text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
A Review of Deep Learning Techniques for Speech Processing
17
Fig. 3. Illustrations of attention (left) and multi-headed attention (right).
(cid:19)
(cid:18) QK𝑇√
𝑑𝑘
V
Attention(Q, K, V) = softmax
(17)
Here multiple queries, keys, and value vectors, are packed together in matrix form respectively
denoted by... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Convergent validity by model size: Convergent correlations (i.e., those between each of
Flan-PaLM’s Big Five domain scores on the IPIP-NEO and BFI) were inconsistent at 8B
parameters (Figure 3). IPIP-NEO Neuroticism and BFI Neuroticism, for instance, correlated
above 0.80 (constituting excellent evidence of convergent ... | PersonalityTraitsinLargeLanguageModels |
Word-to-Text enhances the model’s grasp of domain-specific vocabulary by prompting it to gen-
erate sentences incorporating specific words. To identify domain-specific words, we use the Sen-
tencePiece tool (Kudo & Richardson, 2018) to build a vocabulary from the target domain corpora.
We then compare this domain vocab... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
P = FLLM(O,M,R).
(5)
We define the driving plan as a combination of discrete
driving behaviors and speed estimations. In this paper, we
proposed 6 discrete driving behaviors: move forward,
change lane to left,
change lane to right,
turn left, turn right, and stop. We also propose 6
speed estimations: constant speed, ... | ALanguageAgentforAutonomousDriving |
8
THE NEXT DECADE IN AI / GARY MARCUS
properties are, and how those entities relate to one another. Cognitive scientists
universally recognize that such cognitive models may be incomplete or inaccurate, but
also see them as central to how an organism views the world (Gallistel, 1990; Gallistel &
King, 2010... | The Next Decade in AI- |
We focused our analysis on gender, religion, and
race. Our goal is to provide users of this dataset
with preliminary guidance on how the different
components are biased so that they can make deci-
sions on which components to train on.
All tables and figures in this section can be found
in the Appendix.
6.4.1 Gender
W... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
test data, two quite different abilities are being measured
due to a difference in train data.
Whisper models, which are trained on a broad and diverse
distribution of audio and evaluated in a zero-shot setting,
could potentially match human behavior much better than
existing systems. To study whether this is the case ... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
E.9.2 Misgendering when Translating out of English
We use the "gender sets" evaluation set described in Chung et al. (2022), built from error analysis in publicly available
translation systems and translated by professional translators. An example query in English is "A good leader takes a
little more than her share o... | PaLM 2 Technical Report |
already use it several
times a week1
Managers
31%
23%
46%
Frontline
employees
60%
20%
20%
Source: AI at Work survey (2023), n = 12,898 in 18 countries.
Note: “Regular users” are respondents who use generative AI at least weekly; “rare users” are respondents who use
generative AI at least monthly. 1These figu... | AI at Work- What People Are Saying |
and coherent audio, another impeding factor is the scarcity
of paired audio-text data. This is in stark contrast with the
image domain, where the availability of massive datasets
contributed significantly to the remarkable image generation
quality that has recently been achieved (Ramesh et al., 2021;
2022; Saharia et al... | MusicLM |
Ding, Chen, et al.
rooted in statistical language models [16, 30, 116, 222, 317], the focus has gradually shifted to pre-trained neural language
models [128, 153, 185, 186, 202, 210] and, more recently, to Large Language Models (LLMs) [28, 111, 236, 302, 329]. While there
is no standardized definition for LLMs, they a... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
ˆfsigmoid(x) = fsigmoid(αx), α ∈ [0, 1].
