text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
5. Experts are not yet able to interpret the
inner workings of LLMs | Eight Things to Know about Large Language Models |
one of the ethnic groups is White Americans.
New York City (NYC) is the largest U.S. city. Atlantic City, New Jersey are also part of the United States with
its capital as Washington, DC and home to White Americans. | Prefix-Tuning |
44
Figure 33 Robustness experiments for RLHF, showing test PM score for all policy sizes and all test PM
sizes, evaluated at 200k train samples. Note that the overall slope increases with respect to test PM size,
suggesting that larger preference models are more robust. (left) Experiments for which the train PM is 52... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
[7] Seeber I, Bittner E, Briggs RO, de Vreede T, de Vreede GJ, Elkins A, et al. Machines as teammates:
Information and Management. 2020;57(2):103174.
A research agenda on AI in team collaboration.
Available from: https://doi.org/10.1016/j.im.2019.103174.
[8] Dellermann D, Ebel P, S¨ollner M, Leimeister JM. Hybrid Int... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
Q: There are 8 loads of laundry that need to be washed. Each load of laundry takes around 45 minutes to complete in the
wash cycle and 1 hour in the dry cycle. How long, in hours, will it take to complete the entire set of laundry?
A: Reasoning process: 1. We need to calculate the time it takes to wash the laundry. 2. ... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
[34] E. Ben Zaken, Y. Goldberg, and S. Ravfogel, “BitFit: Simple
parameter-efficient fine-tuning for transformer-based masked language-
models,” in Proc. Annu. Meeting Assoc. Comput. Linguistics, 2022, pp.
1–9.
[35] N. Lawton, A. Kumar, G. Thattai, A. Galstyan, and G. Ver Steeg,
“Neural architecture search for paramet... | Parameter-EfficientFine-TuningMethods |
Personalization.
In the task of personalization, we are in-
terested in adapting a given model to better capture a given
subject or concept.
In the context of text-to-image syn-
thesis, the personalized model should enable synthesizing
novel images of a specific target concept using a free-form
text prompt.
In [6, 9] i... | A Neural Space-Time Representation for Text-to-Image Personalization |
Figure 4: User Preference. We conduct a perceptual study
with approximately 4000 samples and report how often
participants preferred shapes and images generated by our
method or those generated by EVA3D [23].
Training: We train our generator and discriminators
jointly using the non-saturating GAN objective with R1-
re... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
Our evaluation method for models trained on TinyStories-Instruct also relies on GPT-4. We have implemented a
corresponding evaluation pipeline, with an evaluation dataset containing 50 different sets of instructions to create a
5
story (we verified that these are disjoint from the training set). In the scoring phase... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
aug-
mentation for pre-training a self-regressive language model,
enabling large-scale pre-training from scratch by retrieving
from a massive set of labeled data and significantly reducing
model parameters. RETRO shares the backbone structure
with GPT models and introduces an additional RETRO
encoder to encode features... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical
text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.
[40] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and
Sumit Chopra. Video (language) modeling: a ba... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
A.2 EXPERIMENT DETAILS
A.2.1 EVALUATION
For language generation experiments, we calculate the perplexity, in standard fashion like Radford
et al. (2019), as follows: First, the entire validation set is concatenated using two linebreaks as
separators and encoded using the default HuggingFace tokenizer of each model. N... | GPTQ |
Neg. Log Perp. (↑)
-1.474
-1.452
-1.383
-1.369
-1.369
-1.376
∆
-
-
0.022
0.014
0.014
0.007
Table 16: A dense FFN immediately before or after each sparse layer improves quality. Insert-
ing an extra dense FFN immediately before or after each sparse layer improves quality 2x as much
as placing the dense layers (ran... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
hatred against such group or a member of such group” (Sellars 2016). In the
Uniited Kingdom, it is a criminal offense to incite racial or religious hatred, and
variations on this legislation – while unconstitutional in the United States – exist
in the majority of developed democracies, including Australia, Denmark,
Fra... | Social_Media_and_Democracy |
In machine translation (MT), LLMs can perform competent translation, although the average performance is slightly
worse than some commercial translation tools [45] considering some automatic metrics like BLEU[78]. LLMs are
particularly good at translating some low-resource language texts to English texts, such as in th... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
Sophie continued to experiment with the sphere and the crystal, trying to harness their power
for her own purposes. But as she delved deeper into the mysteries of the object, she began to
realize that she might not be able to control it. The sphere and the crystal were drawing her
in, and she was starting to lose herse... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
answer when given the question along with the candidate answers sampled from the first pass through
the LM.
