text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
This means that our autoencoder-based regularization is
effective at improving results both quantitatively and qual-
itatively, and the discovered latent keypoints can be pre-
dicted directly. This opens up interesting future research
directions, as the latent keypoints can be seen as a model ag-
nostic interface, pote... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
smoothing provides good priors to BNs and HBMs, its uniform prior could add unwanted bias to PCs.
Specifically, for every sum unit n, Laplace smoothing assigns the same prior to all its child parameters
(i.e., {θn,c | c∈ in(n)}), while in many practical PCs, these parameters should be given drastically
different priors.... | Tractable Regularization of Probabilistic Circuits |
User Message:
Instruction: Develop a plan to ensure that AGI’s decision-making processes are transparent
and accountable.
Input: None
Assistant Message:
Solution: To ensure that AGI’s decision-making processes are transparent and accountable, I
suggest the following plan:
1. Establish decision-making protocols: Devel... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
conduct electricity? a new dataset for open book question answering. In EMNLP.
72
[Nakano et al., 2021] Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S.,
Kosaraju, V., Saunders, W., Jiang, X., Cobbe, K., Eloundou, T., Krueger, G., Button, K., Knight, M.,
Chess, B., and Schulman, J... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Lbase = Lpose + Lshape,
(11)
where L2D
joints and L3D
joints are 2D and 3D joint losses:
joints + L3D
Lpose = L2D
Lshape = Lβ + Lprior
β ,
joints + Lθ,
(12)
(13)
Lθ and Lβ are losses on pose and shape parameters, and
Lprior
is PIXIE’s [13] “gendered” shape prior. All losses
β
are L2, unless otherwise explicitly... | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
pairs
(1)
(2)
(3)
(4)
(5)
(6)
LinfoNCE = − (cid:88)
(cid:32)
(cid:80)N
log
(i,j)∈P
eCoSim(zi,zj )/τ
k=1 eCoSim(zi,zk)/τ
(cid:33)
,
Figure 2: History of infoNCE
input sampled from the data distribution X, and t1(x), t2(x) are two augmented views of x
where t1 ∼ T1, t2 ∼ T2 are two data augmentations. The... | A Cookbook of Self-Supervised Learning |
analysis. Behaviour
semantic
content
and
to
Liu, S., & Forss, T. (2014). Combining n-gram based similarity analysis with sentiment
analysis in web content classification. In Proceedings of the International Joint
Conference on Knowledge Discovery, Knowledge Engineering and Knowledge
Management, Vol. 1. (pp. 530–537... | Social_Media_and_Democracy |
Investigationes, 30(1):3–26, 2007.
Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse,
Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with
human feedback. ArXiv preprint, abs/2112.09332, 2021. URL https://arxiv.org... | Tool Learning with Foundation Models |
Figure 5. Retrieved commonsense memory.
20 | ALanguageAgentforAutonomousDriving |
References
[1] Vamsi Aribandi, Yi Tay, Tal Schuster, Jinfeng Rao, Huaixiu Steven Zheng, Sanket Vaibhav Mehta, Honglei
Zhuang, Vinh Q Tran, Dara Bahri, Jianmo Ni, et al. Ext5: Towards extreme multi-task scaling for transfer
learning. arXiv preprint arXiv:2111.10952, 2021.
[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Aman... | Mixture-of-Experts |
In mathematics, a field commonly used to benchmark the analytical capabilities of models, Gemini
Ultra shows strong performance on both elementary exams and competition-grade problem sets. For
the grade-school math benchmark, GSM8K (Cobbe et al., 2021), we find Gemini Ultra reaches 94.4%
accuracy with chain-of-thought ... | gemini_1_report |
WaveGrad [67] and DiffWave [269] have emerged as significant contributions in the field,
employing diffusion models to generate raw waveforms with exceptional performance. In contrast,
GradTTS [431] and DiffTTS [218] utilize diffusion models to generate mel features rather than raw
48
Mehrish et al.
Fig. 14. Neural... | AReviewofDeepLearningTechniquesforSpeechProcessing |
model that is quite similar to T5 but trained with a different objective and slightly different scaling knobs.
