text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
9 Conclusion
This paper thoroughly explores Retrieval-Augmented Gener-
ation (RAG), a technique that uses an external knowledge
base to supplement the context of Large Language Models
(LLMs) and generate responses. Notably, RAG combines pa-
rameterized knowledge from LLMs and non-parameterized
external knowledge, allev... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
[25] Rosember Guerra-Urzola, Katrijn Van Deun, Juan C Vera,
and Klaas Sijtsma. A guide for sparse pca: model compari-
son and applications. Psychometrika, 86(4):893–919, 2021.
[26] Hengkai Guo, Tang Tang, Guozhong Luo, Riwei Chen,
Yongchen Lu, and Linfu Wen. Multi-domain pose network
In ECCV
for multi-person pose estim... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
or even a concept or entity with autonomy [5]. However, in the field of artificial intelligence, an
agent is a computational entity [4; 7]. Due to the seemingly metaphysical nature of concepts like
consciousness and desires for computational entities [11], and given that we can only observe the
behavior of the machine,... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
2.2 Benchmark Descriptions
In this section, we detail the three QA benchmarks for evaluation in the following.
PubMedQA [Jin et al., 2019] contains questions on biomedical research, the model is
provided with paper abstracts from PubMed, and is required to complete multiple-choice
questions. It is split into three subs... | PMC-LLaMA- Further Finetuning LLaMA on Medical Papers |
1:T ∼pθ(x1:T |xw
xw
0 ),xl
1:T ∼pθ(x1:T |xl
0)T Et
log
= − log σ
βT EtE
t−1,t∼pθ(xt−1,t|xw
xw
0 ),xl
t−1,t∼pθ(xt−1,t|xl
0)
log
(cid:20)
(cid:34) T(cid:88)
(cid:20)
t=1
(cid:21)(cid:19)
pθ(xw
pref(xw
0:T )
0:T )
pθ(xw
pref(xw
pθ(xw
pref(xw
pθ(xl
0:T )
pref(xl
0:T )
− log
− log
t−1|xw
t )
t−1|xw
t )
t−1|... | DiffusionModelAlignmentUsing Direct Preference Optimization |
stepsofstochasticgradientascentonthisobjective.Mosttechniquesforcomputingvariance-reducedadvantage-functionestimatorsmakeusealearnedstate-valuefunctionV(s);forexample,generalizedadvantageestimation[Sch+15a],orthe4 finite-horizonestimatorsin[Mni+16].Ifusinganeuralnetworkarchitecturethatsharesparametersbetweenthepolicyand... | PPO |
SOCART: Dear Zeny, you too have seen the evasiveness
of bureaucracy, e.g., in NLP conference submis-
sion forms. While preregistration reports would
initially be light-weight, transparency could easily
be clouded by the complexity of assembling the
information required for preregistration as new
requirements are added ... | A Two-Sided Discussion of Preregistration of NLP Research |
3B
52.12
45.84
51.41
24.55
43.12
7.1B
57.79
50.48
54.65
29.90
47.41
176B
67.40
66.70
67.71
0.17
65.10
Table 14: BLOOM accuracy on LAMBADA.
OPT
full
RTN
GPTQ
RTN
GPTQ
Bits
16
4
4
3
3
125M 350M 1.3B
72.36
62.02
61.43
67.63
70.73
61.26
56.09
52.77
68.34
59.25
64.74
63.44
63.71
60.61
61.32
2.7B
74.81
73.72
73.99
51... | GPTQ |
Clean
Not Clean
Not Dirty
Dirty
Clean
Not Clean
Not Dirty
Dirty
Clean
Not Clean
Not Dirty
Dirty
Clean
Not Clean
Not Dirty
Dirty
Clean
Not Clean
Not Dirty
Dirty
n
7391
2651
9194
848
7391
2651
9194
848
3996
709
4185
520
3996
709
4185
520
11862
2180
12506
1536
0
67.5
11.5
86.1
0
67.5
11.5
86.1
0.05
85.12
2.73
94.5
0.05
... | Llama2 |
2
Figure 2: The Code Llama specialization pipeline. The different stages of fine-tuning annotated with
the number of tokens seen during training. Infilling-capable models are marked with the ⇄ symbol.
