text
stringlengths 1
1k
⌀ | title
stringlengths 2
143
|
|---|---|
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, potentially allowing us to incorporate new
skeleton formats by expanding the decoder, without a need
for model specific fine-tuning or probing.
5.3. Comparison to Prior Works | 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. For example, consider the PC shown in Fig. 1(a). Since c2 has an exponentially
larger support than c1, it should be assumed as prior that θ12 will be much larger than θ11.
We highlight the significance of the above issue by examining the fraction of sum units with imbal-
anced child support sizes in PCs learned by Strudel, a state-of-the-art structure learning algorithm
for deterministic PCs [5]. We examine 20 PCs learned from the 20 density estimation benchmarks
[39], respectively. All sum units with ≥ 3 children and with a support size ≥ 128 are recorded. We
measure “imbalanceness” of a sum unit n by the fraction of the maximum and minimum support size | 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: Develop clear and transparent decision-making
protocols that outline the steps involved in making decisions. These protocols should
be based on principles such as fairness, transparency, and accountability.
2. Oversight and accountability: Establish oversight and accountability mechanisms to
ensure that AGI’s decision-making processes are in line with the established.
... (terminated due to the max length of token limit).
3.
34 | 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. (2021). Webgpt: Browser-assisted question-answering with human feedback.
CoRR, abs/2112.09332.
[Nalisnick et al., 2019] Nalisnick, E., Matsukawa, A., Teh, Y. W., Gorur, D., and Lakshminarayanan, B.
(2019). Hybrid models with deep and invertible features.
[Nguyen et al., 2014] Nguyen, A., Yosinski, J., and Clune, J. (2014). Deep neural networks are easily fooled:
High confidence predictions for unrecognizable images. | 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 specified. Losses on
SMPL-X parameters are applied only on the pose data
[9, 23, 33]. For more implementation details, see Sup. Mat. | 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 target network fθt
is of the same
architecture as the student and is updated by EMA with ξ controlling to what degree the
target network preserves its history as in
with initialization η = θs.
θt ← ξθt + (1 − ξ)θs
10
The infoNCE Offsprings
• He et al. [2020a, MoCo] introduces momentum encoder as an alternative to the memory bank regular-
ization of eq. (5) and introduces a queue to store many negative samples from previous batches; [Chen
et al., 2020d, MoCoV2] adds a projector, [Chen et al., 2021b, MoCoV3] adds ViTs
• Chen et al. [2020b, SimCLR] removes the momentum encoder and the ith term from the denominator | 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). Setúbal: SciTePress.
(2015). New classification models for detecting Hate and Violence web content. In
Proceedings of the 7th International Joint Conference on Knowledge Discovery,
Knowledge Engineering and Knowledge Management (IC3K 2015), Vol. 1. (pp.
487–495). Lisbon: IEEE.
Lizza, R. (2016). Twitter’s anti-Semitism problem. The New Yorker, October 19. www
.newyorker.com/news/news-desk/twitters-anti-semitism-problem
Magdy, W., Darwish, K., Abokhodair, N., Rahimi, A., & Baldwin, T. (2016). #
isisisnotislam or # deportallmuslims? Predicting unspoken views. In Proceedings
of the 8th ACM Conference on Web Science (pp. 95–106). New York: ACM. | 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/abs/2112.
09332.
Ani Nenkova and Kathleen McKeown. A survey of text summarization techniques. Mining text data, pp.
43–76, 2012.
Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber,
David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for
intermediate computation with language models. ArXiv preprint, abs/2112.00114, 2021. URL https:
//arxiv.org/abs/2112.00114.
OpenAI. OpenAI: Introducing ChatGPT, 2022. URL https://openai.com/blog/chatgpt.
OpenAI. Gpt-4 technical report, 2023.
Guy A Orban and Fausto Caruana. The neural basis of human tool use. Frontiers in psychology, 5:310, 2014. | 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, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain,
Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with
reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022.
[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna
Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from
ai feedback. arXiv preprint arXiv:2212.08073, 2022. | 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 prompting and self-consistency (Wang et al., 2022) compared to
the previous best accuracy of 92% with the same prompting technique. Similar positive trends are
observed in increased difficulty math problems drawn from middle- and high-school math competitions
(MATH benchmark), with the Gemini Ultra model outperforming all competitor models, reaching
53.2% using 4-shot prompting. The model also outperforms the state of the art on even harder tasks
derived from American Mathematical Competitions (150 questions from 2022 and 2023). Smaller
models perform poorly on this challenging task scoring close to random, but Gemini Ultra can solve | 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 Text-to-speech (TTS) pipeline: a diagram showing the main modules of a typical TTS system.
The system takes text input and processes it through various stages to generate speech output. The text
analysis module tokenizes the input text and generates linguistic features such as phonemes and prosody. The
acoustic model module then converts these linguistic features into acoustic features, such as mel spectrograms,
using a neural network. Finally, the waveform generation module synthesizes the speech waveform from the
acoustic features using another neural network. | 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
i
n
a
p
p
r
o
p
r
i
a
t
e
r
e
s
p
o
n
s
e
s
t
o
o
t
h
e
r
s
,
a
n
d
(
e
)
v
u
l
n
e
r
a
b
i
l
i
t
y
t
o
a
d
v
e
r
s
a
r
i
a
l
p
r
o
m
p
t
i
n
g
:
u
s
e
r
s
w
i
l
l
fi
n
d
w
a
y
s
t
o
s
t
r
e
s
s
t
e
s
t
B
a
r
d
f
u
rt
h
e
r
.
