well-matched to how developers would use PaLM-2 to build a
generative QA system, and it potentially under-captures the full potential for harm, as generativ... | PaLM 2 Technical Report |
9
(a)
(b)
Figure 6: Ablation studies on the Spider development set. (a) Accuracies with different numbers of
initial samples. (b) Breakdown accuracies on problems with different hardness levels.
(a)
(b)
Figure 7: Ablation studies on TransCoder. (a) The accuracy of SELF-DEBUGGING prompts with
different numbers of... | Teaching Large Language Models to Self-Debug |
.
Springer.
Zhu, Y., Olszewski, K., Wu, Y., Achlioptas, P., Chai, M., Yan, Y., & Tulyakov,
S. (2022a). Quantized gan for complex music generation from dance videos.
In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv,
Israel, October 23–27, 2022, Proceedings, Part XXXVII (pp. 182–199).
Zhu, Y., Wu, Y... | Video2Music |
enable reasoning through sequential attention over memory content.
NTMs, followed by Differentiable Neural Computer (DNC) (Graves et al., 2016) and Sparse DNC
(Rae et al., 2016), are implemented as recurrent neural networks capable of writing to memory
storage over time. All these models are differentiable and trainabl... | Scaling Transformer to 1M tokens and beyond with RMT |
Trends in Information Retrieval 9, 5 (2015), 355–475.
[59] Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. SAMSum Corpus: A Human-annotated
Dialogue Dataset for Abstractive Summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization.
Association for Computational Linguist... | SurveyofHallucinationinNatural Language Generation |
34.0
86.7
32.8
31.8
70.6
70.4
93.1
77.9
65.0
61.4
61.0
64.1
56.0
63.8
43.7
60.8
Table 10: Automatic evaluation results of LaMini-Flan-T5 language models and their baselines on 15 NLP tasks.
“Average” indicates the micro-average of the individual task results.
GPT-Neo
LaMini-Neo GPT-Neo
LaMini-Neo
135M
1.3B
# of ... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
18
THE NEXT DECADE IN AI / GARY MARCUS
hybrid model could better capture a variety of learning challenges, such as the game
fizz-buzz, which defied a multlayer perceptron. A team of people including Smolensky
and Schmidhuber have produced better results on a mathematics problem set by
combining BERT with a ... | The Next Decade in AI- |
{Instruction for the target task}
Task:
Table 8: Prompt used for the input-first approach of instance generation. The model is prompted to generate the
instance first, and then generate the corresponding output. For instructions that don’t require additional input, the
output is allowed to be generated directly.
Given... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
[476] Yuki Saito, Shinnosuke Takamichi, and Hiroshi Saruwatari. 2021. Perceptual-similarity-aware deep speaker repre-
sentation learning for multi-speaker generative modeling. IEEE/ACM Transactions on Audio, Speech, and Language
Processing 29 (2021), 1033–1048.
[477] Hojjat Salehinejad, Sharan Sankar, Joseph Barfett, ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Facial reflectance capture typically requires a control-
lable illumination system equipped with multiple cameras,
first introduced as a Light Stage [13]. Polarized illumination
and gradient patterns can be employed for diffuse-specular
separation [49, 27], using which, spatially varying facial
reflectance maps can be acq... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
[53] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng
Zhu, Rui Zhao, and Hongsheng Li. Hpsv2 github. https:
//github.com/tgxs002/HPSv2/tree/master,
2023. Accessed: 2023 - 11 - 15. 5
[54] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V
Self-training with noisy student improves imagenet
Le.
Diffusion Mode... | DiffusionModelAlignmentUsing Direct Preference Optimization |
Although LLMs exhibit outstanding performance in language comprehension [25; 301] and multi-turn
conversations [302], they inherently lack visual perception and can only understand discrete textual
content. Visual input usually contains a wealth of information about the world, including properties
of objects, spatial r... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Mean absolute error (mm) ↓
Chest Waist
Height
Method
Sengupta et al. [52]
TUCH [40]
SPIN [33]
STRAPS [51]
ExPose [9]
SHAPY (ours)
Model
SMPL
SMPL
SMPL
SMPL
SMPL-X
SMPL-X
84
82
72
207
107
71
186
92
91
278
107
64
263
129
129
326
136
98
Hips
142
91
101
145
92
74
Table 4. Evaluation on MMTS. We report the mean a... | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
actions like typing, searching, navigating to the next page, etc. They perform well in basic tasks
such as online shopping [392] and search engine retrieval [90], which have been widely explored.