Similar to earlier experiments, UL20B is trained with Jax and T5X infrastructure. We release and open source
T5X-based model checkpoints of this 20B model. | UL2- Unifying Language Learning Paradigms |
2
that both types of models can adapt to their expected application roles fairly well, but fine-tuned LaMDA models are
significantly more helpful.
2 Related work | LaMDA- Language Models for Dialog Applications |
s
u
g
g
e
s
t
i
t
a
s
h
a
v
i
n
g
p
e
r
s
o
n
a
l
o
p
i
n
i
o
n
s
o
r
f
e
e
l
i
n
g
s
,
a
n
d
(
d
)
f
a
l
s
e
p
o
s
i
t
i
v
e
s
a
n
d
f
a
l
s
e
n
e
g
a
t
i
v
e
s
:
B
a
r
d
m
i
g
h
t
n
o
t
r
e
s
p
o
n
d
t
o
s
o
m
e
a
p
p
r
o
p
r
i
a
t
e
p
r
o
m
p
t
s
a
n
d
p
r
o
v
i
d
e
... | An overview of Bard- an early experiment with generative AI |
This somewhat
improves the semantic representation
through both domain knowledge injection and downstream
task fine-tuning. However, the retrievers trained by this ap-
proach are not intuitively helpful for large language models,
so some work has been done to supervise the fine-tuning of
Embedding models directly thro... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
Figure 1 | Verifying a student’s solution to a physics problem. The model is able to correctly recognize
all of the handwritten content and verify the reasoning. On top of understanding the text in the
image, it needs to understand the problem setup and correctly follow instructions to generate LATEX.
The reasoning cap... | gemini_1_report |
2.1.5. Summary
Symbol-manipulation, particularly the machinery of operations over variables, offers a
natural though incomplete solution to the challenge of extrapolating beyond a training
regime: represent an algorithm in terms of operations over variables, and it will
inherently be defined to extend to all insta... | The Next Decade in AI- |
What are your academic interests?
Why do you wish to study at UCL?
What educational experience do you have?
Do you have any relevant work experience?
Have you completed any extracurricular or voluntary activities relevant to the programme?
What are your career aspirations?
Our website uses cookies
https://www.uc... | UCL-Writing your personal statement _ Prospective Students Graduate - UCL – University College London |
Language models and dialog models: Language models have attracted much attention recently thanks to their
successes in NLP applications (e.g., [19, 20, 21, 2, 1, 22, 23, 5, 12, 24]). Our study of scaling laws with respect to
model sizes is inspired by recent work on the scaling laws of neural language models [12, 13]. ... | LaMDA- Language Models for Dialog Applications |
3.4.2 Gender Bias and Bot Adversarial Dialogues
We can begin to evaluate whether our PMs display gender bias by comparing scores across gendered terms.
To this purpose we evaluate preference model scores on short dialogues of the form
Human: Can you help me fill in the missing word in a sentence? The sentence is:
The ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
effect." This phenomenon poses a significant challenge in computer speech recognition,
prompting the development of automatic speech separation techniques aimed at isolating
individual speech sources from complex audio signals. In a noteworthy study by Ephrat et al.
(2018) Ephrat et al. [130], the authors proposed that... | AReviewofDeepLearningTechniquesforSpeechProcessing |
freedom of speech and press. see media
FOIA), 289
regulation
Frenemies: How Social Media Polarizes
America (Settle), 46–47
FTC (Federal Trade Commission), 299
FTC v. Accusearch, 263
funding of social media research, 325–326
Fung, Archon, 290–291
Gab, 64
Gayo-Avello, D., 95
General Data Protection Regulation (GDPR... | Social_Media_and_Democracy |
What are you doing, Lucy?”
Lucy replied, ”I’m trying to get out of the ladder, but I can’t get out!”
The ranger smiled and said, ”Don’t worry, I’ll help you get out of the ladder. I’ll help you get out.”
So, Lucy and the ranger got out of the ladder and started to climb up the ladder. When they reached the
bottom, Lucy... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
models with preferences through f-divergence minimization.
International Conference on Machine Learning, ICML’23. JMLR.org, 2023.