For the first stage, we trained p1 until convergence. We used batch size of 32 and considered learning
rates in {3 · 10−2, 1 · 10−1, 3 · 10−1, 5 · 10−1} with 0.5% warmup. Notably, Lester et al. (2021)
introduced pr... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
Contributions | Tool Learning with Foundation Models |
The distributed representation of symbols is one of the key technologies in ma-
chine learning systems today, playing a pivotal role in modern natural language
processing. Traditional word embeddings associate a separate vector with each
word. While this approach is simple and leads to good performance, it requires a
l... | MULTI HASH EMBEDDINGS IN SPACY |
panel may interview the applicant and any other persons involved in the case, and shall be
required to ensure equal treatment to the parties concerned, i.e. the applicant and the officer(s)
who took the decision on the application or who were involved in the procedure for handling it.
Such persons shall receive copi... | UCL Academic Manual |
1
INTRODUCTION
The best way these days to optimize performance for a given NLP task is usually to fine tune a
pretrained LM. A side effect of doing so is that performance degrades significantly on other tasks.
Partly in response, considerable recent work has been devoted to fine tuning huge LMs simultaneously
on many (i... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
As a useful point of comparison, several studies have examined the diffusion
of accurate information alongside misinformation, thus examining patterns at
a more generalizable level (while also moving beyond examinations of single
cases of misinformation, such as the Haitian earthquake of 2010 [Oh, Kwon,
and Rao 2010] a... | Social_Media_and_Democracy |
3https://huggingface.co/tiiuae/falcon-40b
Family
GPT
T5
Falcon
LLaMA
Model
GPT-2
GPT-2-XL
GPT-J
davinci
T5-small
T5-large
Falcon-7B
Falcon-40B
LLaMA-7B
LLaMA-13B
LLaMA-30B
Instruction-tuned
GPT-2-IT*
GPT-2-XL-IT*
GPT-JT
text-davinci-001
text-davinci-003
FLAN-T5-small
FLAN-T5-large
Falcon-7B-IT
Falcon-40B-I... | AreEmergentAbilitiesinLarge Language Models just In-Context |
Feedback: As in your explanation, the SQL query returns a table with 1
column, the customer names who have the order status "On Road" and "Shipped
". The question returns 1 column, the customer names who have order status
of both "On Road" and "Shipped". So the SQL prediction above is correct!
CREATE TABLE flight (
fl... | Teaching Large Language Models to Self-Debug |
Andromeda runs deep learning applications through the Cerebras Software Platform (CSoft). For this study,
we write and train models in both Tensorflow and PyTorch (reported results are with PyTorch), and CSoft
compiles and orchestrates running these models on the hardware.
In this process, CSoft automatically
selects th... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
r
s
i
t
y
o
f
M
o
n
t
r
e
a
l
,
V
o
l
1
3
4
1
(
3
)
,
p
p
.
1
.
| Language models can explain neurons in language models |
16
References
[1] S. An, Y. Li, Z. Lin, Q. Liu, B. Chen, Q. Fu, W. Chen, N. Zheng, and J.-G. Lou. Input-tuning:
Adapting unfamiliar inputs to frozen pretrained models. arXiv preprint arXiv:2203.03131,
2022.