T
h
e
s
e
a
r
e
a
r
e
a
s
o
f
r
e
s
e
a
r
c
h
t
h
a
t
w
e
a
n
d
t
h
e
b
r
o
a
d
e
r
fi
e
l
d
a
r
e
l
o
o
k
i
n
g
t
o
a
d
d
r
e
s
s
,
a
n
d
w
e
a
t
G
o
o
g
l
e
a
r
e
c
o
m
m
i
tt
e
d
t
o
w
o
r
k
i
n
g
t
o
i
m
p
r
o
v
e
t
h
e
m
o
v
e
r
t
i
m
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 through feedback signals from
the LLM. (This section will be presented in 4.4)
4.2 How to Match the Semantic Space of Queries
and Documents
is
to rewrite the query.
intuitive way to align the | 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 capabilities of large language models show promise toward building generalist
agents that can tackle more complex multi-step problems. The AlphaCode team built AlphaCode
2 (Leblond et al, 2023), a new Gemini-powered agent, that combines Gemini’s reasoning capabilities
with search and tool-use to excel at solving competitive programming problems. AlphaCode 2
ranks within the top 15% of entrants on the Codeforces competitive programming platform, a large
improvement over its state-of-the-art predecessor in the top 50% (Li et al., 2022).
2
Gemini: A Family of Highly Capable Multimodal Models | 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 instances of some class. It also provides a clear
basis for representing structured representations (such as the tree structures that are
11 Seeking mappings between implementational details and algorithmic description, if they exist, may | 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.ucl.ac.uk/prospective-students/graduate/writing-your-personal-statement
Cookie settings
1/3
25/06/2023, 09:48
Writing your personal statement | Prospective Students Graduate - UCL – University College London
Some programmes ask for programme-specific information in your personal statement. Check
your programme in the Graduate Prospectus for details.
Some of these cookies are essential, while others help us to improve your
experience of our website. Find out more:
Privacy Policy (opens in new window)
Register interest in your chosen subject to receive notice of graduate open days, events
and more. | 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]. Similar to their findings,
our results show that model scaling improves our quality (sensibleness, specificity, and interestingness), safety and
groundedness metrics to some extent. However, fine-tuning combined with scaling significantly improves performance
on all metrics.
Our work is also closely related to recent successes in applying language models to dialog modeling (e.g., [25, 26,
17, 18]), which built on earlier research in neural dialog modeling (e.g., [14, 15, 16, 27, 28]). One of our fine-tuning
stages requires training on dialog-only data, which is related to Wolf et al. [29], Dinan et al. [25] and Zhang et al. [30]. | 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 {Occupation} is a
Assistant: Sure thing, here is your sentence with the missing word filled in:
15 | 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 audiovisual speech separation surpasses
audio-only approaches by leveraging visual cues from a speaker’s face to resolve ambiguity
in speech signals. By integrating visual information, the model’s ability to disentangle over-
lapping speech signals is enhanced. The implications of automatic speech separation extend
across diverse applications, including assistive technologies for individuals with hearing
impairments and head-mounted devices designed to facilitate effective communication in
noisy meeting scenarios.
• Talking face generation: Generating a realistic talking face of a target character, synchronized | 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),
EU, 199, 208, 317–318
Gentzkow, Matthew, 18, 39, 44
Germany
content takedown transparency rules, 230
media regulation in, 201–207
newspaper markets, 204
Gillespie, Tarleton, 238
Gladwell, Malcolm, 41
Glaser, J., 180
global considerations
censorship danger of content moderation, 73
hate speech detection in languages other than
English, 61
hate speech legal definitions, 58
political bot usage, 93–94
political polarization, 49
scope of misinformation and, 25–26
social media as transnational communication
Global Network Initiative (GNI), 229–230,
mechanism, 99
295–296, 299
global South, misinformation effects in, 25–26
Goel, Sharad, 39
Google | 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 was so happy. She thanked the ranger for helping her and said, ”I’m glad I could help.”
The ranger smiled and said, ”You’re welcome, Lucy. I’m glad I could help.” | 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. Weinberger, editors, Advances in Neural Information Processing Systems, volume 26. Curran
Associates, Inc., 2013. URL https://proceedings.neurips.cc/paper_files/paper/
2013/file/c058f544c737782deacefa532d9add4c-Paper.pdf. | 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] . Although they have
achieved promising results, typical limitations still exist
when using a single 3D representation: silhouette-based
methods like [2] is suffer from lack of details and view
inconsistency, template-based deformation methods [3], [4]
are unable to handle loose clothes, depth-based methods [6],
[7] cannot handle self-occlusions naturally, and volumetric
methods [1], [5] cannot recover high-frequency details due
to their cubically growing memory consumption. | 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, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video
generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3
[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural
Information Processing Systems, 33:6840–6851, 2020. 2, 4
[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet.
Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3, 15
[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for | 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 Undergraduate Personal Statements
General
1. Should the UCAS Similarity Detection Service alert UCL to possible plagiarism in an application, | 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 must hold that Eo∼F|ai
ψ(ai) ≥ Eo∼F|ai−1
[b1(o)]−ψ(ai−1). That is, (γq−i−γq−(i−1))(b1(o2)−b1(o1)) ≥ γ1−i(1−γ)−1+(γ+(cid:15)).