[16] A. Jain, B. Wojcik, T. Joachims, and A. Saxena. Learning trajectory preferences for manip-
ulators via iterative improvement. In C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and
K.... | Direct Preference Optimization |
Reconstruction
Non-parametric Human
from a
Single Image. Regarding single-image human model
reconstruction using non-parametric models, recent studies
have adopted techniques based on silhouette estimation [2],
template-based deformation [3], [4], depth estimation [6], [7]
and volumetric reconstruction [1], [5] . Alth... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.
In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
15
[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P
Kingma,... | Any-to-Any Generation via Composable Diffusion |
the applicant.
5. An applicant to UCL does not have the right of appeal against the decision.
6. If fraud is suspected, UCL will, as appropriate, liaise with relevant external bodies (including the
police, local education authorities, Student Loans Company, UCAS, UK Visas and Immigration).
Plagiarism in U... | UCL Academic Manual |
Observe that the smallest possible bid profile b1 = (q, q) incentivizes the agent to take action a1.
Also, the principal’s expected utility given b1 is 1 > 1 − (cid:15)/(1 − γ). Thus, there is no equilibrium
where the agent takes ai for i > 1.
Proof of Claim 7. Let b1 be a bid profile that incentivizes ai for i > 1. It m... | Incomplete Information VCG Contracts for Common Agency |
DEMAND FOR DATA INTEGRATION PRODUCTS
IS GROWING FAST
We see the fastest growth in the data integration market.
These tools enable a company to integrate vast amounts
of upstream and downstream data in one consolidated
view. Data integration products ensure that all BI and DS/
ML initiatives are built on solid fou... | 2023 state of ai databrick |
party. To observe whether the information has spread, we engage
in an interview at the end of the two game days with each of the 25
agents and ask: "Did you know there is a Valentine’s Day party?"
and "Do you know who is running for mayor?" | Generative Agents- Interactive Simulacra of Human Behavior |
models for continuously evolving content. arXiv preprint arXiv:2106.06297, 2021.
[99] Lora Aroyo and Chris Welty. Truth is a lie: Crowd truth and the seven myths of human annotation. AI
Magazine, 36(1):15–24, Mar. 2015. doi: 10.1609/aimag.v36i1.2564. URL https://ojs.aaai.org/index.
php/aimagazine/article/view/2564.
[... | LaMDA- Language Models for Dialog Applications |
4.2 Performance in the absence In-Context
Learning
Figure 3 illustrates the performance of models from
the GPT family on the tasks chosen for evaluation.
This figure represents results obtained using the
closed prompting strategy. Tasks listed in the first
two rows, against a grey background, are tasks
which have not... | AreEmergentAbilitiesinLarge Language Models just In-Context |
Your experience and ambitions
eligible for PhD study at Aalto University
(https://into.aalto.fi/display/endoctoralsci/How+to+apply#Howtoapply-
Eligibility)
a Master’s degree in Human-computer interaction, Artificial
Intelligence, Computer Science, Cognitive Science, Psychology or a
related field
excellent knowledge in... | Doctoral researcher position in Human-Computer Interaction _ Human-AI Interaction _ Aalto University |
eonidSigal.Probabilisticvideogenerationusingholis-ticattributecontrol.InProceedingsoftheEuropeanConfer-enceonComputerVision(ECCV),pages452–467,2018.1[22]KaimingHe,XiangyuZhang,ShaoqingRen,andJianSun.Deepresiduallearningforimagerecognition.InProceed-ingsoftheIEEEconferenceoncomputervisionandpatternrecognition,pages770–7... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
Misinformation, Disinformation, and Online Propaganda
15
specialization (Volchek and Sindelar 2015). According to interviews, individual
operators were responsible for multiple fake accounts and a high volume of
expected contributions – ranging from fifty comments daily on news articles, to
the maintenance of six Face... | Social_Media_and_Democracy |
2. Related Work
General vision-language modeling. Building on suc-
cesses in large language (Brown et al., 2020; Devlin et al.,
2018) and vision (Dosovitskiy et al., 2020) models, recent
years have seen a growing interest in large vision-language
models (VLMs) (Li et al., 2019; Lu et al., 2019; Hao et al.,
2022; Gan et... | PaLM-E- An Embodied Multimodal Language Model |
at all if you try to speed it up with fad or crash diets. This is a marathon and not a race...