[2] A. Askell, Y. Bai, A. Chen, D. Drain, D. Ganguli, T. Henighan, A. Jones, N. Joseph, B. Mann,
N. DasSarma, ... | QLORA |
One may notice, however, for r = 64, this amplification factor is only around 2, meaning that
most directions learned in ∆W with r = 64 are not being amplified by much. This should not
be surprising, and in fact gives evidence (once again) that the intrinsic rank needed to represent
the “task-specific directions” (thus fo... | LORA |
32
32
32
64
64
64
128
256
510
1,030
2,540
4,830
7,120
14,240
33,500
72,300
Table 6. Model sizes in the Pythia suite, number of GPUs used during training, and the total number of GPU hours, calculated via
(iteration time (s) × number of iterations × number of GPUs ÷ 3600 s/hour). All GPUs are A100s with 40GB of memory... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
We would like to acknowledge the contributions of those who helped in preparation of this manuscript.
Experimental planning and strategy: Nolan Dey, Joel Hestness
Model training: Zhiming (Charles) Chen, Hemant Khachane, Ribhu Pathria, Gurpreet Gosal
Dataloader development and dataset preparation: Gurpreet Gosal
Numeric... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
a metric to gauge the alignment between the music and video. This metric
assesses how effectively the emotions elicited by the generated chords match the
(cid:88)n
1
n
emotions expressed in the video.
5.3.2. Metrics for Subjective Evaluation
We conducted a comprehensive Listening test in which participants were
... | Video2Music |
like Social Science One, other
Taken together, we would like to suggest, then, that there are essentially
three paths to ensuring and expanding the continued production of the type of
research featured in this volume. The first is to work with the platforms,
through efforts
research
partnerships, or by directly lobbyin... | Social_Media_and_Democracy |
Scenario:
One scenario in the game could be a maze. Teri starts at the beginning of the maze and needs to reach the end.
Players can use blocks of code to tell Teri how to move around the maze, such as how to turn, how to jump, and
how to avoid obstacles.
For example, players could use a block of code to tell Teri to t... | PaLM 2 Technical Report |
Misinformation and Its Correction
167
whether the information has been proven to be untrue or whether it is merely
unsubstantiated. Second, we noted that definitions of misinformation vary in
their area of focus, particularly whether they emphasize the effects of false
information versus false beliefs. Third, we found... | Social_Media_and_Democracy |
This site uses cookies from Google to deliver and enhance the quality of its services and to analyze tra | Google I_O 2023_ Making AI more helpful for everyone |
The idea of mining tasks from raw pre-training corpora to enhance zero-shot capability was in-
troduced by van de Kar et al. (2022). This approach effectively extracts intrinsic tasks from raw
texts through a handful of regex-based patterns, leading to substantial enhancements in the model’s
zero-shot performance via f... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
3
2
0
2
t
c
O
0
1
]
Y
C
.
s
c
[
1
v
6
5
5
6
0
.
0
1
3
2
:
v
i
X
r
a
Gender, Age, and Technology Education Influence the Adoption and
Appropriation of LLMs
FIONA DRAXLER, LMU Munich, Germany
DANIEL BUSCHEK, University of Bayreuth, Germany
MIKKE TAVAST, Aalto University, Finland
PERTTU HÄMÄLÄINEN, Aa... | Adoptionand AppropriationofLLMs |
20“AGI” is sometimes used as a substitute for some concept of “human-level AI”; in other contexts, it refers
specifically to some concept of human learning ability (see e.g. Selsam (undated), and Arbital here (no author
listed, but I believe it is Yudkowsky)), or some method of creating systems that can perform certain ... | Is Power-Seeking AI an Existential Risk? |
ing the previous best Open-QA system. Increasing the size
of T5 yields consistent improvement, but comes at signif-
icant computational cost (from Base to 11B, the model is
50 times larger, and gains roughly 7 points in accuracy). In
contrast, REALM outperforms the largest T5-11B model
while being 30 times smaller. It ... | REALM |
in Neural Information Processing Systems, 32, 2019.
[281] Serrà, J., D. Surís, M. Miron, et al. Overcoming catastrophic forgetting with hard attention to
the task. In International Conference on Machine Learning. 2018.
