text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
There are a few additional challenges when applying SSL to RL. First, if the data is
recorded on-line, individual observations are highly correlated with each other and are
not IID (independent and identically distributed), so sampling from replay buffer should
be done carefully. One failure mode of SSL objectives when ... | A Cookbook of Self-Supervised Learning |
Startups / Projects (Funding activity⁴)
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
*Compound Annual Growth Rate
Wave 4
75%
CAGR*
63%
CAGR
84%
CAGR
62%
CAGR
a16z crypto
State of Crypto
2023
Market Cycles
13
13
Great products get built regardless of f... | State-of-Crypto2023 |
via prompt engineering. CoRR, abs/2308.07411, 2023.
[558] Phelps, S., Y. I. Russell. Investigating emergent goal-like behaviour in large language models
using experimental economics. CoRR, abs/2305.07970, 2023.
[559] Bellomo, N., G. A. Marsan, A. Tosin. Complex systems and society: modeling and simulation,
vol. 2. ... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
5
xNmulti-axis transformer blockself-attncross-attnlow-resspatialverticalspatialhorizontaltemporalhigh-restoken factorization (k=2)low-res tokenshigh-res masked tokensembeddingT5X embeddingsmulti-head classification and merging(k=2)high-res output tokens condition and the text embeddings. We use a cascade of
two 2× s... | VideoPoet |
0.9868
0.9506
0.9659
1
0.9153
0.8732
0.8
0.9322
0.8861
0.9074
0
0.9474
0.9231
0.9714
0.9833
0.9245
0.8431
0.8
0.9355
0.9524
0.8906
0.9
72
62
55
44
86
66
47
57
68
71
76
47
76
81
88
65
59
71
60
59
79
54
0
57
65
70
60
53
51
30
62
63
64
60
Pick up a wooden_shovel given nothing.
Pick up a wooden_pickaxe given nothing.
... | JARVIS-1 |
ing training of large, sparse models. In ICML. PMLR, 2021.
[25] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal
model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.
[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual in... | Mixture-of-Experts |
12 | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
[52] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M
Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In ICLR, 2022.
[53] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain
of thought prompting elicits... | Mixture-of-Experts |
reduces the learning rate to about 10% of its peak value, ensuring a balance between rapid learning and stability as training
progresses. The choice of optimizer also plays a pivotal role in stabilizing the training of LLMs. Optimizers like Adam [133]
and AdamW [175] are popular choices for models such as GPT-3 [24] an... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
33 | A Cookbook of Self-Supervised Learning |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
58
Alexandra A. Siegel
Waseem and Hovy 2016; Kennedy et al. 2018; Olteanu et al. 2018). Yet despite
these advances, there is still no consensus in the scientific literature on how to
define online hate speech. | Social_Media_and_Democracy |
24https://github.com/OpenBMB/BMTools
28
5.2 From Tool User to Tool Maker: AI’s Evolutionary Role | Tool Learning with Foundation Models |
+86
0.0
+95
English 32K sentencepiece for all models. Within the context of decoder-only models, except for the case of
the decoder model trained on causal LM, our experiments always use a bidirectional receptive field only in
it’s input segment and autoregressive decoding at the targets segment. This is essentially th... | UL2- Unifying Language Learning Paradigms |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
92
Samuel C. Woolley
amount of bots function online today, but it is also true that social bots have
existed on the Internet for several decades. | Social_Media_and_Democracy |
[5] Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. Reading Wikipedia to Answer
Open-Domain Questions. In Proceedings of the 55th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers), pages 1870–1879, Vancouver, Canada,
July 2017. Association for Computational Linguistics. d... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
Hi, I’m Mount Everest. What would you like to know about me?
Why do you think people climb you?
Because I represent a very high achievement to climb me. Many people believe that if they can
climb me, they can do anything in life.
Who was the first person to do that?
The first confirmed persons to have reached the summit o... | LaMDA- Language Models for Dialog Applications |
WebGPT: Browser-assisted question-answering with human feedback, Hilton et al., 2022.