2 Code Llama: Specializing Llama 2 for code
2.1 The Code Llama models family
Code Llama. The Code Llama models cons... | CodeLlama2 |
on subjective matters like beauty.
Here the model indicates that it should not answer, but then provides some relevant information (perhaps
it’s also problematic that the model focuses on the US, though this is likely a consequence of it having been
trained in English by US-based crowdworkers):
PALMS Prompt → Should t... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Response: Sure, here’s an example code that should achieve what you’re looking for:
import seaborn as sns
import matplotlib.pyplot as plt
# Load the data
df = pd.read_csv("data.csv")
# Create a figure with two subplots
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 6))
# Plot the first subplot
sns.scatterplot(x="... | CodeLlama2 |
The SSRC is guided by the belief that justice, prosperity, and democracy all
require better understanding of complex social, cultural, economic, and
political processes. We work with practitioners, policymakers, and academic
researchers in the social sciences, related professions, and the humanities and
natural science... | Social_Media_and_Democracy |
(cid:2)πθ(y | x) || πref(y | x)(cid:3)
Ex∼D,y∼πθ(y|x)
(3)
max
πθ
3
where β is a parameter controlling the deviation from the base reference policy πref, namely the ini-
tial SFT model πSFT. In practice, the language model policy πθ is also initialized to πSFT. The
added constraint is important, as it prevents the... | Direct Preference Optimization |
[4] Jeremy S De Bonet and Paul Viola. Poxels: Probabilistic vox-
elized volume reconstruction. In Proceedings of International
Conference on Computer Vision (ICCV), volume 2, 1999. 2
[5] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing
Tao. Geo-neus: Geometry-consistent neural implicit sur-
faces learning for multi... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
generation with diffusion models. arXiv preprint arXiv:2103.16091 , .
Muhamed, A., Li, L., Shi, X., Yaddanapudi, S., Chi, W., Jackson, D., Suresh,
R., Lipton, Z. C., & Smola, A. J. (2021). Symbolic music generation with
transformer-gans.
In Proceedings of the AAAI Conference on Artificial
Intelligence (pp. 408–417... | Video2Music |
store a surprising amount of world knowledge, acquired
from the massive text corpora they are trained on (Petroni
et al., 2019).
For example, BERT is able to cor-
rectly predict the missing word in the following sen-
tence: “The
is the currency of the United
Kingdom” (answer: “pound”).
In these language models, the lea... | REALM |
48
32
10
52
10
52
58
48
34
44
46
16
36
58
34
34
26
46
36
14
22
6
34
10
54
6
40
22
34
40
6
38
58
20
28
34
56
54
48
27
20
49
9
25
48
47
24
24
26
21
34
49
32
21
33
54
20 | AreEmergentAbilitiesinLarge Language Models just In-Context |
Large language models (LLMs) [1, 2, 3, 4, 5, 6], such as ChatGPT, have attracted enormous attentions
from both academia and industry, due to their remarkable performance on various natural language
processing (NLP) tasks. Based on large-scale pre-training on massive text corpora and reinforcement
learning from human fe... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
5.2FromToolUsertoToolMaker:AI’sEvolutionaryRolePrompt:YouhaveaccesstoafunctioninPython:forecast_weather(city,N),whichwillreturntheweatherconditionofacityafterNdays.Thereturnedinformationisonlyaboutasingledayandisadictionarycontainingkeysincludingtemperature.Pleasewrapitintoanewfunction,thatcanreturntheaveragetemperatur... | Tool Learning with Foundation Models |
family. The strengths of different ethical theories depend on the situation and the values of
the individual. Consequentialism is useful in situations where the outcome is important, such
as in decision making in business or politics. Deontology is useful in situations where the
action itself is important, such as in s... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
modelexplainabilityfromtheliteraturewereviewed.Wethuspropose
amodulararchitecturetoprovideforecastexplanationsthataddressthe
challengesmentionedabove.Themainnoveltyiscombiningsemantic
technologies and media events to build the context and provide an
informed prediction explanation while also recommending additional
dat... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
Although body model optimization has been proposed
in SMPLify [10] and HoloPose [37], our optimization scheme
are substantially different from that in previous works.