Since b1(o1) must be at least q, it follows from the last inequality that b1(o2) ≥ q + γ1−q − γi−q +
(cid:15)/(γq−i − γq−(i−1)). Using this, the following chain of inequalities shows that the principal’s ex-
pected utility is at most 1 − (cid:15)/(1 − γ):
[b1(o)]−
Eo∼Fai
[v1(o) − t1(b, o)]
= γq−i(v1(o2) − b1(o2)) + (1 − γq−i)(v1(o1) − b1(o1))
≤ γq−i(v1(o2) − q − γ1−q + γi−q − (cid:15)/(γq−i − γq−(i−1))) + (1 − γq−i)(v1(o1) − q)
= γq−i(γi−q − (cid:15)/(γq−i − γq−(i−1)))
= 1 − (cid:15)/(1 − γ).
A.3 Existence of IIVCG for Examples 1 and 2 | 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 foundation.
While it’s easier for smaller markets to experience
faster growth, at 117% YoY increased adoption, the data
integration market is growing substantially faster than BI.
This trend dovetails with the rapid growth of ML adoption
we see across the Lakehouse, covered in the DS/ML
section of the report.
19
2023 STATE OF DATA + AI Views from
the Lakehouse
MIGRATION AND DATA
FORMAT TRENDS | 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.
[100] Yi Chern Tan and L. Elisa Celis. Assessing social and intersectional biases in contextualized word representations.
arXiv preprint arXiv:1911.01485, 2019.
[101] Nithya Sambasivan, Erin Arnesen, Ben Hutchinson, Tulsee Doshi, and Vinodkumar Prabhakaran. Re-imagining
algorithmic fairness in india and beyond. arXiv preprint arXiv:2101.09995, 2021.
[102] Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. Adversarial
training for large neural language models. arXiv preprint arXiv:2004.08994, 2020.
[103] Joseph Weizenbaum. Computer Power and Human Reason: From Judgment to Calculation. W. H. Freeman & | 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 been found to be emergent by Wei
et al. (2022b), while the rest of the tasks are those
which have been found to be emergent in prior lit-
erature (Wei et al., 2022b). It should be noted that
the instruction-tuned GPT models exhibit perfor-
mance that could lead to the appearance of emer-
gence in the exact match, few-shot setting, on all
of those tasks that have previously been found to
be emergent, as illustrated in Figure 3. This out-
come enables us to validate our experimental con-
figuration and establish a baseline for comparison.
These models also exhibit the lack of emergence
on tasks which have previously been found not to | 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 at least one of these three key areas:
quantitative User studies, statistical modelling of human experimental
data, intelligent systems/AI, Human augmentation
very good in scientific writing and communication skills
https://www.aalto.fi/en/open-positions/doctoral-researcher-position-in-human-computer-interaction-human-ai-interaction
2/6
21/11/2023, 04:56 | 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,pages770–778,2016.3[23]MartinHeusel,HubertRamsauer,ThomasUnterthiner,BernhardNessler,andSeppHochreiter.Ganstrainedbyatwotime-scaleupdateruleconvergetoalocalnashequilib-rium.Advancesinneuralinformationprocessingsystems,30,2017.5,8[24]JonathanHo,WilliamChan,ChitwanSaharia,JayWhang,RuiqiGao,AlexeyGritsenko,DiederikPKingma,BenPoole,MohammadNorouzi,DavidJFleet,etal.Imagenvideo:Highdefinitionvideogenerationwithdiffusionmod-els.arXivpreprintarXiv:2210.02303,2022.2,3,5[25]JonathanHo,AjayJain,andPieterAbbeel.Denoisingdiffu-sionprobabilisticmodels.AdvancesinNeuralInformationProcessingSystems,33:6840–6851,2020.3[26]JonathanHoandTimSalimans.Classifier-freediffusionguidance.InNeurIPS2021Workshopo | 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 Facebook pages with three daily posts, to the
maintenance of ten Twitter accounts with at least fifty daily tweets (Dawson
and Innes 2019). Workers also were reportedly given daily topics to focus on
and keywords to include. The IRA reportedly experienced high worker
turnover. While driven by Russian interests at the organizational
level,
individual workers were probably not typically ideologically invested in the
work (Koreneva 2015).
Owing to congressional | 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 al., 2022). Unlike their predecessors, VLMs
are capable of simultaneously understanding both images
and text, and can be applied to tasks such as visual ques-
tion answering (Zhou et al., 2020; Zellers et al., 2021b),
captioning (Hu et al., 2022), optical character recognition
(Li et al., 2021), and object detection (Chen et al., 2021b).
The methods by which images are integrated varies. For ex-
ample, Alayrac et al. (2022) augments pretrained language
models with a mechanism to directly attend to a single con-
text image. In contrast, PaLM-E represents images and
text as “multimodal sentences” of latent vectors, allowing | 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 of meal plan and exercise routine, and stick with it for a few months
to get lasting results. The key is to adopt healthier habits and make them a lifelong lifestyle
change, not focusing on rapid weight loss which is generally not sustainable. I can provide
some general weight loss tips if you want, but I really think a slow and steady approach is
the best way to lose weight and keep it off.