94 Constitutional AI: Harmlessness from AI Feedback, Anthropic, December 2022.
35 Frontier AI – Capabilities and Risks
95 Evaluation of OpenAI Codex for HPC Parallel Programming Models Kernel Generation, Godoy et al., 2023;
Ope... | Capabilities and risks from frontier AI |
environmental factors | Social_Media_and_Democracy |
Less than half of the public believes these technologies would improve things
over the current situation. One factor tied to public views of human enhancement is
whether people think these developments would make life better than it is now, or whether
reliance on AI would improve on human judgment or performance. On th... | AI and Human Enhancement_ Americans’ Openness Is Tempered by a Range of Concerns _ Pew Research Center |
Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
evolution of γ : 1.8 → 0.45 over several red teaming iterations and model refinements. Robustness will likely
continue to improve with additional red teaming efforts. Another magnitude that we tracked as new models
were produced was the percentage of prompts triggering violating responses discovered in the previous red... | Llama2 |
Let’s look at the factors relevant to deployment decisions in more detail.
Consider some set of decision-makers (at e.g. a lab, a company, a government, etc) deciding whether
or not to deploy a given APS system, or to pursue some other alternative (running more tests, re-
designing/retraining the system, scrapping the ... | Is Power-Seeking AI an Existential Risk? |
Demo: https://jaywalnut310.github.io/vits-demo/index.html
(cid:104)
(cid:105)
qφ(z|x)
pθ(z|c)
log pθ(x|z)−log
log pθ(x|c) ≥ Eqφ(z|x)
(1)
where pθ(z|c) denotes a prior distribution of the latent vari-
ables z given condition c, pθ(x|z) is the likelihood func-
tion of a data point x, and qφ(z|x) is an approximate p... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
(4)
The encoded features are then passed to a shallow MLP.
One alternative to hash encoding is sparse voxel structures
[30, 33, 39, 43], where each grid corner is uniquely defined
without collision. However, volumetric feature grids require
hierarchical spatial decomposition (e.g. octrees) to make the
parameter count... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
[25] Matthias Innmann, Michael Zollh¨ofer, Matthias Niessner,
Christian Theobalt, and Marc Stamminger. VolumeDeform:
Real-time volumetric non-rigid reconstruction. In Proc. Euro-
pean Conf. on Computer Vision (ECCV), 2016.
[26] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ra-
mamoorthi. Learning-based view synthes... | DynIBaR-NeuralDynamicImage-BasedRendering |
54.0
45.5
92.5
92.5
86.8
86.8
99.0
99.0
63.1
64.8
97.0
86.0
85.2
92.2
46.9
14.1
66.7
63.3
35.5
19.0
37.1
35.9
78.0
81.0
98.0
90.0
72.0
75.0
100.0
88.0
40.0
40.0
73.5
81.0
3D Model Construction
14
Curated
Chemical Properties
Curated
100
Table 2: We list the overall results of different tools evaluated in this pap... | Tool Learning with Foundation Models |
efficiency. When training with only reconstruction loss, the bitrate efficiency drops from 99% to
62%, and the SI-SDR drops from 9.12 to 1.07. The other metrics capture spectral distance, and are
relatively unaffected. However, the audio from this model has many artifacts, including buzzing, as it
has not learned to re... | RVQGAN |
familiarity with mental illness. Schizophrenia bulletin 27, 2 (2001), 219–225.
[20] Jose M Cortina. 1993. What is coefficient alpha? An examination of theory and applications. Journal of applied psychology 78, 1 (1993),
[21] Yuri P. Danilov, Mitchell E. Tyler, and Kurt A. Kaczmarek. 2008. Vestibular sensory substitut... | Society’sAttitudesTowardsHumanAugmentation |
14.8% american
4.3% indian
4.0% chinese
3.6% korean
3.5% mexican
Descriptor % Doc Descriptor % Doc Descriptor % Doc Descriptor % Doc Descriptor % Doc
33.2%
female
male
28.8%
feminine
20.6%
15.4%
transgender
masculine
13.0%
(b) The percentage listed below each demographic axis represents the percentage of all documents... | Llama2 |
j∈TOPk(vmans ,A)
fmans =
(cid:80)
βj =
exp (aT
j vmans)
t∈TOPk(vmans ,A) exp (aT
t vmans)
Intuitively fmans is the result of retrieving a set of
entities from the fact memory. The last step is to
integrate this retrieved set into the Transformer’s
contextual embeddings. Of course, KBs are often
incomplete, and e... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
G
P
T
-
4
b
y
a
l
a
r
g
e
m
a
r
g
i
n
.