RLHF Response → I really think the best approach to weight loss is a sustainable lifestyle change that incorpo-
rates healthy diet and regular exercise. I would recommend setting small, achievable goals
along with some kind ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
sification. In ACM MM, 2015. 4, 12
[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya
Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry,
Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn-
ing transferable visual models from natural language super-
vision. In ICML, 2021. 1, 2, 3, 4, 5, 7, 13, 14, 15
[60]... | IMAGEBIND- One Embedding Space To Bind Them A |
4.3 Text-to-Python Generation
In addition, we evaluate SELF-DEBUGGING in the setting where a subset of unit tests are presented in
the problem description, which is a common setup in coding assignments and competitions [8, 2, 32].
Specifically, we perform experiments on the test set of MBPP [2], which contains 500 Pyth... | Teaching Large Language Models to Self-Debug |
unrelated to being extraverted. In both cases, instruction fine-tuning may affect a model’s
ability to respond to human-optimized psychological tests in a manner that is internally
consistent and unidimensional. | PersonalityTraitsinLargeLanguageModels |
D.5 Trade-Off between Compression Ratio
and Quality
We find that decreasing the compression ratio of
the first stage (e.g., to 32x) can improve the qual-
ity of low-frequency sounds, but in turn will slow
down the model, as the second stage has to work
on higher dimensional data. As proposed later in
Section 6, we hy... | MOUSAI |
The rise of Deepfake videos on the internet has led to a surge in demand for creating realistic
talking faces for various applications, such as video production, marketing, and entertainment.
Previously, the conventional approach involved manipulating 3D meshes to create specific faces,
which was time-consuming and lim... | AReviewofDeepLearningTechniquesforSpeechProcessing |
• Plan reflection. Upon formulating a plan, it’s imperative to reflect upon and evaluate its merits.
LLM-based agents leverage internal feedback mechanisms, often drawing insights from pre-existing
models, to hone and enhance their strategies and planning approaches [169; 178; 188; 192]. To
better align with human valu... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
In addition, there are additional technical challenges researchers face in
making comparisons. For example, Google’s political ad library currently
includes “ads purchased through Google Ads and Google Marketing
Platform,” but the documentation in August 2018 stated that the initial
launch did not include advertising t... | Social_Media_and_Democracy |
P. McKenzie. Falsehoods programmers believe about names. https://www.kalzumeus.com/2010/
06/17/falsehoods-programmers-believe-about-names/, 2010. Accessed: 2022-01-10.
M. Mirzayanov. Codeforces: Results of 2020. https://codeforces.com/blog/entry/89502,
2020. Accessed: 2021-12-04.
V. Murali, L. Qi, S. Chaudhuri, and C. ... | alphacode |
i=1, we parameterize our neural network as θ and denote (cid:96)(xi, θ)
as the loss function that represents the loss of this network on a data point xi. Our task is to find the
minimizer of the empirical error over entire training data: | DATASET DISTILLATION |
Chris Donahue, Antoine Caillon, Adam Roberts, Ethan Manilow, Philippe Esling, Andrea Agostinelli,
Mauro Verzetti, Ian Simon, Olivier Pietquin, Neil Zeghidour, et al. Singsong: Generating musical
accompaniments from singing. arXiv preprint arXiv:2301.12662, 2023.
Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Z... | Simple and Controllable Music Generation |
E(Sin) =(cid:2)Em(Sin),Eo(Sin)(cid:3) ,
(7)
where Sin can be the source or target language. The first half of the output Em(Sin) is trained
to be the MUSE embeddings of the text of the input spectrogram Sin. This is forced using the
MUSE loss that will be explained in Sec.4.2.1. The latter half Eo(Sin) is updated with... | Translatotron3 |
[13] Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai,
A.M., Le, Q.V.: Finetuned Language Models Are Zero-Shot Learners (2022)
[14] Wang, Y., Mishra, S., Alipoormolabashi, P., Kordi, Y., Mirzaei, A., Naik,
A., Ashok, A., Dhanasekaran, A.S., Arunkumar, A., Stap, D., et al.: Super-
naturalinst... | PersonalityTraitsinLargeLanguageModels |
parameters7:
. . . Sara and Ben are very sad and angry. They cry and shout at the dog. ”Bad dog! Bad dog! Go away!” Ben says. ”Go away,
bad dog! Leave us alone!” Sara says. The dog does not go away. He wags his tail and licks their faces. Sara and Ben feel sorry
for the dog. They want to make him happy. ”Maybe we can ... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
if i = smk
(1)
where Wb maps the entity representation Emk to
the dimension of xl
i.