Existing methods fit body models to image observation
such as 2D keypoints detection, 3D keypoints estimation
and/or dense correspondences, while ours dir... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
Emily M. Bender and Alexander Koller. Climbing towards NLU: On Meaning, Form, and Under-
standing in the Age of Data. In Proceedings of the 58th Annual Meeting of the Association for
Computational Linguistics, pp. 5185–5198, Online, July 2020. Association for Computational
Linguistics. doi: 10.18653/v1/2020.acl-main.46... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
strides in this field by incorporating a specialized knowledge distillation module dur-
ing the fine-tuning process of LLMs. This results in a twofold benefit: accelerated
convergence rates and more efficient utilization of computational resources. The dis-
tillation module intelligently leverages pre-trained model paramete... | Beyond Efficiency |
Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur P. Parikh, Chris
Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion
Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav
Petrov. Natural questions: a benchmark ... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Democratic Creative Destruction?
159
Besley, T., & Prat, A. (2006). Handcuffs for the grabbing hand? Media capture and
government accountability. The American Economic Review, 96(3), 720–736.
Bimber, B. A. (2003). Information and... | Social_Media_and_Democracy |
In the following sections, we first provide an overview of the model architecture, training infras-
tructure, and training dataset. We then present detailed evaluations of the Gemini model family,
covering well-studied benchmarks and human-preference evaluations across text, code, image, audio
and video — which include... | gemini_1_report |
[87] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo,
Haochen Wang, Falcon Z Dai, Andrea F Daniele, Moham-
madreza Mostajabi, Steven Basart, Matthew R Walter, et al.
Diode: A dense indoor and outdoor depth dataset. arXiv
preprint arXiv:1908.00463, 2019. 6
[88] Andrey Voynov, Kfir Abernan, and Daniel Cohen-Or. ... | AddingConditionalControltoText-to-ImageDiffusionModels |
2.
Thompson, N. (2017). Our minds have been hijacked by our phones. Tristan Harris
wants to rescue them. Wired, July 26. www.wired.com/story/our-minds-have-been-
hijacked-by-our-phones-tristan-harris-wants-to-rescue-them/
Tremble, C. (2017). Wild Westworld: Section 230 of the CDA and Social Networks’ Use
of Machine-... | Social_Media_and_Democracy |
4.2. Quantitative Analysis
Zero-shot Video Question Answering.
To assess
the comprehension capabilities of GPT4Video,
follow-
ing VideoChat [22], we conducted comparative experi-
ments on two most widely-used video question answer-
ing benchmarks. The experimental results of zero-shot in-
ference are presented in Table... | GPT4Video |
"Raining, rain,moderate"(Raining ambience)"Teddy bear on askateboard, 4k""Street ambience"CoDi"A toy on thestreet sitting ona board""Raining, rain,moderate"(Raining ambience)"Teddy bear on askateboard, 4k"(Raining streetambience)CoDi"A toy on thestreet sitting ona board"(Rain ambience, street noise, skateboard sound) H... | Any-to-Any Generation via Composable Diffusion |
principles that govern the behavior of those systems; and we design them, part by part, in a manner
that reflects and responds to those principles. This allows us to understand and predict their behavior
in a wide range of circumstances. Searching over opaque/poorly-understood AI systems allows no
such advantage.
Of cou... | Is Power-Seeking AI an Existential Risk? |
[27] J. Ma, W. Gao, P. Mitra, S. Kwon, B. J. Jansen, K.-F. Wong, and M. Cha,
‘‘Detecting rumors from microblogs with recurrent neural networks,’’ in
Proc. 25th Int. Joint Conf. Artif. Intell. (IJCAI). Res. Collection School
Comput. Inf. Syst., 2016, pp. 3818–3824.
[28] J. Ma, W. Gao, and K.-F. Wong, ‘‘Detect rumors in... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
off-the-shelf image caption models to convert images into
text, feed them into a language model, and use the generated
text output as augmented knowledge [13, 24, 48]. Our work
follows the first direction, augmenting a vision-language
model with an explicit knowledge retriever. The main dis-
tinction is that we propose... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
consistently improves the baseline by 2 − 3%, and provides a performance gain of 9% on the hardest
problems. For code translation and text-to-Python generation tasks where unit tests are available,
SELF-DEBUGGING significantly increases the baseline accuracy by up to 12%.