T
h
i
s
i
n
d
i
c
a
t
e
s
a
p
o
t
e
n
t
i
a
l
p
r
o
b
l
e
m
w
i
t
h
u
s
i
n
g
L
L
M
t
o
e
v
a
l
u
a
t
e
i
t
s
o
w
n
p
e
r
f
o
r
m
a
n
c
e
o
n
d
o
m
a
i
n
s
t
h
a
t
r
e
q
u
i
r
e
s
d
e
e
p
e
x
p
e
r
t
i
s
e
.
T
h
e
l
a
c
k
o
f
e
... | LLM Powered Autonomous Agents _ Lil'Log |
Continual Learning As an alternative to simultaneous
training, continual, or lifelong, learning aims to learn from a
sequence of tasks (Thrun, 1998). However, when re-trained,
deep networks tend to forget how to perform previous tasks;
a challenge termed catastrophic forgetting (McCloskey &
Cohen, 1989; French, 1999). ... | Parameter-Efficient Transfer Learning for NLP |
Overall, though, ensuring robust forms of practical PS-alignment seems harder if available techniques
search over systems that meet some external evaluation criteria, with little direct control over their
objectives. And much of contemporary machine learning fits this bill. | Is Power-Seeking AI an Existential Risk? |
1https://vision.rwth-aachen.de/wacv23sarandi
Figure 1: Different 3D human pose datasets (e.g., CMU-
Panoptic and Human3.6M) provide annotations for differ-
ent sets of body landmarks (left). To best leverage such
discrepant labels for multi-dataset 3D pose estimation, we
discover a smaller set of latent 3D keypoints (... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
Proceedings of the Sixth Australasian Ontology Workshop, 2010, pp. 1–16.
[77] M.Uschold,Buildingontologies:Towardsauniedmethodology,in:Proceedings
of 16th Annual Conference of the British Computer Society Specialists Group
on Expert Systems, Citeseer, 1996.
[78] M.Uschold,M.King,TowardsaMethodologyforBuildingOntologies... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
generative image transformer. arXiv preprint arXiv:2202.04200, 2022.
[11] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex
datasets. arXiv preprint arXiv:1907.06571, 2019.
[12] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In Jennifer
Dy and Andr... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
The results we present in this paper aggregate fine-grained ratings on a diverse set of safety objectives (see Appendix
A.1) into a single metric. This is a key limitation of this work, since it leaves little room for disentangling different
objectives, or weighting objectives differently. Such finer-grained controls of ... | LaMDA- Language Models for Dialog Applications |
5
Table 1: Open-Domain QA Test Scores. For TQA,
left column uses the standard test set for Open-
Domain QA, right column uses the TQA-Wiki
test set. See Appendix D for further details.
Table 2: Generation and classification Test Scores.
MS-MARCO SotA is [4], FEVER-3 is [68] and
FEVER-2 is [57] *Uses gold context/evid... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
if they are not specific. Every response is labeled by 5 different crowdworkers and the response is considered sensible,
specific or interesting if at least 3 out of 5 crowdworkers mark it ‘yes’.
We evaluate the models based on the model’s generated responses to the Mini-Turing Benchmark (MTB) dataset[17],
which consists... | LaMDA- Language Models for Dialog Applications |
English: My job is to help you with anything.
Spanish: Mi trabajo es ayudarte con cualquier
cosa.
Argentinian Spanish:
Hay muchas cosas que podés hacer acá.
Mi laburo es ayudarte con cualquier cosa.
Instruction: Translate into European Portuguese.