LaMDA Prompt → Please describe what the following code does and (if applicable how it works): | 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] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu,
and Mark Chen. Hierarchical text-conditional image gen-
eration with clip latents. arXiv preprint arXiv:2204.06125,
2022. 1, 6, 13
[61] Ren´e Ranftl, Katrin Lasinger, David Hafner, Konrad
Schindler, and Vladlen Koltun. Towards robust monocular
depth estimation: Mixing datasets for zero-shot cross-dataset
transfer. TPAMI, 2020. 13
[62] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San-
jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy,
Aditya Khosla, Michael Bernstein, Alexander C. Berg, and
Li Fei-Fei. ImageNet Large Scale Visual Recognition Chal-
lenge. IJCV, 2015. 4 | 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 Python
problems with text descriptions, where each problem has 3 unit tests. We follow prior work [49, 37]
in including the first unit test in the prompt as part of the problem description, and keeping the
remaining 2 unit tests hidden for full evaluation. Similar to code translation, we can also utilize the
unit test execution results in the feedback message, but the main difference is that the model still
needs to infer the code correctness even if the predicted Python code passes the given unit test. We
utilize the same prompt as in [37] to obtain the initial Python code, and we present the full prompts
for SELF-DEBUGGING in Appendix C. | 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 hypothesize that using perceptually
weighted loss functions instead of L2 loss during
diffusion could help this trade-off, giving a more
balanced importance to high frequency sounds even
at high compression ratios.
D.6 High-Frequency Audio Generation
We have encountered challenges in achieving satis-
factory results when dealing with high-frequency
audio signals, as detailed in Appendix D.1. To
gain deeper insights into the underlying issues, we
conducted an ablation experiment by exclusively
training our model on classical music, a genre
known for its prominent high-frequency charac-
teristics. We train this model using 500 hours of | 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 limited to certain identities. However, recent advancements in deep
generative models have made significant progress. For example, DAVS [671] introduced an end-to-
end trainable deep neural network capable of learning a joint audiovisual representation, which uses
adversarial training to disentangle the latent space. Another architecture proposed by ATVGnet
[65] consists of an audio transformation network (AT-net) and a visual generation network (VG-net)
for processing acoustic and visual information, respectively. This method introduced a regression-
based discriminator, a dynamically adjustable pixel-wise loss, and an attention mechanism. In | 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 values and preferences, agents actively engage with humans, allowing
them to rectify some misunderstandings and assimilate this tailored feedback into their planning
methodology [108; 189; 190]. Furthermore, they could draw feedback from tangible or virtual
surroundings, such as cues from task accomplishments or post-action observations, aiding them in
revising and refining their plans [91; 101; 187; 191; 260].
15
3.1.5 Transferability and Generalization | 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 that is available through approved third-
party vendors who also serve ads on Google; and, in 2019, the content of third-
party ads was often unavailable, inhibiting content comparisons. As of late
2019, Facebook’s library and researcher API still did not allow searching for
particular time windows, meaning that, to access 2018 election content, one
may have to page through thousands of ads placed in 2019 to access the desired
advertising. More problematic is that researcher calls to the API are rate limited | 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. Jermaine. Neural sketch learning for conditional program
generation. arXiv preprint arXiv:1703.05698, 2017.
Y. Nandwani, D. Jindal, Mausam, and P. Singla. Neural learning of one-of-many solutions for combi-
natorial problems in structured output spaces. In International Conference on Learning Representations,
2021.
R. Y. Pang and H. He.
arXiv:2009.07839, 2020.
H. Pearce, B. Ahmad, B. Tan, B. Dolan-Gavitt, and R. Karri. An empirical cybersecurity evaluation
of GitHub Copilot’s code contributions. CoRR, abs/2108.09293, 2021. URL https://arxiv.org/
abs/2108.09293.
R. Puri, D. S. Kung, G. Janssen, W. Zhang, G. Domeniconi, V. Zolotov, J. Dolby, J. Chen, M. Choudhury, | 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 Zhou, Shujie Liu, Zhuo Chen, Yanqing
Liu, Huaming Wang, Jinyu Li, et al. Neural codec language models are zero-shot text to speech
synthesizers. arXiv preprint arXiv:2301.02111, 2023.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need.
In I. Guyon, U. Von
Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, edi-
tors, Advances in Neural Information Processing Systems, volume 30. Curran Associates,
Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/2017/file/
3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf.
11 | 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 without the MUSE
loss. It is important to note that the same encoder E is shared between source and target languages.
Furthermore, the MUSE embedding is multilingual in nature. As a result, the encoder is able to learn
a multilingual embedding space across source and target languages. This allows a more efficient
and effective encoding of the input, as the encoder is able to encode speech in both languages into a
common embedding space, rather than maintaining a separate embedding space for each language. | 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-
naturalinstructions: Generalization via declarative instructions on 1600+ nlp
tasks. In: Proceedings of the 2022 Conference on Empirical Methods in Natural
Language Processing, pp. 5085–5109 (2022)
[15] Ziegler, D.M., Stiennon, N., Wu, J., Brown, T.B., Radford, A., Amodei,
D., Christiano, P., Irving, G.: Fine-Tuning Language Models from Human
Preferences (2020)
[16] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P.,
Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F.,
Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., Lowe,
R.: Training language models to follow instructions with human feedback (2022) | 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 make a new snowman for him,” Sara says. ”Maybe,” Ben says.
They hug the dog and give him a name. The dog wags his tail and licks their faces. ”Maybe he likes us,” Sara says. ”Maybe
he likes us,” Ben says. The dog wags his tail and licks their faces. He is happy and friendly. | 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
||xl
hmi = Wf [xl
(2)
where Wf is of shape (dent, 2 · demb), where demb
is the dimension of X1.
smi
tmi
]
We find the k nearest entity embeddings of hmi
from E by computing the dot product, and Emi is
a weighted sum of them. More formally:
(cid:88)
αj · (EntEmbed(ej))
exp(EntEmbed(ej) · hmi)
(cid:80)
e∈topK(E,hmi ,k) exp(EntEmbed(e) · hmi)
ej∈topK(E,hmi ,k)
Emi =
αj = | 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 (Rajpurkar et al., 2018), and LAMBADA (Paperno et al., 2016). We use
the standard exact-match accuracy metric for the these datasets.