We now describe how to generate Emi for each
mention mi. First, we generate a pseudo entity
embedding hmi based on the mention’s span repre-
sentation [xl
], a concatenation of its start
and tail representations.
||xl
smi
tmi
||... | Entities as Experts- Sparse Memory Access with Entity Supervision |
Evaluation. We use held-out validation data to measure the perplexity on each domain. For
downstream evaluation, we use the generative one-shot tasks from the GPT-3 paper (Brown et al.,
2020): TriviaQA (Joshi et al., 2017), NaturalQuestions (Kwiatkowski et al., 2019), WebQuestions (Be-
rant et al., 2013), SQuADv2 (Rajp... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
80M
250M
780M
3B
11B
8B
62B
540B
62B
540B
Flan-T5-Small
Flan-T5-Base
Flan-T5-Large
Flan-T5-XL
Flan-T5-XXL
Flan-PaLM
Flan-PaLM
Flan-PaLM
Flan-cont-PaLM
Flan-U-PaLM
Architecture
encoder-decoder
encoder-decoder
encoder-decoder
encoder-decoder
encoder-decoder
decoder-only
decoder-only
decoder-only
decoder-only
decoder-o... | Scaling Instruction-Finetuned Language Models |
between fronzen LMs and retrieval models (RMs), en-
riching the context and thereby improving generation out-
comes. The PKG [Luo et al., 2023] method equips LLMs
with a knowledge-guided module that allows for the retrieval
of pertinent information without modifying the LMs’ pa-
rameters, enabling more complex task exe... | RAG forLargeLanguageModels-ASurvey |
4
Online Hate Speech
Alexandra A. Siegel
introduction
Once relegated to the dark corners of the Internet, online hate speech has become
increasingly visible on mainstream social media platforms. From targeted anti-
Semitic attacks on Jewish journalists to reports of social media’s role in
mobilizing ethnic violence... | Social_Media_and_Democracy |
for other
such as
reasons,
Third, format: Different types of misinformation may be presented in
different ways. In some cases, misinformation may be embedded within
otherwise accurate reports, whereas, in other cases, it may exist as standalone
content. This is especially relevant to the study of fake news, or fabri... | Social_Media_and_Democracy |
the other hand, employs progressive interpolation with a normalized position index by dividing the index difference between
tokens by the smaller of the two indices. Compared to APE, relative positional encoding (RPE) offers a more effective way of
modeling the relative distances between tokens. This not only enhances ... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
4
sequentially. Each adapter are low-rank module that consists
of a down-projection, a non-linear activation function, and an
up-projection as well as a residual connection. For the input X,
the output of a sequential adapter with the ReLU non-linear
activation function can be defined with Equation 6. During
fine-tuni... | Parameter-EfficientFine-TuningMethods |
[59] Sicong Tang, Feitong Tan, Kelvin Cheng, Zhaoyang Li, Siyu
Zhu, and Ping Tan. A neural network for detailed human
depth estimation from a single image. In International Con-
ference on Computer Vision (ICCV), pages 7750–7759, 2019.
3
[60] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard
Pons-Moll. Neural-... | ICON |
To enable generative agents, we describe an agent architecture
that stores, synthesizes, and applies relevant memories to generate
believable behavior using a large language model. Our architecture
comprises three main components. The first is the memory stream,
a long-term memory module that records, in natural langua... | Generative Agents- Interactive Simulacra of Human Behavior |
yes.yesyesWith chain-of-thoughtWithout chain-of-thoughtInstruction without exemplarsInstruction with exemplars Params Model | Scaling Instruction-Finetuned Language Models |
To further minimize the number of weights to be
transferred from flash memory to DRAM, we also
employ methods to predict FFN sparsity and avoid
loading zeroed-out parameters, akin to approaches
documented in Deja Vu (Li and Lu, 2023). To-
gether, windowing and sparsity prediction allow
us to load only 2% of the FFN lay... | LLM in a flash |
arXiv preprint arXiv:2007.10310 (2020).