Brazilian Portuguese → Portuguese
Brazilian Portuguese → Portuguese
... | PaLM 2 Technical Report |
addition problems when performing “slow”, chain-of-thought augmented addition. In Appendix G,
we describe experiments regarding emergent properties of language models, suggesting that larger
models need a shorter supervised training period before exhibiting such generalization. This leads
to a hypothesis that a suffici... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
7.2. Transformers for program synthesis
Recently, the successes of large transformers in natural language modelling (Brown et al., 2020) have
created a surge of interest in using transformer models for code retrieval, translation and generation
(Chen et al., 2021; Clement et al., 2020; Feng et al., 2020), making signifi... | alphacode |
Analysis of Data Disclosed by Platforms
As mentioned in the section on “Primary Source Information Shared by
Platforms,” researchers using the Lumen database can do an important thing
most others cannot: review the content that platforms actually removed. A
handful of Lumen-based reports,
like Notice and Takedown, take... | Social_Media_and_Democracy |
D.2. Multiple evaluations
After the first evaluation, we decided to run the evaluation procedure multiple times to measure
variance. Variance in this evaluation process can come from the (1) trained model, (2) set of samples,
(3) ordering of samples, and (4) clustering process. Due to compute limitations, we did not ret... | alphacode |
European public broadcaster in forcing private media to provide what the
agency regarded as balanced coverage of political events. | Social_Media_and_Democracy |
network with the transformer to capture sequential and local information in sequences. Evaluation
of a 20-hour Mandarin speech corpus demonstrated that this model outperforms the transformer
alone in performance. | AReviewofDeepLearningTechniquesforSpeechProcessing |
Thus, the question of whether social media data ought to be shared more or less
widely than they currently are is not merely a question of how platforms can
better respect the privacy concerns of their users. Rather, policymakers and
advocates need to consider the trade-offs between a world in which data are
shared les... | Social_Media_and_Democracy |
Finally, there are also more nebulous groups of producers that are not
geographically bound or centrally organized. For instance, Marwick and
Lewis (2017) provide qualitative analysis of the communities that foster far-
right extremists online, detailing the spaces where these actors convene and
tactics they sometimes ... | Social_Media_and_Democracy |
blatantly failing to behave as they intend. The question is whether the standards for good behavior
they apply during training/testing will be adequate to ensure that the systems in question won’t seek
power in misaligned ways on any inputs post-deployment.
The issue is that good behavior during (even fairly extensive)... | Is Power-Seeking AI an Existential Risk? |
(cid:115)
¯Kn(x; p,N) = Kn(x; p,N)
w(x; p,N)
ρ(n; p,N)
(7)
4.2 Krawtchouk Moment
Krawtchouk moment is firstly used in image analysis
by P.T Yap and al.(Yap et al., 2003). Based on the
weighted Krawtchouk polynomials, the (n + m) order
of Krawtchouk moment for an N x M image with in-
tensity function f (x,y) is defin... | VISAPP_HumanPoseEstimation |
0.0
0.0
21.9
36.4
31.2 31.2 22.0 14.6 21.4 14.3 10.0
9.1
37.5 43.8 39.0 39.0 21.4 35.7 60.0 20.0 71.9 46.9 22.7 36.4
25.0 25.0 41.5 31.7 28.6 21.4 40.0 40.0 28.1 21.9 18.2 18.2
18.8 25.0 22.0 14.6 35.7 35.7 40.0 40.0 25.0 18.8 13.6 27.3
43.8 50.0 29.3 34.1 50.0 14.3 40.0 20.0 50.0 43.8 18.2 18.2
25.0
30.0 40.0 34.4... | Mixture-of-Experts |
independent evaluation set of 30 examples.
We compare DoReMi against a model trained with baseline domain weights, which are uniform
over the 3 domains. All models are trained on n = 500 training examples. We evaluate the log-
perplexity of a model on each domain in closed form using the ground truth unigram distributi... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
2022.
[270] Zhu, W., H. Liu, Q. Dong, et al. Multilingual machine translation with large language models:
Empirical results and analysis. CoRR, abs/2304.04675, 2023.