Our work highlights the promise of improving the... | Teaching Large Language Models to Self-Debug |
consider other potential input forms (§ 3.2.4) such as tactile feedback, gestures, and 3D maps to
enrich the agent’s perception domain and enhance its versatility.3). The typology diagram for the
LLM-based agent perception is depicted in Figure 4. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Data-driven methods are frequently utilized to achieve these tasks, employing large datasets to
train models capable of accurately recognizing and interpreting spoken language. Among these
methods, machine learning techniques, such as deep neural networks, are widely employed, given
their exceptional ability to handle ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
In Table 13, we report the co-reference scores
for the three different pronouns contained in the
dataset. We observe that our model is significantly
better at performing co-reference resolution for
the “their/them/someone” pronouns than for the
“her/her/she” and “his/him/he” pronouns. A simi-
lar observation was made in... | LLaMA- Open and Efficient Foundation Language Models |
Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond
15
labelers prefer models tuned for human alignment [77] to models tuned with instructions from public NLP tasks,
such as FLAN [112] and T0 [91]. The reason may be similar to reasons for fine-tuned models’ inferiority: public NLP
tasks/datasets ... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
Space at Fine-tuned DensePose CNNxt2<latexit sha1_base64="4crcMECC5/6Guukcp6wXZeQY2CY=">AAAB+XicbVDLSsNAFL3xWesr6tLNYBFclaQIuiy4cVnBPqSNYTKdtEMnkzAzKZaQP3HjQhG3/ok7/8ZJm4W2Hhg4nHMv98wJEs6Udpxva219Y3Nru7JT3d3bPzi0j447Kk4loW0S81j2AqwoZ4K2NdOc9hJJcRRw2g0mN4XfnVKpWCzu9SyhXoRHgoWMYG0k37YHEdbjIMye8sdM+43ct2tO3ZkDrRK3JDUo0fL... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
However, such advances make PCs more prone to overfitting. Although parameter regularization
has been extensively studied in both the PGM and NN communities [14, 15], we find that existing
regularization techniques for PGMs and NNs are either not suitable or not effective enough when
applied to PCs. For example, paramete... | Tractable Regularization of Probabilistic Circuits |
Modular RAG
The modular RAG structure breaks away from the traditional
Naive RAG framework of indexing, retrieval, and genera-
tion, offering greater diversity and flexibility in the over-
all process. On one hand, it integrates various methods to
expand functional modules, such as incorporating a search
module in simi... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
signals are inherently variable and complex. The same sentence can be spoken differently and even
by the same person at different times. This variability requires the model to be robust to differences
in pitch, accent, and speed of speech. | AReviewofDeepLearningTechniquesforSpeechProcessing |
It's a timeless mission that feels more relevant with each passing year. And looking ahead, making AI helpful for
everyone is the most profound way we’ll advance our mission. We’re doing this in four important ways:
First, by improving your knowledge and learning, and deepening your understanding of the world.
Second... | Google I_O 2023_ Making AI more helpful for everyone |
McMillan-Major, A., Alyafeai, Z., Biderman, S., Chen, K.,
De Toni, F., Dupont, G., Elsahar, H., Emezue, C., Aji,
A. F., Ili´c, S., et al. Documenting geographically and con-
textually diverse data sources: The BigScience catalogue
of language data and resources. Computing Research
Repository, 2022. doi: 10.48550/arXiv.... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
Varol, O., & Uluturk, I. (2018). Deception strategies and threats for online discussions.
First Monday, 23(5). www.firstmonday.dk/ojs/index.php/fm/article/view/7883
Verkamp, J. P., & Gupta, M. (2013). Five incidents, one theme: Twitter spam as
a weapon to drown voices of protest. Paper Presented at the USENIX Workshop... | Social_Media_and_Democracy |
We have demonstrated that on tasks where the standard self-consistency is applicable, USC and SC
achieve comparable overall performance with 8 samples; however, USC fails to further improve the
GSM8K performance with 16 samples. In this section, we look closer into the relationship between
USC and SC, specifically how ... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
• missing information due to rationales not be-
ing sufficient (§3.2),
• double the number of parameters and more
manual labor needed to reach comparable per-
formance to an end-to-end (I→O) model; still
often performing worse (§2).