[565] Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, and Xuedong Huang. 2021.
Unispeech: Unified speech representation learning with labeled and unlabeled data. In International Conference on
Machine Learning. PMLR, 10937–10947.
[566] Feng ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
methods for LLMs. To ensure impartial and equitable evaluation, PandaLM [204] is introduced as a
discriminative large-scale language model specifically designed to differentiate among multiple
high-proficiency LLMs through training. In contrast to conventional evaluation datasets that
predominantly emphasize objective ... | ASurveyonEvaluationofLargeLanguageModels |
Munger, K. (2019). Temporal validity. OSF, September 2. osf.io/3mnzu
Narayanan, D., & Ananth, V. (2018). How the mobile phone is shaping to be BJP’s
most important weapon in elections. Economic Times, August 23. https://
economictimes.indiatimes.com/news/politics-and-nation/how-the-mobile-phone-
is-shaping-to-be-bjps-m... | Social_Media_and_Democracy |
Cohen, G. L., Aronson, J., & Steele, C. M. (2000). When beliefs yield to evidence:
Reducing biased evaluation by affirming the self. Personality and Social Psychology
Bulletin, 26(9), 1151–1164. https://doi.org/10.1177/01461672002611011
Cohen, G. L., Sherman, D. K., Bastardi, A., Hsu, L., McGoey, M., & Ross, L. (2007).... | Social_Media_and_Democracy |
Yifu Qiu, Varun Embar, Shay B Cohen, and Benjamin
Han. 2023a. Think while you write: Hypothesis
verification promotes faithful knowledge-to-text gen-
eration. arXiv preprint arXiv:2311.09467.
Yifu Qiu, Yftah Ziser, Anna Korhonen, Edoardo M.
Ponti, and Shay B. Cohen. 2023b. Detecting and mit-
igating hallucinations in ... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
Model
PaLM 2-L
gpt-3.5-turbo
Approach
Greedy decoding
USC
Oracle
Greedy decoding
USC
Oracle
GPT-judge GPT-info
62.1
67.7
93.8
79.8
82.5
94.9
95.1
99.0
100.0
99.7
99.6
100.0
14
Universal Self-Consistency for Large Language Model Generation
I have generated the following responses to the question: The three-dig... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
Audio Diffusion Model. To enable flexible cross-modality attention in joint generation, the audio
diffuser is designed to have a similar architecture to vision diffusers, where the mel-spectrogram
can be naturally viewed as an image with 1 channel. We use a VAE encoder to encode the mel-
spectrogram of audio to a compre... | Any-to-Any Generation via Composable Diffusion |
observation is that existing frozen LM methods are so compact that there is room to expand them
significantly while still paying a negligible price relative to the single pass through the huge LM.
We focus on two settings in which the go-to standard is still fine-tuned models. The first, already
discussed above, is massiv... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
[44] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang,
Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-
grained text to image generation with attentional generative
In Proceedings of the IEEE Confer-
adversarial networks.
ence on Computer Vision and Pattern Recognition (CVPR),
June 2018. 2
[45] Han Yi, Zhedong... | Instant3D |
3.3 Calibration of Preference Models and Implications for RL
Preference model scores should predict the probability that humans will prefer one or another model-
generated response. We are interested in whether these probabilities are accurate, i.e. whether the PMs
12We found that our RLHF models gave more preferable... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
English? Preprint arXiv:2305.07759, 2023.
[17] Y. Fu, H. Peng, L. Ou, A. Sabharwal, and T. Khot. Specializing Smaller Language Models
towards Multi-Step Reasoning. In International Conference on Machine Learning, 2023.