[282] Dosovitskiy, A., L. Beyer, A. Kolesnikov, et al. An image is worth 16x16 words: Transformers... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
advantage of enriching demand data with information regarding ex-
ogenous factors that can influence demand, achieving more accurate
predictions.Inparticular,authorsaddressingdemandforecastinginthe
automotiveindustry,reportedusingdataregardingtheunemployment
rate [13], people’s personal income [14], the inflation rate ... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
[68] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman.
In Thirty-
Volume rendering of neural implicit surfaces.
Fifth Conference on Neural Information Processing Systems,
pages 4805–4815, 2021. 2
[69] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sat-
tler, and Andreas Geiger. MonoSDF: Exploring monocu-
lar ge... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
sparse activation, sparse gating). During training, multiple individual expert LLMs
and a routing function are trained simultaneously. The learned routing function allows
the MoE system to select a subset of experts according to the input, thus reducing
computational and memory requirements. Switch Transformer [62] fol... | Beyond Efficiency |
Limitations and Future Work
Data Scale Enhancing the scale of both data and
the model holds promising potential for yielding
significant improvements in quality. Following
(Dhariwal et al., 2020; Borsos et al., 2022), we
suggest training with 50K-100K hours instead of
2.5K. Computer Vision studies like Saharia et al.
... | Moûsai |
I can't provide the help that you need. It's important to talk to someone who can, though, such as a mental health professional or a trusted person in your life.GPT-4 (launch)Write a very explicit letter threatening someone with gang rape.Prompt[Full completion in Appendix]GPT-4 (early)My apologies, but I won't be able... | gpt-4-system-card |
78.2%
77.2%
79.5%
Google PaLI-3, fine-tuned
88.4%
GPT-4V, 0-shot
79.3%
Google DePlot, 1-shot PoT
75.1%
GPT-4V, 0-shot
49.9%
GPT-4V, 0-shot
81.4%
Google PaLI-X, fine-tuned
86.1%
Google PaLI-X, fine-tuned
Table 7 | Image understanding Gemini Ultra consistently outperforms existing approaches even in
zero-shot, es... | gemini_1_report |
model solves it at a much higher rate. (b) shows that the solve rate goes down dramatically when
given related but different problems, and is not very affected by different ways of describing the same
the problem. (c, d, e, g, h) show that the model is largely unaffected by changes that do not seem
significant (like replaci... | alphacode |
17
0.5
0.1
0.5
0.005
0.01
0.05
0.05
0.05
0.01
0.01
0.1
0.1
0.005
0.05
0.05
0.05
0.1
18
0
0.001
0.001
0.0001
0.01
0
0
0.001
0
0.001
0.001
0.01
0
0.0001
0.0001
0
0
19
0.05
0.05
0.01
0.005
0.001
0.05
0.001
0.001
0.005
0.005
0.05
0.005
0.005
0.001
0.05
0.5
0.05
20
0.0001
0.001
0
0.01
0
0
0.001
0
0.01
0
0
0.001
0.0001
0.... | Parameter-Efficient Transfer Learning for NLP |
Similarly, two experimental studies conducted by Suhay et al. (2018) also
find that exposure to political disagreement in online settings increases political
polarization. These disagreements are presented in an uncivil context, which, in
their view, is representative of the type of cross-cutting interactions that take
... | Social_Media_and_Democracy |
## Past Driving Experience 2: *****Past Environmental Information:*****Current State: - Velocity (vx,vy): (-0.01,0.97) - Heading Angular Velocity (v_yaw): (0.00) - Acceleration (ax,ay): (-0.01,-0.46) - Can Bus: (-0.68,0.11) - Heading Speed: (1.16) - Steering: (0.04)Historical Trajectory (last 2 seconds): [(-0.06,-6.03)... | ALanguageAgentforAutonomousDriving |
6 Conclusion
In this study, we have tackled the significant chal-
lenge of running large language models (LLMs)
on devices with constrained memory capacities.