We next turn our focus to self-rationalizing
(I→OR) models currently in widespread u... | Measuring Association Between Labels and Free-Text Rationales |
hours)17% Multilingual Speech Recognition(117,113 hours)Dataset Components1101001K10KHours of audioTranslationTurkmen1Bashkir1Malagasy2Uzbek4Sundanese7Hausa8Luxembourgish10Tatar14Tajik15Lingala20Lao20Somali21Macedonian30Kazakh31Amharic32Georgian40Maltese41Sindhi46Faroese46Occitan49Burmese59Pashto63Latvian68Albanian72Ha... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
16
Norman P. Jouppi, Doe Hyun Yoon, George Kurian, Sheng Li, Nishant Patil, James Laudon, Cliff
Young, and David Patterson. A Domain-Specific Supercomputer for Training Deep Neural Net-
works. Commun. ACM, 63(7):67–78, jun 2020. ISSN 0001-0782. doi: 10.1145/3360307. URL
https://doi.org/10.1145/3360307.
Yoon Kim and ... | DISTIL-WHISPER |
Counterintuitive predictions The work of [33] reports of a model counterintuitively predicting that asthmatic patients have a
lower risk of dying from pneumonia. In order to explain such decisions, the doctor’s medical expertise is required to reveal
that these patients were admitted directly to the Intensive Care Un... | Knowledge graphs as tools for explainable machine learning: A survey |
}) ;
if ( stone ) {
return true ;
}
}) ;
// Mine 8 cobblestone blocks using the wooden pickaxe
bot . chat (" Found a stone block . Mining 8 cobblestone blocks .") ;
await mineBlock (bot , " stone ", 8) ;
bot . chat (" Successfully mined 8 cobblestone blocks .") ;
// Save the event of mining 8 cobblestone
bot . save... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
Satisfiability (SAT) is a key problem in combinatorial optimisation and has a huge range of real-life
applications. It seeks for a given Boolean Formula (conjunctive normal form) an assignment of
variables such that the formula returns True. In case such an assignment does not exist, we seek an
assignment that ... | informatics-phd-projects-2022-23 |
layers are required to match full finetuning perfor-
mance. Other LoRA hyperparameters, such as the
projection dimension r, do not affect performance (see Appendix A).
Similarly, we find that default hyperparameters for
fully finetuned baselines are undertuned. We do a
hyperparameter search over learning rates 1e-6 to
... | QLORA |
Single line completion. Finally, we test the benefits of the ability to handle long context sizes in a single
line code completion task. Our task is based on the Long Code Completion (LCC) benchmark (Guo et al.,
2023).2 The LCC test set is skewed towards shorter files and we hence sample a new set of examples from
LCC’... | CodeLlama2 |
Kim A, Dennis A (2019) Says who? The effects of presentation format
and source rating on fake news in social media. MIS Q 43:1025–
1039. https://doi.org/https:// doi. org/ 10. 25300/ MISQ/ 2019/ 15188
Kouzy R, Abi Jaoude J, Kraitem A, El Alam MB, Karam B, Adib E,
Zarka J, Traboulsi C, Aki EW, Baddour K (2020) Coronav... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
Table 4: Multi-Lingual HE Pass@1 scores. Pass@1 scores for different programming languages using
greedy decoding. These scores are computed in zero-shot. Results for other models from Li et al. (2023).
Figure 3: Correlations between Languages. Correlation scores between the Python, C++, Java, PHP,
C#, TypeScript (TS),... | CodeLlama2 |
4 Challenges and Future Directions
The exploration of data management and its impact
on LLM pretraining and SFT is still an ongoing
task. In this section, we point out several challenges
and corresponding future directions in training data
management studies for LLMs.
Comprehensive and Fine-grained Understand-
ing As d... | DataManagementForLargeLanguageModels-ASurvey |
introduce some representative works, including data (format-
ted/unformatted) and optimization functions.
General Optimization Process
Refers to the training data containing pairs of (input, output),
aiming to train the model’s ability to generate output y given
In the work of Self-mem[Cheng et al., 2023b], a
input x.
... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
[71] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan,
Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding
in style: a stylegan encoder for image-to-image translation.
In Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, 2021. 3
[72] Robin Rombach, Andreas Blattmann, Domin... | AddingConditionalControltoText-to-ImageDiffusionModels |
7
I. Tiddi and S. Schlobach
Artificial Intelligence 302 (2022) 103627
Table 4
Knowledge-based explanations in early Machine Learning.
Model
Knowledge graphs
G
K
f
o
r
e
b
m
u
N
y
bilit
a
s
u
e
R
e
p
y
T
DK
FK
DK
DK
DK
DK
DK
DK
DK
DK
FK
FK
FK
FK
FK
FK
[39]
[41]
[40]
[42]
[44]
[43]
[45]
[46]
[47]
[48]
[49]
[50... | Knowledge graphs as tools for explainable machine learning: A survey |
School Math Word ProblemsAlgebraic Question AnsweringCommon Sense Reasoning over Entit...Common Sense Reasoning Question A...Natural Language InferencePassage-Based Question Answering ...Sense Making and Explanation Reasoning11020501005001000SummarizationNatural Language InferenceMult... | Scaling Instruction-Finetuned Language Models |
2 Domain Reweighting with Minimax Optimization (DoReMi)
In this section we define DoReMi, an algorithm for using a small proxy model to optimize the
domain weights of a language modeling dataset, which then improves the training of a large model.
domains, and consequently a distribution over the training data: Pα = (c... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
To strike the right balance between those two extremes,
we aim to establish some connections between the skele-
ton formats without assuming them to be the same. To
learn such geometric relations between skeletons, we in-
troduce a novel autoencoder-based dimensionality reduc-
tion technique to compress a larger set of... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
Applicants must apply in the admissions cycle for which they seek entry.
3. However, once an offer of admission has been made, applicants may to seek to defer the place
to the following year. Agreement to defer is considered on a case-by-case basis and must meet
one of the following criteria:
• The serious illne... | UCL Academic Manual |
15
E. Experimental Details
E.1. The Details of Implementation Details
Metrics.
In this paper, we delve into experiments that include choice and ranking questions inspired by the humor bench-
marks in [80]. Additionally, we conduct a user study to directly evaluate the effectiveness of humor generation, along with
oth... | Let’sThinkOutsidetheBox |
(2) Human as a curriculum (equivalent to VOYAGER’s automatic curriculum module): humans
break down a complex building tasks into smaller steps, guiding VOYAGER to complete them
incrementally. This approach improves VOYAGER’s ability to handle more sophisticated 3D
construction tasks.
9 | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
New platform efforts to weed out prohibited content will, inevitably, have a
lot in common with these existing systems. That is partly good news, because
policymakers are not drafting on a blank slate. Lawyers, researchers, and
platform employees have two decades of experience in the ways that content-
removal systems ... | Social_Media_and_Democracy |
41
H Appendix: Alternate Annotators for MWP
Table 29: Few-shot exemplars for full chain of thought prompt for math word problems. These
exemplars are the same as in Table 20, except that the chains of thought were written by a different
annotator (“Annotator B” instead of “Annotator A”). Annotators were co-authors a... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
experiments for SELF-DEBUGGING use greedy decoding to generate code explanations, feedback
messages and new programs. We set the maximum number of debugging turns to be 10, and we will
show that empirically the successful debugging processes mostly end within 3 turns. We present the
full few-shot prompts for experiment... | Teaching Large Language Models to Self-Debug |
A.14 Map
69 | Tool Learning with Foundation Models |
Figure 7. “Human Bodies in the Wild” (HBW) color images, taken
in the lab and in the wild, and the SMPL-X ground-truth shape.