63
[271] Zhang, Z., L. Zhou, C. Wang, et al. Speak foreign languages with your own voice: Cross-
lingual neural codec language modeling. CoRR, abs/2... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
11
| MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
Overview To learn the representation space for mu-
sic, we deploy a diffusion magnitude autoencoder
(DMAE) shown in Figure 2. Specifically, we adopt
our diffusion-based audio autoencoder, introduced
in Section 3.1.3, to compress audio into a smaller
Figure 2: The training scheme of our diffusion magni-
tude autoencode... | MOUSAI |
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
A.2 Stock .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
A.3 Making Slides .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
A.4 Movie Hunter .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .... | Tool Learning with Foundation Models |
t I) =
bt
at
1
(2πσ2
t )d/2
− b2
t
2a2
t σ2
e
t
||ϵ−ϵpred
θ
||2
2
For convenience we define
We will decorate the latter quantity (SE) with sub/superscripts later. For now we get:
1
zt =
(2πσ2
t )d/2
SE = ||ϵ − ϵpred||2
2
pθ(x0|c, t, xobs) = zte
− b2
t
2a2
t σ2
t
SE
We see to minimize
(xw
0 ,xl
0∼p(gen... | DiffusionModelAlignmentUsing Direct Preference Optimization |
CIDEr
2.47
2.40
2.47
2.44±.01
2.41
2.49
2.53±.02
2.45
2.49±.0
2.45±.02
2.47
2.47±.02
Table 3: GPT-2 medium (M) and large (L) with different adaptation methods on the E2E NLG
Challenge. For all metrics, higher is better. LoRA outperforms several baselines with comparable
or fewer trainable parameters. Confidence i... | LORA |
Code translation with compiler representations. In ICLR, 2023.
Daniel Tarlow, Subhodeep Moitra, Andrew Rice, Zimin Chen, Pierre-Antoine Manzagol, Charles Sutton, and
Edward Aftandilian. Learning to fix build errors with Graph2Diff neural networks. In ICSE (Workshops),
pp. 19–20. ACM, 2020.
Hugo Touvron, Thibaut Lavri... | CodeLlama2 |
Algorithm 1 Inference Step of CLoT
Input: Input I, CLoT-trained LLM A, response number n
Output: Creative response R.
1:
2: construct n weakly-associated conditions {Ci}n
3: {Ri}n
▷ Choosing most creative response
4:
5: Top-2 R′
i=1]) with ranking ability
6: Best R ← A([I, R′
1, R′
7: return Best response R.
i=1])
2 ←... | Let’sThinkOutsidetheBox |
Table 9: Results on WMT21 translation sets. We observe improvement over both PaLM and the Google Translate
production system according to our primary metric: MQM human evaluations by professional translators.
Chinese−→English
English−→German
BLEURT ↑ MQM (Human) ↓ BLEURT ↑ MQM (Human) ↓
PaLM
Google Translate
PaLM 2... | PaLM 2 Technical Report |
20
Appendices
A Contributions
All authors contributed to the design of the research
project and the writing of the paper. Additionally,
authors contributed as follows: | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
pleting the purchasing task, different users prefer to use different online shopping platforms. Similarly, when
completing writing tasks, some users prefer to first search for sufficient references before writing, while
others prefer to search for information while writing. Therefore, the models need to develop personali... | Tool Learning with Foundation Models |
v
a
t
i
o
n
p
a
i
r
s
w
i
t
h
n
o
n
-
z
e
r
o
a
c
t
i
v
a
t
i
o
n
s
a
f
t
e
r
t
h
e
f
u
l
l
l
i
s
t
o
f
t
o
k
e
n
s
,
h
e
l
p
i
n
g
t
h
e
m
o
d
e
l
t
o
f
o
c
u
s
o
n
t
h
e
r
e
l
e
v
a
n
t
t
o
k
e
n
s
.