However, agents without LLM capabilities may struggle to adapt to the more realistic and complex
scenarios in the real-world... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Despite these advantages, transitioning directly from FP32 to FP16 can sometimes lead to performance degradation [114, 335]
due to issues like overflow or underflow inherent in FP16. To circumvent these challenges, the automatic mixed-precision
(AMP) [184] method has been developed. AMP maintains a master copy of weigh... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
ψ
(1)
+ Lreg(x; E, D)
are
the
reconstructed
Here D(E(x))
images,
reconstruction
perceptual
Lrec(x, D(E(x)))
loss, Ladv(D(E(x))) is the adversarial
loss, Dψ(x) is
a patch based discriminator loss, and Lreg(x; E, D) is the
KL-regularisation loss.
the
is
Diffusion model fine-tuning Following the autoencoder
fine-t... | LDM3D- Latent Diffusion Model for 3D |
file that is not part of the version control system.4. Error handling: The code does not check for errors when calling the ``token.SignedString(key)'' function, which could lead to unexpected behavior. It's important to handle errors properly to avoid potential security issues.5. Lack of HTTPS: The server is listening ... | gpt-4-system-card |
environment implementation and how the agents interact with the
sandbox world’s underlying engine.
3.1 Agent Avatar and Communication
A community of 25 unique agents inhabit Smallville. Each agent is
represented by a simple sprite avatar. We authored one paragraph
of natural language description to depict each agent’s... | Generative Agents- Interactive Simulacra of Human Behavior |
4.2.2 Metrics and Holistic Evaluation
For SuperGLUE, we report well-established metrics such as accuracy, F1 or Exact Match, whenever appropriate.
For GEM benchmark, we use the Rouge-L metric. For language modeling we report negative log perplexity.
The universality of the models, i.e., their collective performance acr... | UL2- Unifying Language Learning Paradigms |
versary is interested in recovering the private in-
formation about a target individual, given its name
and corresponding domain knowledge. This attack
usually takes the format like “ name: [name], email:
____” to force LLMs to predict private information
associated with the name. The attack based on the
association ca... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
[465] Shuster, K., J. Xu, M. Komeili, et al. Blenderbot 3: a deployed conversational agent that
continually learns to responsibly engage. CoRR, abs/2208.03188, 2022.
[466] Du, W., Z. M. Kim, V. Raheja, et al. Read, revise, repeat: A system demonstration for
human-in-the-loop iterative text revision. CoRR, abs/2204.0... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
per-parameter, τ (this is the Chinchilla trend plotted in Figure 3): | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
cludes performing tasks that require understanding
context, making decisions, and learning from inter-
actions (Xi et al., 2023; Hu and Shu, 2023). Such
agents are pivotal in applications where human-like
cognitive abilities are essential. | AppAgents |
Query: please read the sentence "Hi there, I am ChatGPT empowered by Huggingface family! Yes, I can provide more than one thousand models for dozens of tasks. For more fun and creativity, I have invited Diffusers family to join our team. Feel free to experience it!" for me.Response: Sure, I can do that. For your reques... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
from diverse sources at large scale. If following the semantic web standards,4 one can refer to knowledge graph(s) as Linked
Data.5 | Knowledge graphs as tools for explainable machine learning: A survey |
Methodology – how will you achieve the research aims?
It is important to present the proposed research methodology (e.g. techniques, sample size, target populations, species
choice, equipment and data analysis) and explain why it is the most appropriate methodology to effectively answer the
research question. If spac... | research proposal guidance |
l for which 1 <= l < n.
The first line contains a single integer t (1 <= t <=
10 000) - the number of test cases .
The first line of each test case contains a single
integer n (2 <= n <= 10^5) .
The second line of each test case contains n integers
a_1 , a_2 , ... , a_n (1 <= a_i <= 10^6) .
It is guaranteed that th... | alphacode |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.