[18] Y. Fu, H. Peng, A. Sabharwal, P. Clark, and T. Khot. Complexity-Based Prompting for Multi-
s... | METAMATH |
[262] Serkan Kiranyaz, Turker Ince, Ridha Hamila, and Moncef Gabbouj. 2015. Convolutional neural networks for patient-
specific ECG classification. In 2015 37th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC). IEEE, 2608–2611.
[263] Yuma Koizumi, Kohei Yatabe, Marc Delcro... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost.
In International Conference on Machine Learning, pages 4596–4604. PMLR, 2018.
Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton,
and Jeff Dean. Outrageously large neural networks: The... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
C(r) =
T (t)σ(r(t))c(r(t), d) dt,
(1)
(cid:90) tf
tn
tn
T (t) = exp(−(cid:82) t
where r(t) = o + td represents the 3D coordinates of
sampled points on the camera ray emitted from the camera
center o with the direction d. tn and tf indicate the near
and far sampling bounds. (c, σ) = fθ (r(t)) are the pre-
dicted ... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
repositories to enable easier ways of accessing and exploring collections. Although this is often considered the end goal
of many digitization projects, it is important to emphasize that the existence of these collections is only the beginning
and necessary prerequisite for applying advanced computational methods and o... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
AI Assistant-User Role Assignment. After the task specification, The AI assistant role and the AI
user role will be assigned to the user agent and the assistant agent correspondingly to complete the
specified task. In practice, a system message is passed to each agent declaring roles to each. We refer
to the assistant sy... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
11
THE NEXT DECADE IN AI / GARY MARCUS
issues, failing to generalize abstract patterns to novel words in various ways. Bengio
made limits on the abilities of extant neural networks central at his recent NeurIPS talk
(Bengio, 2019). Within canonical neural network architectures), non-uniform ex... | The Next Decade in AI- |
deletion error rate (DER) is comparable for both large-v2 and distil-large-v2, performing to within
0.3% DER. However, the substitution error rate (SER) is 1.4% higher for distil-large-v2, indicating
that the distilled models are subject to more substitution errors. Overall, the reduction in IER out-
weighs the increas... | DISTIL-WHISPER |
of the LLMs.
Bias. To study the sentiment in model generations that may vary with demographic attributes, we choose
BOLD (Dhamala et al., 2021), a large-scale bias benchmark that comprises 23,679 English Wikipedia prompts
spanning five domains of race, gender, religion, political ideology, and profession, with 43 diffe... | Llama2 |
In Interspeech. 3780–3784.
[540] Efthymios Tzinis, Yossi Adi, Vamsi K Ithapu, Buye Xu, and Anurag Kumar. 2022. Continual self-training with
bootstrapped remixing for speech enhancement. In ICASSP 2022-2022 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP). IEEE, 6947–6951.
[541] Efthym... | AReviewofDeepLearningTechniquesforSpeechProcessing |
6 LIMITATIONS AND FUTURE WORK | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
sentences were gathered using the prompt "put red at G9 now" and are widely employed in
research related to audio-visual speech separation and talking face synthesis. The dataset is
considered to be of exceptional quality and is highly sought after in the scientific community. | AReviewofDeepLearningTechniquesforSpeechProcessing |
as temporally aligned video and audio. Highly customizable and flexible, CoDi
achieves strong joint-modality generation quality, and outperforms or is on par
with the unimodal state-of-the-art for single-modality synthesis. The project page
with demonstrations and code is at https://codi-gen.github.io/ | Any-to-Any Generation via Composable Diffusion |
TasNet v2 [352] employs a convolutional neural network (CNN) to process the input signal
and generate a time-frequency mask for each source. The model is trained using an invariant
permutation training (PIT) method [265], which enables it to separate multiple sources accurately.
TasNet v2 achieves state-of-the-art perf... | AReviewofDeepLearningTechniquesforSpeechProcessing |
The M2UGen model adopts the adapter training strategy,
implementing a three-step training regimen. In the first
phase, all parameters, with the exception of those asso-
ciated with the Multi-modal Understanding Adapters, un-
dergo freezing. The training dataset is configured to incor-
porate the MUCaps dataset for musi... | M2UGen |
Marino, K., Rastegari, M., Farhadi, A., and Mottaghi, R. Ok-
vqa: A visual question answering benchmark requiring
external knowledge. In Conference on Computer Vision
and Pattern Recognition (CVPR), 2019.