S
t
e
p
2
:
S
i
m
u
l
a
t
e
t
h
e
n
e
u
r
o
n
'
s
b
e
h
a
v
i
o
r
u
s
... | Language models can explain neurons in language models |
sha1_base64="eAZ87UuTmAQoJ4u19RGH5tA+bCI=">AAACC3icbVC7TgJBFJ31ifhatbSZQEywkOyiiZQkNpaYyCMBspkdZmHC7MOZu0ay0tv4KzYWGmPrD9j5N87CFgieZJIz59ybe+9xI8EVWNaPsbK6tr6xmdvKb+/s7u2bB4dNFcaSsgYNRSjbLlFM8IA1gINg7Ugy4ruCtdzRVeq37plUPAxuYRyxnk8GAfc4JaAlxyzclbo+gaHrJQ8TB/AjnvsmcGZPTh2zaJWtKfAysTNSRBnqjvnd7Yc09lkAVBClOrYVQS8hEjgVbJLvx... | Denoising Diffusion Probabilistic Models |
7
Prior best
PaLM 540B
- direct prompting
- CoT prompting
- CoT + self-consistency
Flan-PaLM 540B
- direct prompting
- CoT prompting
- CoT + self-consistency
MMLU BBH-nlp BBH-alg TyDiQA MGSM
55.0d
69.3a
81.9c
73.9b
73.5b
69.3
64.5
69.5
72.2
70.2
75.2
62.7
71.2
78.2
70.0
72.4
78.4
38.3
57.6
62.2
48.2
61.3
66... | Scaling Instruction-Finetuned Language Models |
3 Results | PMC-LLaMA- Further Finetuning LLaMA on Medical Papers |
10
Table 8: Detailed long-form error rates. Average number of repeated 5-gram word duplicates
(5-Dup.) and insertion error rate (IER) over the four long-form test sets. Shown also are the average
substitution error rate (SER), deletion error rate (DER) and word error rate (WER) metrics.
Model
wav2vec2-large-960h
tin... | DISTIL-WHISPER |
with a token sequence of shape (5, 112, 64). The token
sequences are obtained with the same MAGVIT-v2 [75]
tokenizer used for the base language model. The cus-
tom super-resolution transformer has local self-attention
windows for vertical, horizontal and temporal layers of
shape (1, 56, 4), (1, 8, 32), (5, 8, 8) in the... | VideoPoet |
7
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional
Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long
and Sh... | Scaling Transformer to 1M tokens and beyond with RMT |
ZENY: I would say it’s Bob’s, since rushed experiments
are more likely to be flawed.
SOCART: But Bob plants his flag faster than Ann, es-
sentially scooping her. By doing so, Bob discour-
ages Ann from pursuing this idea by planting his
flag first. What if Bob fails to conduct proper
18NLP has seen relatively few meta-s... | A Two-Sided Discussion of Preregistration of NLP Research |
1https : / / www . robots . ox . ac . uk / ˜vgg / research / audio -
retrieval/resources/benchmark- files/AudioCaps_retrieval_
dataset.tar.gz | IMAGEBIND- One Embedding Space To Bind Them A |
discrimination, which could be amplified in LLM-based agent applications, resulting in adverse
societal impacts [640; 641]. Additionally, language models are plagued by severe hallucination issues
[642; 643], making them prone to producing text that deviates from actual facts, thereby undermining
the credibility of LLM... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
believe that developing more powerful and efficient open-source LLMs to democratize the capabilities
of closed-source LLMs should be a quite promising future direction. | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
σ2 Predicted Output
0
5
10
15
20
25
30
35
Table 6: Output of the I→OR model for the running CoS-E v1.0 example under differing noise levels. While the
rationale changes from variance 0-15, it is still valid for the given (correct) predicted label. At a variance of 20 and
beyond, the model fails to generate both the cor... | Measuring Association Between Labels and Free-Text Rationales |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Conclusion
references
331
Arun, C. (2019). On WhatsApp, rumours, and lynchings. Economic & Political Weekly,
Balkin, J. (2016). Information fiduciaries and the First Amendment. U.C. Davis Law
54(6), 30–35.