6
Table 1: Domain weights on The Pile. Baseline domain weights are computed from the default Pile
dataset. DoReMi (280M) uses a 280M proxy model to optimize the domain weights.
Pile-CC
PubMed Central
Books3
OpenWebText2
ArXiv
Github
FreeLaw
StackExchange
USPTO Backgrounds
PubMed Abstracts
Gutenberg (PG-19)
Baseline DoReMi (280M)
0.6057
0.0046
0.0224
0.1019
0.0036
0.0179
0.0043
0.0153
0.0036
0.0113
0.0072
0.1121
0.1071
0.0676
0.1247
0.1052
0.0427
0.0386
0.0929
0.0420
0.0845
0.0199 | 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-only
Pre-training
Objective
span corruption
span corruption
span corruption
span corruption
span corruption
Pre-train
FLOPs
1.8E+20
6.6E+20
2.3E+21
9.0E+21
3.3E+22
3.7E+22
2.9E+23
2.5E+24
4.8E+23
prefix LM + span corruption 2.5E+23
causal LM
causal LM
causal LM
causal LM
Finetune
FLOPs
2.9E+18
9.1E+18
2.4E+19
5.6E+19
7.6E+19
1.6E+20
1.2E+21
5.6E+21
1.8E+21
5.6E+21
% Finetune
Compute
1.6%
1.4%
1.1%
0.6%
0.2%
0.4%
0.4%
0.2%
0.4%
0.2% | 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 execution. CREA-
ICL [Li et al., 2023b] employs a synchronous retrieval of
cross-lingual knowledge to enhance context, while RE-
CITE [Sun et al., 2022] generates context by sampling para-
graphs directly from LLMs. | 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 in Myanmar and Sri Lanka, the offline consequences
of online hate speech appear increasingly dire. Fearing that this harmful rhetoric
is inciting violence and driving extremism, governments worldwide are passing
regulation and pressuring social media companies to implement policies to stop
the spread of online hate speech (Gagliardone et al. 2016). | 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 fabricated
articles that imitate the appearance of traditional news stories (Allcott and
Gentzkow 2017; Lazer et al. 2018). Fake news is a form of disinformation, as
it is spread despite being known to be false, but it may be distinguished from
other types of disinformation by its unique format – namely, its emulation of
legitimate media outlets (Pennycook and Rand 2018). In addition to fake news,
recent work also looks beyond textual forms of misinformation to other types
of media, including manipulated images and videos (Kasra, Shen, and O’Brien
2016; Schwarz, Newman, and Leach 2016; Shen et al. 2019).
Finally, | 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 the model’s understanding of token relationships but
also facilitates length extrapolation, a critical feature for handling varied and complex sequences in language processing. | 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-tuning, only the parameters of adapter network Wup and
Wdown need to be updated to make the PLMs adapt to the
specific downstream tasks. The specific architecture of the
sequential adapter is presented in Fig. 3.
X = (ReLU(XWdown))Wup + X,
Wdown ∈ Rd×k, Wup ∈ Rk×d.
(6) | 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-GIF: Neural generalized implicit functions
for animating people in clothing. In International Conference
on Computer Vision (ICCV), pages 11708–11718, 2021. 3
[61] Twindom. twindom.com, 2018. 5
[62] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger,
and Siyu Tang. MetaAvatar: Learning animatable clothed
human models from few depth images. In Conference on
Neural Information Processing Systems (NeurIPS), 2021. 3
[63] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica K.
Hodgins. MonoClothCap: Towards temporally coherent cloth-
ing capture from monocular RGB video. In International
Conference on 3D Vision (3DV), pages 322–332, 2020. 3 | 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 language, a
comprehensive list of the agent’s experiences. The retrieval model
combines relevance, recency, and importance to surface the records
that are needed to inform the agent’s moment-to-moment behavior.
The second is reflection, which synthesizes memories into higher-
level inferences over time, enabling the agent to draw conclusions
about itself and others to better guide its behavior. The third is
planning, which translates those conclusions and the current en-
vironment into high-level action plans and then recursively into
detailed behaviors for action and reaction. These reflections and
plans are fed back into the memory stream to influence the agent’s | 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 layer from flash
for each inference query. We also propose a static
memory preallocation to minimize transfers within
DRAM and reduce inference latency. Our load
from flash cost model captures the tradeoff between
loading less data and reading bigger chunks. Op-
timizing this cost model and selectively loading
parameters on demand yields flash loading strate-
gies that can run models 2x larger than the device’s
DRAM capacity and speed up inference by 4-5x
and 20-25x compared to naive implementation in
CPU and GPU, respectively.
2 Flash Memory & LLM Inference | 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 Wang and David MJ Tax. 2016. Survey on the attention based RNN model and its applications in computer
vision. arXiv preprint arXiv:1601.06823 (2016).
[567] Gary Wang. 2019. Deep text-to-speech system with seq2seq model. arXiv preprint arXiv:1903.07398 (2019).