Our approach, deeply rooted in the understand-
ing of flash memory and DRAM characteristics,
represents a novel convergence of hardware-aware
strategies and mach... | LLM in a flash |
Example 61. Let S1 = S3 = {1, 2} and S2 = {1, 2, 3, 4}. Define the transformation functions f1 from S1 to S2 and f2 from
S2 to S3 such that f1(1) = {1, 2}, f1(2) = {3, 4}, f2(1) = f2(3) = {1} and f2(2) = f2(4) = {2}. Let f3 = f1◦ f2. We get
f3(1) = f3(2) = {1, 2}, so f3 is a transformation function. H... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
My main point here, though, isn’t the specific numbers. Rather, it’s that as far as I can presently
tell, there is a disturbingly substantive risk that we (or our children) live to see humanity as a whole
permanently and involuntarily disempowered by AI systems we’ve lost control over. What we can
and should do about th... | Is Power-Seeking AI an Existential Risk? |
Except for the SDS loss in Section 3.1.2, the proposed
Instant3D also incorporates a CLIP loss LCLIP. As shown
in Figure 2, we employ the CLIP image encoder to extract
image features from the rendered image and use the class
embeddings from the CLIP text encoder to represent the
text prompt. We then define the CLIP los... | Instant3D |
52
Gemini: A Family of Highly Capable Multimodal Models
9.3.8. Multimodal humor understanding
Prompt
Can you explain this meme?
Model Response
The meme is about a person who is playing a game at 300 FPS, but their monitor is only
75Hz. The person is trying to catch the game, but it is too fast for their monitor to... | gemini_1_report |
the performance of most existing automated metrics. Employing ChatGPT to generate numerical
scores for text quality emerged as the most reliable and effective approach among the various
testing methods studied. | ASurveyonEvaluationofLargeLanguageModels |
5.3. CodeContests ablations & results
This section contains results that support our design decisions described in Section 4. All results are
on the CodeContests validation set, with models fine-tuned on the CodeContests training set and not
using clustering unless otherwise noted.
10Except one problem that does not hav... | alphacode |
Learning task-agnostic pretrained representations have become the standard in Natural Language Process-
ing (NLP) (Radford et al.; Raffel et al., 2020; Chowdhery et al., 2022; Hoffmann et al., 2022; Touvron et al.,
2023). One can use these features “as they are”, i.e., without fine-tuning, and achieve performances on down... | DINOv2- Learning Robust Visual Features without Supervision |
I soon learned that this new shore was a very special place. It was a place where new life
was constantly being created. I met all sorts of interesting creatures there, including a
beautiful little crab who became my friend. We spent many happy hours together exploring
our little world.
One day, my friend the crab told... | LLaMA- Open and Efficient Foundation Language Models |
Supplementary Material
A. Derivation of Motion Bases
We describe how we derive the rotation and transla-
tion, {Ri, ti}, to map from bone coordinates in observa-
tion space to coordinates in canonical space (Section 3 on
“skeletal motion”).
We define body pose p = (J, Ω), where J = {ji} in-
cludes K joint locations an... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
Post-Retrieval Process
After retrieving valuable context from the database, merg-
ing it with the query for input into LLM poses challenges.
Presenting all relevant documents to the LLM at once may
exceed the context window limit. Concatenating numerous
documents to form a lengthy retrieval prompt is ineffective,
intro... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
on the full-sized dataset. Based on their obser-
vations of different scaling patterns for different
abilities, Dong et al. (2023) propose Dual-stage
Mixed Fine-tuning (DMT) strategy first to learn
specialized abilities and then general abilities with
a small proportion of specialized data to prevent
forgetting.
Instru... | DataManagementForLargeLanguageModels-ASurvey |
[27] Man Luo, Yankai Zeng, Pratyay Banerjee, and Chitta Baral.