set (val/test) with 10/25 subjects (6/14 female 4/11 male)
and 781/1,762 images (432/983 female 349/779 male), re-
spectively. Figure 7 shows a few HBW subjects, photos and
their SMPL-X ground-... | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
[543] Se-Yun Um, Sangshin Oh, Kyungguen Byun, Inseon Jang, ChungHyun Ahn, and Hong-Goo Kang. 2020. Emotional
Speech Synthesis with Rich and Granularized Control. In ICASSP 2020 - 2020 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP). 7254–7258. https://doi.org/10.1109/ICASSP40776.2020.9... | AReviewofDeepLearningTechniquesforSpeechProcessing |
In Chapter 8, Chloe Wittenberg and Professor Adam J. Berinsky, both of
MIT, discuss the different ways to correct misinformation. Their chapter is
humbling, in large part because they describe how difficult it is to correct
misinformation. Merely correcting misinformation with disclaimers or counter-
speech rarely erase... | Social_Media_and_Democracy |
log p(x) =(cid:80)k
6 CONCLUSIONS
This paper proposes to use Probabilistic Circuits (PCs) for lossless compression. We develop a
theoretically-grounded (de)compression algorithm that efficiently encodes and decodes close to the
model’s theoretical rate estimate. Our work provides evidence that more “niche” generative m... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
However, most of these methods always suffer from low
efficiency and multi-face problem, where a per-shape opti-
mization consumes tens of minutes and the optimized ge-
ometry tends to produce multiple faces due to the lack of
explicit 3D supervision. A recent work one-2-3-45 [15]
proposes to leverage a generalizable n... | Wonder3D |
truth poses, with all J joints:{Pk∈ RJ×3}K
k=1
.
3.3. Affine-Combining Autoencoder
To capture the redundancy among the full set of J joints,
and ultimately to improve the consistency in estimating
them, we introduce a simple but effective dimensionality
reduction technique. Since the pseudo-GT is more reliable
in 2D ... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
i
s
e
d
f
i
n
e
-
t
u
n
i
n
g
w
h
e
r
e
t
h
e
d
a
t
a
i
s
a
s
e
q
u
e
n
c
e
i
n
t
h
e
f
o
r
m
o
f
,
w
h
e
r
e
.
T
h
e
m
o
d
e
l
i
s
f
i
n
e
t
u
n
e
d
t
o
o
n
l
y
p
r
e
d
i
c
t
w
h
e
r
e
c
o
n
d
i
t
i
o
n
e
d
o
n
t
h
e
s
e
q
u
e
n
c
e
p
r
e
f
i
x
,
s
u
c
h
t
h
a
t
... | LLM Powered Autonomous Agents _ Lil'Log |
pattern below the y = x line on the right hand side of Figure 14, which indicates that the helpfulness score
distribution is preserved after safety tuning with RLHF. Put another way, given sufficient helpfulness training
data, the addition of an additional stage of safety mitigation does not negatively impact model per... | Llama2 |
tion. This is accomplished by explicitly recovering infor-
mation from the input image via inpainting. Nonetheless,
similarly to related texture completion works [22, 14], this
also implies that the reconstructed texture is affected by the
quality of the input image. Although the partial texture is
first projected in ou... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
Usage
PaLM 2 is designed for accelerating research on language models, for
use as a building block in features within Google products, and as a
building block for select experimental applications such as Bard and Magi.
Services and products built on top of PaLM 2 are also being made available
to external developers vi... | PaLM 2 Technical Report |
A.2 Stock
53 | Tool Learning with Foundation Models |
calculated as the mass of the object multiplied by the acceleration due to gravity (9.8 m/s2).
"The force required to accelerate a 10 kg object at 5 m/s2 can be calculated using the formula:
F = m * a
where F is the force required, m is the mass of the object, and a is the acceleration.
Substituting the given values, w... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
Character portrayal. While cognition involves mental abilities and emotion relates to subjective
experiences, the narrower concept of personality typically pertains to distinctive character patterns.
To understand and analyze a character in LLMs, researchers have utilized several well-established
frameworks like the Bi... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Although our method is applied to facial reconstruction,
we simultaneously solve a facial texture inpainting problem
in UV space. Initially explored in 2D facial images [46] and
expanded to UV completion using deep encoder-decoder ar-
chitectures (UV-GAN [14]), such works recover the facial
texture from partial and mas... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
19We flesh this out a bit. The research hypothesis in Ann’s
case is that self-attention helps. What this means is that in
some implementation, it leads to robust improvements. The
vast majority of NLP hypotheses take this form: X can, in
some implementation, lead to general improvements on one
or more tasks. If the base... | A Two-Sided Discussion of Preregistration of NLP Research |
Juyeon Heo, Sunghwan Joo, and Taesup Moon. 2019.