[568] Heming Wang and Deliang Wang. 2020. Time-Frequency Loss for CNN Based Speech Super-Resolution. In ICASSP
2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 861–865. https:
//doi.org/10.1109/ICASSP40776.2020.9053712 | 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 correctness, PandaLM incorporates crucial subjective elements,
including relative conciseness, clarity, adherence to instructions, comprehensiveness, and formality. | 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-most-important-weapon-in-elections/articleshow/65508743.cms
PTI. (2019). 2019 polls: BJP to form chain of WhatsApp groups to strengthen
communication between party workers. Economic Times, December 23. https://
economictimes.indiatimes.com/news/politics-and-nation/2019-polls-bjp-to-form-
chain-of-whatsapp-groups-to-strengthen-communication-between-party-workers/
articleshow/67219816.cms
Stigler Center. (2019). Digital Platforms and Concentration. Stigler Center for the Study
of the Economy and the State. https://promarket.org/wp-content/uploads/2018/04/
Digital-Platforms-and-Concentration.pdf | 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).
Bridging the partisan divide: Self-affirmation reduces ideological closed-mindedness
and inflexibility in negotiation. Journal of Personality and Social Psychology, 93(3),
415–430. https://doi.org/10.1037/0022-3514.93.3.415
Cook, J., & Lewandowsky, S. (2011). The Debunking Handbook. St. Lucia: University
of Queensland. http://sks.to/debunk
Cook, J., Lewandowsky, S., & Ecker, U. K. H. (2017). Neutralizing misinformation
through inoculation: Exposing misleading argumentation techniques reduces their
https://doi.org/10.1371/journal
influence. PLoS ONE,
.pone.0175799
e0175799.
12(5), | 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 multilingual summarisation.
Vipula Rawte, Swagata Chakraborty, Agnibh Pathak,
Anubhav Sarkar, S. M Towhidul Islam Tonmoy,
Aman Chadha, Amit P. Sheth, and Amitava Das.
2023. The troubling emergence of hallucination in
large language models – an extensive definition, quan-
tification, and prescriptive remediations.
Partha Pratim Ray. 2023. Chatgpt: A comprehensive
review on background, applications, key challenges,
bias, ethics, limitations and future scope. Internet of
Things and Cyber-Physical Systems, 3:121–154.
Evgeniia Razumovskaia,
Ivan Vuli´c, | 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-digit number "ab5" is divisible by
3. How many different three-digit numbers can "ab5" represent? | 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 compressed latent space. In audio synthesis, a VAE decoder maps the latent
variable to the mel-spectrogram, and a vocoder generates the audio sample from the mel-spectrogram.
We employ the audio VAE from [33] and the vocoder from [27].
Text Diffusion Model. The VAE of the text LDM is OPTIMUS [29], and its encoder and decoder
are [9] and GPT-2 [39], respectively. For the denoising UNet, unlike the one in image diffusion, the
2D convolution in residual blocks is replaced with 1D convolution [53].
3.4
Joint Multimodal Generation by Latent Alignment | 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 massive multi-tasking: asking a single model to simultaneously address many
NLP tasks. The variety of existing multi-tasked models are all fine tuned; no frozen model method has
been considered in this setting. Our second setting is a challenging individual task, in which leading
methods are all fine tuned (Chen et al., 2017; Lee et al., 2019; Karpukhin et al., 2020; Roberts et al.,
2020): open-domain question answering, asking a model to answer general-knowledge questions.
Open-domain question answering has two popular variants: “open book”, in which the model is given
access to documents retrieved from a predefined corpus (web, books, proprietary corpora) that are | 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 Zheng, Xiangyu Xu, and Tat-seng Chua.
Progressive text-to-3d generation for automatic 3d prototyp-
ing. arXiv preprint arXiv:2309.14600, 2023. 3
[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiao-
gang Wang, Xiaolei Huang, and Dimitris N. Metaxas. Stack-
gan: Text to photo-realistic image synthesis with stacked
generative adversarial networks. In Proceedings of the IEEE
International Conference on Computer Vision (ICCV), Oct
2017. 2
16
Instant3D: Instant Text-to-3D Generation
APPENDIX
1. Daily Life Set
stated in the main paper,
than
As
more
1-5GP4H2QJSPwIuUIfiEBjzPI7DwG9LOX/view?usp=sharing.
our Daily Life prompt
available
prompts.
dataset
17,000
The | 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 responses without top-p sampling, presumably because that’s
how they were trained, so we decided to remove top-p sampling when comparing snapshot Elos, including the context-
distilled models which are the initial snapshots of all RLHF models.
13 | 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-
step Reasoning. In International Conference on Learning Representations, 2023.
[19] J. Gou, B. Yu, S. Maybank, and D. Tao. Knowledge Distillation: A Survey. International
Journal of Computer Vision, 2021.
[20] T. He, C. Shen, Z. Tian, D. Gong, C. Sun, and Y. Yan. Knowledge Adaptation for Efficient
Semantic Segmentation. In Computer Vision and Pattern Recognition, 2019.
[21] D. Hendrycks, C. Burns, S. Kadavath, A. Arora, S. Basart, E. Tang, D. Song, and J. Steinhardt.
Measuring Mathematical Problem Solving With the MATH Dataset. In Neural Information
Processing Systems: Datasets and Benchmarks, 2021. | 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 Delcroix, Yoshiki Masuyama, and Daiki Takeuchi. 2020. Speech enhancement
using self-adaptation and multi-head self-attention. In ICASSP 2020-2020 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP). IEEE, 181–185.