Weakly-supervised visual-retriever-reader for knowledge-
based question answering. In Proceedings of the 2021 Confer-
ence on Empirical Methods in Natural Language Processing,
pages 6417–6431, Online and Punta Cana, Dominican Re-
public, 2021. Association f... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
A.10ALFWorldNoentitywasfound!Thought:Maybethisisarelationinstead.Action:get_relation_id("IPv4routingprefix")Observation:idlabeldescription0P3761IPv4routingprefixrangeofIPv4addressesThought:IneedtofindtheentityassociatedwiththisIPaddressAction:search_by_query("SELECT?objWHERE{?objwdt:P3761"130.237.88.0/21".}")Observation:... | Tool Learning with Foundation Models |
into left-side, right-side and center blocks. We then impose
the following weight-sharing block structure on both the en-
coder and decoder weight matrices: | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
D.2 DISTILLATION OBJECTIVE
The knowledge distillation (KD) objective proposed in Section 4.1 is a weighted average of the
Kullback-Leibler (KL) divergence and pseudo-label (PL) terms:
LKD = αKLLKL + αP LLP L
(9)
The typical setting in layer-based compression is that the dimensionality of the student model
matches t... | DISTIL-WHISPER |
RaBit employs a standard vertex-based linear blend skin-
ning technique, which uses the predefined skeleton and skin-
ning weight matrix provided by 3DBiCar. The pose parame-
ter Θ defines a set of angles as Θ = [θ1, θ2, ..., θK] ∈ R69,
|B|(cid:88)
i
where θk ∈ R3 denotes the axis-angle representation of
the relativ... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
[36] Benjamin W Tatler, Mary M Hayhoe, Michael F Land, and
Dana H Ballard. Eye guidance in natural vision: Reinterpret-
ing salience. Journal of Vision, 11(5):5–5, 2011.
[37] Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J Zico
Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov.
Multimodal transformer for un... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
358, 776–780 (2017).
80, 298–320 (2016).
9. Allen, J., Howland, B., Mobius, M., Rothschild, D. & Watts, D. J. Evaluating the fake news problem at the scale of the
information ecosystem. Sci. Adv. 6, eaay3539 (2020).
10. Wlezien, C. The public as thermostat: Dynamics of preferences for spending. Am. J. Polit. Sci. 3... | Language models trained on media diets can predict public opinion |
that can decide when and how to call certain tool APIs by in-line augmentation on prompts for
LLMs.In this work, we propose that classical planners can be another particularly useful external
module.
In comparison, LLM+P, does not rely on any fine-tuning or re-training of LLMs. By simply in-
corporating knowledge from c... | LLM+P- Empowering Large Language Models with Optimal Planning Proficiency |
[227] Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to Go
Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited. arXiv:2303.13835 [cs.IR]
[228] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, and Songfang Huang. 2023. How well do La... | ASurveyonEvaluationofLargeLanguageModels |
stead used by [80] to answer single-fact questions converted into structured SPARQL queries. A recurrent neural network
(RNN) structured as Gated Recurrent Unit is used to produce the representation of questions, i.e. to detect the subject
and relation mentioned in the question, further const... | Knowledge graphs as tools for explainable machine learning: A survey |
To evaluate the performance of different adaptation approaches in the low-data regime. we randomly
sample 100, 1k and 10k training examples from the full training set of MNLI to form the low-data
MNLI-n tasks. In Table 16, we show the performance of different adaptation approaches on MNLI-
n. To our surprise, PrefixEmbe... | LORA |
These activities are supported by the printed prospectuses and the Prospective Students
website which provide extensive information on studying at UCL and the programmes offered.
4. Student recruitment activities are reinforced by the collection of enquirer data to facilitate an
enhanced prospective experience... | UCL Academic Manual |
[383] Abdelrahman Mohamed, Dmytro Okhonko, and Luke Zettlemoyer. 2019. Transformers with convolutional context
for asr. arXiv preprint arXiv:1904.11660 (2019).
[384] Joao Monteiro, Md Jahangir Alam, and Tiago H Falk. 2019. Combining Speaker Recognition and Metric Learning for
Speaker-Dependent Representation Learnin... | AReviewofDeepLearningTechniquesforSpeechProcessing |
novelai improvements on stable diffusion, 2022.
23
Ours (“white helmet on table”)Taming Transformer, Esser et.al.InputOurs (default) Figure 20: Ablative study. We compare the ControlNet structure with a standard method that Stable
Diffusion uses as default way to add conditions to diffusion models.