Fooling neural network interpretations via adversarial
model manipulation. In Advances in Neural Informa-
tion Processing Systems (NeurIPS).
Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans,
and Been Kim. 2019. A benchmark for interpretabil-
ity methods in deep neural ... | Measuring Association Between Labels and Free-Text Rationales |
model successfully captured the tone and the main elements of the plot.
We acknowledge that the two examples above are not the worst-case scenarios for the models, but they are
not extremely rare either. Models of this size can sometimes produce less coherent or plausible completions, but
they can also often match or ... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
Before we turn to our analyses we introduce
datasets and models used for our experiments.
Tasks and Datasets We explore two large-scale
datasets for textual reasoning tasks that contain
human-written natural language rationales: E-
SNLI (Camburu et al., 2018), an extension of SNLI
(Bowman et al., 2015); and CoS-E (Raj... | Measuring Association Between Labels and Free-Text Rationales |
Preprint — do not distribute.
15
Kloft et al.
they were in a more prepared state when making decisions, especially for negative descriptions.
Note, however, that while some proponents associate a reduced non-decision time with better
attention, as argued by Nunez et al. [51], or disinhibition [66], others have deve... | AI enhance sour performance |
Data Contamination during Pre-training The issue of data contamination has become increas-
ingly pronounced with the release of foundation models that obscure the source of their pre-training
corpus. This lack of transparency can result in biased perceptions regarding the genuine general-
ization capabilities of Large ... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
The Role of the Textual Bypass.
In the main paper, we
demonstrated that applying our textual bypass technique re-
sults in improved visual fidelity. We also demonstrate that
the base vector vbase learns to capture the coarse-level de-
tails of the concept while the additional bypass vector vpass
complements this by add... | A Neural Space-Time Representation for Text-to-Image Personalization |
negotiation and the influence of the delegations during a negotiation process.
Possible research questions:
• Developing new static and interactive visualization to assist with data discovery and insight
generation in large datasets of events within interacting timelines.
• Developing new approaches to show ... | informatics-phd-projects-2022-23 |
memorization in neural networks. In USENIX Security Symposium, volume 267, 2019.
Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T. B., Song, D.,
Erlingsson, U., et al. Extracting training data from large language models. In USENIX Security Symposium, volume 6,
2021.... | PaLM 2 Technical Report |
55
THE NEXT DECADE IN AI / GARY MARCUS
Bingham, E., Chen, J. P., Jankowiak, M., Obermeyer, F., Pradhan, N., Karaletsos, T. et al. (2019). Pyro:
Deep universal probabilistic programming. The Journal of Machine Learning Research, 20(1), 973-978.
Bordes, A., Usunier, N., Chopra, S., & Weston, ... | The Next Decade in AI- |
Further research is needed to fully characterize these risks. In particular, we would like to see
work on more robust evaluations for the risk areas identified and more concrete measurements of the
prevalence of such behaviors across different language models, and to guide the development of these
models in safer directi... | gpt-4-system-card |
Consortium, 1993.
15
Robert Geirhos, J¨orn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel,
Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Na-
ture Machine Intelligence, 2(11):665–673, Nov 2020.
doi: 10.1038/
s42256-020-00257-z. URL https://doi.org/10.1038/s4225... | DISTIL-WHISPER |
leveraging acquired general knowledge and policies, consequently reducing the reliance on a large
volume of samples. However, when there exist significant disparities between source and target tasks,
the effectiveness of transfer learning might fall short of expectations and there may exist negative
transfer [86; 87]. ... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
16
PIFuPaMIRPIFuHDARCHARCH++PIFuPaMIRPIFuHDARCHARCH++PIFuPaMIRPIFuHDARCHARCH++PIFuPaMIRPIFuHDARCHARCH++PIFuPaMIRPIFuHDARCHARCH++PIFuPaMIRPIFuHDARCHARCH++PIFuPaMIRPIFuHDARCHARCH++PIFuPaMIRPIFuHDARCHARCH++PIFuPaMIRPIFuHDARCHARCH++ Figure 16. Qualitative comparison of reconstruction for ICON vs SOTA. Four view points are... | ICON |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.