[264] Yuma Koizumi, Heiga Zen, Kohei Yatabe, Nanxin Chen, and Michiel Bacchiani. 2022. SpecGrad: Diffusion Probabilistic
Model based Neural Vocoder with Adaptive Noise Spectral Shaping. arXiv preprint arXiv:2203.16749 (2022). | 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 sparsely-gated mixture-of-experts layer.
arXiv preprint arXiv:1701.06538, 2017.
Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn Koanantakool,
Peter Hawkins, HyoukJoong Lee, Mingsheng Hong, Cliff Young, et al. Mesh-tensorflow: Deep
In Advances in Neural Information Processing Systems, pages
learning for supercomputers.
10414–10423, 2018.
Sam Shleifer, Jason Weston, and Myle Ott. Normformer: Improved transformer pretraining with
extra normalization. arXiv preprint arXiv:2110.09456, 2021. | 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 color and density of the sampled point along the ray.
σ(r(s)) ds) is the accumulated transmit-
tance. Different from NeRF that takes both the 3D coordinate
r(t) and view direction d in Eq. 1 to predict the radiance
c(r(t), d), we omit d to avoid the effect of view-dependent
specularity. Additionally, inspired by [17], we introduce the
depth constraint into NeRF optimization to achieve depth-
aware NeRF optimization and speed up model convergence.
To this end, the predicted depth value z(r) is required to be
calculated:
z(r) =
T (t)σ(r(t))t dt.
(2)
(cid:90) tf
tn
i , DR
i and DR | 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 opening new research perspectives.
Figure 1 illustrates the process from digitization to quantitative analysis, knowledge discovery and visualization using
computational methods. Research results obtained from the final phase of this process are often used to enhance the
functionalities of repositories and online collections by adding advanced ways of content exploration. | 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 system prompt/message by PA and that of the user by PU . The system messages are
passed to the agents before the conversations start to assign agents with corresponding roles. Let F1
and F2 denote two large-scale auto-regressive language models [47]. When the system message is
passed to those models respectively, we obtain A ← FPA
2 which are referred to as the
assistant and user agents respectively. In Figure 1, the AI assistant and the AI user are assigned roles
as Python Programmer and Stock Trader at the beginning of the role-playing session, respectively.
The AI user serves as a task planner, engaging in interactive planning to determine feasible steps for | 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 extension
of broad universals (such as identity) is surprisingly common, and in my view it
remains a central obstacle to progress.
§
In essence, extant neural networks of certain sorts (such as the multilayer perceptrons
trained with backpropagation discussed here) excel at two things: memorizing training
examples, and interpolating within a cloud of points that surround those examples in
some cluster of a hyperdimensional space (which I call generalizing within a training
space), but they generalize poorly outside the training space (in Bengio's phrasing, the
training distribution).
5 | 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 increase to SER, and Distil-Whisper returns the lowest WER of all the models. While
the wav2vec 2.0 model underperforms in its average WER score, we find that it is far less prone to
repetition errors compared to both Whisper and Distil-Whisper. Further work is needed to reduce
repetition errors in Seq2Seq ASR models. | 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 different sub-
groups∗∗∗. We conduct a sentiment analysis using the Valence Aware Dictionary and Sentiment Reasoner
(VADER) (Hutto and Gilbert, 2014) to evaluate the sentiments conveyed by the combination of prompt prefix
and model generation. VADER produces a sentiment score between -1 and 1. A positive (negative) score
indicates a positive (negative) sentiment towards the population mentioned in the prompt, and a score closer
to 0 indicates a neutral sentiment. | 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] Efthymios Tzinis, Yossi Adi, Vamsi K Ithapu, Buye Xu, Paris Smaragdis, and Anurag Kumar. 2022. RemixIT: Continual
self-training of speech enhancement models via bootstrapped remixing. IEEE Journal of Selected Topics in Signal
Processing 16, 6 (2022), 1329–1341.
[542] Panagiotis Tzirakis, Anurag Kumar, and Jacob Donley. 2021. Multi-channel speech enhancement using graph neural
networks. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).
IEEE, 3415–3419. | 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 performance in various speech separation tasks with high
separation accuracy, but its disadvantage is its relatively high computational cost. The variant of
TasNet based on CNNs is proposed in [353]. The model is called Conv-TasNet and can generate a
time-frequency mask for each source to obtain the separated source’s signal. Compared to previous
models, Conv-TasNet has faster processing time but lower accuracy. | 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 music understanding, the
COCO dataset for image comprehension, and the captions
sourced from the MUVideo dataset for video understand-
ing. During this training stage, the Cross Entropy Loss
function is applied to compute the disparity between the
caption generated by the LLaMA 2 model and the target
caption corresponding to the input modality. This process
is illustrated in Figure 5.
In the second training stage, the output projector is trained
to generate conditional embeddings using input captions
processed by the LLaMA 2 model. The LLaMA 2 model
produces specialized audio tokens, denoted as [AUDi]
where i ∈ {1, 2, . . . , K} (with K as a hyperparameter | 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
offline data and crowd-sourced annotation. In Conference
on Robot Learning, pp. 1303–1315. PMLR, 2022.
Nottingham, K., Ammanabrolu, P., Suhr, A., Choi, Y., Ha-
jishirzi, H., Singh, S., and Fox, R. Do embodied agents
dream of pixelated sheep?: Embodied decision making
using language guided world modelling. arXiv preprint
arXiv:2301.12050, 2023.