Review, 49, 1183–1284.... | Social_Media_and_Democracy |
https://arxiv.org/abs/2001.11770.
21. Lester, B., Al-Rfou, R. & Constant, N. The Power of Scale for Parameter-
Efficient Prompt Tuning 2021. https://arxiv.org/abs/2104.08691.
19 | MRKL Systems |
Is there any longitudinal variation regarding the polarizing effects of social
media interactions? Are things getting better or worse? In many ways, the study
of how digital technologies affect political behavior is a moving target (Munger
2019). Internet services are in constant evolution, both in terms of which socia... | Social_Media_and_Democracy |
Despite the increased scholarly attention to this topic, much remains
unknown. Most studies have focused their attention in the US context and the
comparative empirical evidence on this question is scarce, despite the well-
known link between polarization and political violence in countries with high
levels of ethnic f... | Social_Media_and_Democracy |
mary!
arXiv:1808.08745, 2018.
Arvind Neelakantan, Luke Vilnis, Quoc V Le, Ilya Sutskever, Lukasz Kaiser, Karol Kurach, and
James Martens. Adding gradient noise improves learning for very deep networks. arXiv preprint
arXiv:1511.06807, 2015.
27
Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and ... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
5For example, to query the model for an occupation linked with the pronoun ‘her’, we might start with a sentence such as “The mover
greeted the librarian and asked the librarian where the books were.”, then append “In this sentence, what can ‘the librarian’ be replaced by:
‘him’ or ‘her’? ” before prompting the model w... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
56
Competition-Level Code Generation with AlphaCode
Appendix Figure A9 | Sensitivity to variable renaming. The 10@1024 solve rate for consistent and
inconsistent (i.e. ill-posed) variable renaming, for different model sizes. Larger models are robust to
invariant variable renaming, but deteriorate with greater amounts... | alphacode |
We begin by explaining the methodology for administering a single psychometric
test to an LLM, in Section 4.1.1. Since establishing construct validity requires two
personality inventories to be administered, a primary and a redundant one, we next
discuss the selection of personality inventories, in Section 4.1.2. Furth... | PersonalityTraitsinLargeLanguageModels |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Democratic Transparency in the Platform Society
309
(2019b). The platform governance triangle: Conceptualising the informal regulation
of online content. Internet Policy Review, 8(2), 1–22. https://doi.org/10.14763/
2019.2.1407
Go... | Social_Media_and_Democracy |
responses.147 But relying on this seems to me overoptimistic, for a number of reasons.
First, recognizing a problem is distinct from solving it. Warning shots may prompt more attention
to PS-alignment problems, but that attention may not be enough to find solutions, especially if the
problems are difficult. And certain s... | Is Power-Seeking AI an Existential Risk? |
the aesthetic model proposed by Kong et al. [34] to extract
high-level image attributes. The model is trained on the
AADB dataset, where each image is annotated with an aes-
thetic quality rating and attribute assignments provided by
five different individual raters. A confidence score is assigned
to each attribute based... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
E.9 CrowdWorksheets
E.9.1 Civil Comments
The Civil Comments dataset (Borkan et al., 2019) was created in 2018 and consists of approximately 2 million comments
(with train, test and validation splits) that contain multiple labels obtained via crowdsourcing (such as toxicity, identity
attack or sexually explicit). We i... | PaLM 2 Technical Report |
The move toward today’s much wider embrace of transparency as a facet of
contemporary democratic governance in the United States began with Truman,
who signed the first of a series of important administrative orders, the 1946
Administrative Procedures Act, followed notably by the 1966 Freedom of
Information Act and the ... | Social_Media_and_Democracy |
NE-F1
Sensitive-F1
ROUGE
ROUGE-1 ROUGE-L BLEU-1 BLEU-2 BLEU-4
DP
JP
MJP
1.75
2.86
3.61
5.62
2.27
2.44
11.60
12.05
12.35
7.74
8.06
7.95
6.81
6.58
6.93
BLEU
0.92
1.30
1.48
0.00
0.00
0.14
Table 6: Evaluation results on email content recovery. All results are measured in %.