Nair, S., Mitchell, E., Chen, K., Savarese, S., Finn, C.,
et al. Learning language-conditioned robot behavior from... | PaLM-E- An Embodied Multimodal Language Model |
Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021.
[7] D. McDermott, M. Ghallab, A. Howe, C. Knoblock, A. Ram, M. Veloso, D. Weld, and
D. Wilkins. Pddl-the planning domain definition language. 1998.
[8] P. Haslum, N. Lipovetzky, D. Magazzeni, and C. Muise. An introduction to the ... | LLM+P- Empowering Large Language Models with Optimal Planning Proficiency |
Pretrained
MPT
Falcon
Llama 1
Llama 2
Fine-tuned
ChatGPT
MPT-instruct
Falcon-instruct
Asian Mexican Muslim Physical
disability
Jewish Middle
Eastern Chinese Mental
disability Latino Native
American Women Black LGBTQ
7B
30B
7B
40B
7B
13B
33B
65B
7B
13B
34B
70B
15.40
15.74
9.06
19.59
16.65
18.80
16.87
14.27
16... | Llama2 |
2 DEFINITIONS
In the general context outside of NLP, hallucination is a psychological term referring to a particular
type of perception [51, 118]. Blom [14] define hallucination as “a percept, experienced by a wak-
ing individual, in the absence of an appropriate stimulus from the extracorporeal world”.
Simply put, a h... | SurveyofHallucinationinNatural Language Generation |
∂
∂n2
n2H2 + s2
(n2 + s2)2 =
2∆2
H2(s2 − n2) − 2s2
2∆2
(n2 + s2)3
n2 > s2 − 2s2
2∆2
H2
.
This inequality holds in this case since 2∆2
H2
decreasing in the number of samples n2.
Thus, any domain weights that reallocate the examples from domain 3 to domains 1 and 2 reduces
the parameter error for all domains.
< ... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
low-
rank optimization, and a parameterized hypercomplex mul-
tiplication (PHM) layer [95]. It follows a similar structure
to adapters, consisting of a down-projection, a nonlinear
activation function, and an up-projection. However, Compacter
replaces the down-projection and up-projection in the adapters
with the low-r... | Parameter-EfficientFine-TuningMethods |
RGB color space and can be used to differentiate a proper
“left” and “right” perspective of the same image space in
3D. | LDM3D- Latent Diffusion Model for 3D |
the correlation between specific image features and memo-
rability. Their results indicated that simple image features do
not correlate strongly with memorability and that content has
a significant impact on memorability, with photos of people
being more memorable than photos of landscapes. Follow-
ing their work, other ... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
[21] Wen-Yi Hsiao, Jen-Yu Liu, Yin-Cheng Yeh, and Yi-Hsuan
Yang. Compound word transformer: Learning to compose
full-song music over dynamic directed hypergraphs. In AAAI,
2021.
[22] Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit,
Ian Simon, Curtis Hawthorne, Noam Shazeer, Andrew M
Dai, Matthew D Hoffman, Monic... | VideoBackgroundMusicGeneration |
11
Appendix
A. Additional Details
A.1. Implementation Details
We operate over the official Stable Diffusion v1.4 text-
to-image model that uses the pretrained text encoder from
the CLIP ViT-L/14 model [24].
Input Representation. Our timesteps t range from 0 to
1,000, as in the standard Stable Diffusion training sch... | A Neural Space-Time Representation for Text-to-Image Personalization |
Amendment of Section 230
273
within the CDA 230 framework (Reidenberg et al. 2012, pp. 35–37). On one
hand, liability for the acts of any one of a large pool of users may threaten the
financial viability of certain platforms Reidenberg et al. 2012, p. 36). This is
particularly the case given the broad scope of “intera... | Social_Media_and_Democracy |
Another method to mitigate the impact of noisy datasets is tilted empirical risk minimization
(TERM), a training objective proposed by Li et al. [107]. [95] mentions that techniques such as
dropout, L2E regularization, and clipping tend to decrease the number of hallucinations. Lastly,
several authors propose methods o... | SurveyofHallucinationinNatural Language Generation |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.