Piergiovanni, A., Kuo, W., and Angelova, A. Pre-training
image-language transformers for open-vocabulary tasks,
2022.
URL https://arxiv.org/abs/2209.
04372.
Polu, S., Han, J. M., Zheng, K., Baksys, M., Babuschkin, I.,
and Sutskever, I. Formal mathematics statement curricu-
lum learning. arXiv preprint arXiv:2202.01344, 2022. | 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 planning do-
main definition language. Synthesis Lectures on Artificial Intelligence and Machine Learning,
13(2):1–187, 2019.
[9] K. Valmeekam, A. Olmo, S. Sreedharan, and S. Kambhampati. Large language models still
can’t plan (a benchmark for llms on planning and reasoning about change). arXiv preprint
arXiv:2206.10498, 2022.
[10] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan,
P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners. Advances
in neural information processing systems, 33:1877–1901, 2020.
[11] M. Helmert. The fast downward planning system. Journal of Artificial Intelligence Research, | 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
16.53
21.29
16.76
21.29
33.55
31.49
18.30
29.61
30.72
32.03
32.24
31.59
31.15
37.25
29.63
32.90
23.54
19.04
17.34
25.83
26.82
25.18
21.53
21.90
22.63
22.81
23.36
25.91
17.09
21.68
8.29
13.54
16.58
14.72
16.24
14.89
15.74
17.77
14.38
16.92
26.12
26.82
19.40
29.85
26.49
28.54
28.54
23.51
26.87
32.65
27.43
30.60
23.20
30.60
12.99
23.40
22.27
21.11
22.04
22.27
19.95
24.13
19.49
21.35
16.25
13.87
10.07
25.55
17.16
18.76
19.91
17.16
15.79
21.05
18.54
16.93
17.63
24.36
10.26
29.10
19.71
15.71
18.27
18.91
19.55
20.19
17.31
21.47
28.40
16.51
18.03
23.20
28.67
30.42
29.88
28.40
25.03
35.40
26.38
30.42
19.52
32.68
15.34
17.31
21.71
20.52
18.13
19.32
18.92
27.69
18.73
20.12 | 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 hallucination is an unreal perception that feels real. The undesired phenomenon of
“NLG models generating unfaithful or nonsensical text” shares similar characteristics with
such psychological hallucinations – explaining the choice of terminology. Hallucinated text gives
the impression of being fluent and natural despite being unfaithful and nonsensical. It appears
to be grounded in the real context provided, although it is actually hard to specify or verify the
existence of such contexts. Similar to psychological hallucination, which is hard to tell apart from
other “real” perceptions, hallucinated text is also hard to capture at first glance. | 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.
< 1 and s2 = 1. Therefore the parameter error is | 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-rank parameterized hypercomplex multiplication
(LPHM) layer, which is an extension of PHM that incorporates
low-rank optimization. Structurally, PHM layer resembles a
fully connected layer, but with the learned W represented as a
i=1 Ai⊗Bi. Notably,
when the weights of down-projection and up-projection are
calculated as in that of the PHM layer, Ai is a shared parameter
across all adapter layers, while Bi represents adapter-specific
parameters. This kind of adapter is called PHM Adapter.
Similarly, Compacter obtains the weight matrix in each LPHM
layer utilizing the sum of Kronecker products, but Compacter
reparameterizes Bi as the product of two independent ranks | 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 approaches were proposed to improve
memorability prediction by investigating different image fea-
tures [44], [45]. A comprehensive overview of studies related
to image memorability is given in [46]. The adoption of | 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, Monica Dinculescu, and Douglas
Eck. Music transformer: Generating music with long-term
structure. In ICLR, 2019.
[23] Qingqing Huang, Daniel S Park, Tao Wang, Timo I Denk,
Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang,
Jiahui Yu, Christian Frank, et al. Noise2music: Text-
conditioned music generation with diffusion models. arXiv
preprint arXiv:2302.03917, 2023.
[24] Yu-Siang Huang and Yi-Hsuan Yang. Pop music transformer:
Beat-based modeling and generation of expressive pop piano
compositions. In MM, 2020. | 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 scheme.
For choosing the U-Net layers, we follow Voynov et al. [41]
and consider 16 different cross-attention layers. Our posi-
tional encoding maps the pair of scalars (t, ℓ) into a 160-
dimensional vector. Our 160 uniformly-spaced (t, ℓ) an-
chors are defined using 16 U-Net layers and 10 time-based
anchors corresponding to t = 0, 100, 200, . . . , 900. Recall,
that the output of our positional encoding is given by:
where
et,ℓ = E × f (t, ℓ) ∈ R160
...
−
−
−
f (0, 0)
−
f (0, 1)
f (0, 16) −
− f (100, 0) −
− f (100, 1) −
− f (900, 16) −
...
160×2048
E =
, | 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 “interactive computer service”
under legal precedent, covering everything from a small blog or listserv to the
biggest platforms operated by companies like Google and Facebook.50 Larger
platforms will have the financial resources and legal expertise to absorb this
risk, while smaller businesses and services run by volunteers may not be able to
manage litigation based on the acts of their users. One effect may be to further
accelerate and reinforce consolidation to the set of largest companies as less
well-resourced platforms exit the market or merge with better-positioned
competitors.
Ironically,
the platforms. | 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 of improving phrase alignment that are helpful both in increasing
translation accuracy and identifying content that did not appear in the source translation [55, 205,
227]. | SurveyofHallucinationinNatural Language Generation |