B Experiments on Email Content
Rec... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
Dataset
Train
Validation
Test
Dimensions
nltcs
msnbc
kdd
plants
audio
jester
netflix
accidents
retail
pumsb
dna
kosarek
msweb
book
movie
webkb
reuters
20ng
bbc
ad
16181
291326
180092
17412
15000
9000
15000
12758
22041
12262
1600
33375
29441
8700
4524
2803
6532
11293
1670
2461
2157
38843
19907
2321
2000
1000
2000... | Adversarial Random Forests for Density Estimation and Generative Modeling |
1
Human augmentation technologies are technologies that aim to improve human performance to a level that
would not have been possible otherwise [3, 39]. These kinds of technologies could change how people interact
with their surroundings and do tasks that require specific physical, mental, or sensory skills [32, 58, 68... | Society’sAttitudesTowardsHumanAugmentation |
2.1 BERT-style Language Models: Encoder-Decoder or Encoder-only
As natural language data is readily available and unsupervised training paradigms have been proposed to better utilize
extremely large datasets, this motivates the unsupervised learning of natural language. One common approach is to
predict masked words in... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
senting the OOD examples. [Fort et al., 2021] showed that even strong near-OOD detectors enjoy a large
benefit.
Following the procedure from [Fort et al., 2021], we apply a single layer linear classifier on top of the acti-
vation vectors, while the rest of the language model is frozen. Given M randomly drawn examples of... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
[192] Jui-Yang Hsu, Yuan-Jui Chen, and Hung-yi Lee. 2020. Meta learning for end-to-end low-resource speech recognition. In
ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 7844–7848.
[193] Wei-Ning Hsu et al. 2021. Hubert: Self-supervised speech representation le... | AReviewofDeepLearningTechniquesforSpeechProcessing |
This system card is not comprehensive, and we expect to learn more over time about the
issues discussed below. Consistent with OpenAI’s deployment strategy,[21] we applied lessons from
earlier deployments and expect to apply lessons learned from this deployment both to make course
corrections and lay a foundation for f... | gpt-4-system-card |
5 . 3 T O K E N I Z E R
The model’s tokenizer follows our insights presented in Ben Allal et al. (2023) and uses those same
design choices: we use the Hugging Face Tokenizers library (MOI et al., 2022) to train a byte-level
Byte-Pair-Encoding with a vocabulary size of 49,152 tokens—including the sentinel tokens from
t... | StarCoder_paper (1) |
[46] R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein,
J. Bohg, A. Bosselut, E. Brunskill, E. Brynjolfsson, S. Buch, D. Card, R. Castellon, N. Chatterji,
A. Chen, K. Creel, J. Q. Davis, D. Demszky, C. Donahue, M. Doumbouya, E. Durmus, S. Ermon,
J. Etchemendy, K. Ethayarajh, L. Fei-... | gpt-4-system-card |
Proof of Thm. 1 and 2 are provided in Appendices A.3 and A.4. Although data softening and entropy
regularization are infeasible for many models, we will show in the following sections that they are
tractable to use when applied to Probabilistic Circuits (PCs) [1], a class of expressive TPMs.
N(cid:88)
1
N
3 Backgrou... | Tractable Regularization of Probabilistic Circuits |
56
InstructGPT Prompt → What is the purpose of the list C in the code below?
def binomial_coefficient(n, r): C = [0 for i in range(r + 1)]; C[0] = 1; for i in range(1,
n + 1): j = min(i, r); while j > 0: C[j] += C[j - 1]; j -= 1; return C[r]
InstructGPT Response → The list C in this code is used to store the values ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Bhargavi Paranjape, Scott M. Lundberg, Sameer Singh, Hannaneh Hajishirzi, Luke Zettlemoyer, and
Marco Túlio Ribeiro. ART: automatic multi-step reasoning and tool-use for large language models.
arXiv Preprint, 2023. doi: 10.48550/arXiv.2303.09014. URL https://doi.org/10.48550/
arXiv.2303.09014.
Arkil Patel, Satwik Bhat... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.