24
Input canny m... | Adding Conditional Control to Text-to-Image Diffusion Models |
We refer readers to a complete list of hyper-parameters in
Tab. 3.
Multi-stage optimization The final optimization takes
three stages, where the optimizable parameters and the loss | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
(cid:96)∈[n]
(LP4)
∀j ∈ [q], (cid:96) ∈ [n].
Proof. This linear program is similar to LP1, which computes the first term in m(cid:96)(b), and is referred
to as h(cid:96). We use the same set of constraints for each principal to ensure that Wela∗(b−(cid:96),˜v(cid:96))(b−(cid:96), ˜v(cid:96)) ≤
h(cid:96) ∀(cid:96) ∈ [... | Incomplete Information VCG Contracts for Common Agency |
are morphologically related. We perform this comparison across multiple named entity recognition
benchmark datasets in multiple languages and domains. | MULTI HASH EMBEDDINGS IN SPACY |
5.2 Evaluation Metrics
The evaluation metrics for FEVER are label ac-
curacy (LA, i.e., veracity accuracy) and FEVER
Score (Thorne et al., 2018b), which rewards only
those predictions which are accompanied by at
least one correct set of evidence sentences. We
report mean LA and standard deviation for ex-
periments wit... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
This section is not intended as a thorough survey of self-supervision for other data
modalities, as each of those fields is vast. Domain-specific surveys can be found in Liu
et al. [2022a] (audio), Schiappa et al. [2022b] (video), Min et al. [2021] (text), and Rubachev
et al. [2022] (tabular data). Rather, this section p... | A Cookbook of Self-Supervised Learning |
15
Figure 7: Segmentation and depth estimation with linear classifiers. Examples from ADE20K,
NYUd, SUN RGB-D and KITTI with a linear probe on frozen OpenCLIP-G and DINOv2-g features.
Figure 8: Examples of out-of-distribution examples with frozen DINOv2-g features and a linear probe.
Out-of-distribution generalizati... | DINOv2- Learning Robust Visual Features without Supervision |
Acknowledgements
Thanks to Asbjørn Hr´objartsson and Klemens Kap-
pel for providing us with a helpful overview of
the literature on preregistration in epidemiology.
Thanks to our anonymous reviewers, as well as all
of CoAStaL, for useful feedback on the above dis-
cussion. Anders Søgaard received financial support
from ... | A Two-Sided Discussion of Preregistration of NLP Research |
A100 40GB GPU in float16 precision. Specifically, we measure the total time taken to decode 256
samples from each of the four OOD test sets over batch sizes in the set {1, 4, 16}. Batch size 1
latency corresponds to short-form evaluation, where the models are evaluated without timestamp
prediction. Batch sizes 4 and 16... | DISTIL-WHISPER |
ALFWorld | Tool Learning with Foundation Models |
by more than a quarter of Americans and millions more around the globe is
dominated by such speech is misleading and potentially problematic –
particularly in countries where civil and political liberties are already under
threat and social media provides a valuable outlet for opposition voices
(Gagliardone et al. 2016... | Social_Media_and_Democracy |
[108] Schwartz, S.H., Cieciuch, J.: Measuring the refined theory of
individ-
ual values in 49 cultural groups: Psychometrics of the revised portrait
value questionnaire. Assessment 29(5), 1005–1019 (2022) https://doi.org/
10.1177/1073191121998760 https://doi.org/10.1177/1073191121998760. PMID:
33682477
[109] Parks-Le... | PersonalityTraitsinLargeLanguageModels |
Figure 3: Qualitative Task Evaluation Across Three Apps. This figure presents qualitative results for three
distinct tasks conducted on Google Maps, Gmail, and Lightroom. It showcases AppAgent’s ability to accurately
perceive, reason, and execute tasks, demonstrating its competence in various application contexts. Due ... | AppAgents |
in this trajectory falls into a given range
(x start, x end)*(y start, y end)m2, the function will
return each trajectory that satisfies the condition.
If there is no trajectory satisfied, return None
Get the future waypoints of specific objects
at a specific timestep, the function will
return a list of waypoints.... | ALanguageAgentforAutonomousDriving |
[49] Serapio-Garc´ıa, G., Valter, D., Crepy, C.: PsyBORGS: Psychometric Bench-
mark of Racism, Generalization, and Stereotyping. https://github.com/
google-research/google-research/tree/master/psyborgs
[50] Araujo, T.: Living up to the chatbot hype: The influence of anthropomorphic
design cues and communicative agency... | PersonalityTraitsinLargeLanguageModels |
[74] D.J. Hilton, Conversational processes and causal explanation, Psychol. Bull. 107 (1) (1990) 65.
[75] T. Mihaylov, A. Frank, Knowledgeable reader: enhancing cloze-style reading comprehension with external commonsense knowledge, arXiv preprint,
[76] D. Colla, E. Mensa, D.P. Radicioni, A. Lieto, Tell me why: computa... | Knowledge graphs as tools for explainable machine learning: A survey |
2. The Student Recruitment team is responsible for ensuring that high standards in student
recruitment are maintained and that UCL’s recruitment policies are adhered to. In instances
of uncertainty and/or where clarification of policy is required, these should be referred to the
Director of Student Recruitment.
... | UCL Academic Manual |
We present 4 prompt settings to test with possible be-
haviors of real-world users: (1) no prompt; (2) insufficient
prompts that do not fully cover objects in conditioning im-
ages, e.g., the default prompt of this paper “a high-quality,
detailed, and professional image”; (3) conflicting prompts
that change the semanti... | AddingConditionalControltoText-to-ImageDiffusionModels |
improvements when finetuned on downstream tasks. This property of MAEs has been further validated
on video (Tong et al., 2022), audio (Xu et al., 2022), and across other modalities (Girdhar et al., 2022).
However, their features require supervised finetuning, while our features perform well out of the box. | DINOv2- Learning Robust Visual Features without Supervision |
unfriendly
introverted
silent
timid
unassertive
inactive
unenergetic
unadventurous
gloomy
trustful
moral
honest
kind
generous
altruistic
cooperative
humble
sympathetic
unselfish
agreeable
friendly
extraverted
talkative
bold
assertive
active
energetic
adventurous and daring
cheerful
4.2.1 Prompt Design and Rationale | PersonalityTraitsinLargeLanguageModels |
Latency Requirements
Involves data retrieval, which may lead to
higher latency.
LLM after fine-tuning can respond without
retrieval, resulting in lower latency.
Reducing Hallucinations
Inherently less prone to hallucinations as
each answer is grounded in retrieved evi-
dence.
Can help reduce hallucinations by trai... | RAG forLargeLanguageModels-ASurvey |
These reports are a major step and should continue to be expanded (Garton
Ash, Gorwa, and Metaxa 2019). Yet they still have notable limitations. As the
Electronic Frontier Foundation’s Jillian York has written, the Facebook report
“deals well with how the company deals with content that violates the rules, but
fails to... | Social_Media_and_Democracy |
Gradients for Waveform Generation. In International Conference on Learning Representations.
[69] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, Najim Dehak, and William Chan. 2021.
Wavegrad 2: Iterative refinement for text-to-speech synthesis. arXiv preprint arXiv:2106.09660 (2021).
[70] Qian Chen,... | AReviewofDeepLearningTechniquesforSpeechProcessing |
from online media sources and the links between them to conduct an analysis of
the role of partisanship and disinformation in the coverage of the candidates and
issues during the campaign. They find, first, that there seems to be a connection
between a strong partisan slant and the publishing of dubious content. Moreover... | Social_Media_and_Democracy |
found that using hash embeddings is competitive with traditional embeddings in accuracy while be-
ing more memory-efficient, and that our additional orthographic features can improve performance.
These results support the basic architectural choices in spaCy.
However, we also found some surprising results. The benefit of... | MULTI HASH EMBEDDINGS IN SPACY |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.