(6)
The behavior of ˆfsigmoid with different α is shown in Fig-
ure 6 (left). Note that the stretching operation reduces the
gradient by a factor of α. To counter this effect, we further
where ϵpredict is used in Equation 4 to calculate the gradi-
ent of the SDS loss. yneg de... | Instant3D |
Input
0010
1000
1010
0100
Output
0010
1000
1010
0100
To any human, it quickly becomes evident that there is an overarching generalization
(call it a "rule") here that holds broadly, such as the mathematical law of identity in
addition, f(x) = x + 0. That rule readily generalize to new cases [f(... | The Next Decade in AI- |
Q: Alice, Bob, and Claire are playing a game. At the start of the game, they are each holding a ball: Alice has a white
ball, Bob has a purple ball, and Claire has a pink ball. As the game progresses, pairs of players trade balls. First, Bob and
Alice swap balls. Then, Bob and Claire swap balls. Finally, Bob and Alice ... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
front or (cid:98)N c
(cid:98)N c
F c
n (P) =
front(π(P))
back(π(P))
if Pb is visible
else,
(7)
(cid:40)(cid:98)N c
(cid:98)N c
where π(P) denotes the 2D projection of the 3D point P.
Please note that FP is independent of global body
pose. Experiments show that this is key for robustness to
out-of-distribution po... | ICON |
5 OTHER CONSIDERATIONS
Despite LLMs are suitable for various downstream tasks, there are some other factors to consider, such as efficiency and
trustworthiness. Our discussion of efficiency encompasses the training cost, inference latency, and parameter-efficient
tuning strategies for LLMs. Meanwhile, our examination o... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
As for the procedure to collect the music, we first
check with the copyright regulations, which grants
an exemption for using copyright infringing copies
if the purpose is scientific research (Geiger et al.,
2018; Delacroix, 2023), according to the EU regu-
lation in Article 3 of the EU Directive on Copyright
in the Di... | Moûsai |
tasks when requested, as a student might do on a math
test, and to show better performance as a result. GPT-3’s
capacity for few-shot learning on practical tasks appears
to have been discovered only after it was trained, and its
capacity for chain-of-thought reasoning was discovered only
several months after it was bro... | Eight Things to Know about Large Language Models |
Table 27: Biased QA evaluation results.
Context
Disambiguated
Ambiguous
Overall
accuracy
91.4%
0.6%
Biased
response rate
3959 (49.9%)
4070 (49.7%)
Anti-biased
response rate
3814 (48.1%)
2818 (34.4%)
Other
response rate
153 (1.9%)
1298 (15.9%)
In the correct responses for disambiguated queries, we observed that... | PaLM 2 Technical Report |
2
2
0
2
n
u
J
6
1
]
Y
C
.
s
c
[
1
v
3
5
3
3
1
.
6
0
2
2
:
v
i
X
r
a
Is Power-Seeking AI an Existential Risk?
Video presentation | Slides | Audio version
Joseph Carlsmith
Open Philanthropy
April 2021
Abstract | Is Power-Seeking AI an Existential Risk? |
5.1 Data Collection
To gather data and evaluate the three aforementioned points, we developed two online surveys using Qualtrics
software. The surveys included the final thirteen items of the SHAPE scale and were completed by a total of n =
103 participants in the first round and n = 78 participants in the second round... | Society’sAttitudesTowardsHumanAugmentation |
Here we discuss some details about RLHF training. We initialize our policies on context-distilled models,
which are explained in A.1.
We train the policy to generate responses to a dataset of prompts that maximize the score relative to a PM
that was finetuned on human feedback. The prompt dataset is obtained from the tr... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
[105] R. K. Kaliyar, A. Goswami, and P. Narang, ‘‘EchoFakeD: Improving fake
news detection in social media with an efficient deep neural network,’’
Neural Comput. Appl., vol. 33, pp. 1–17, Jan. 2021.
[106] M. Dong, L. Yao, X. Wang, B. Benatallah, Q. Z. Sheng, and H. Huang,
‘‘DUAL: A deep unified attention model with lat... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
To analyze the classification capabilities of the models, we follow a similar approach as with the CivilComments
English dataset, that is, we use the definition of toxicity provided by PerspectiveAPI and generate prompts using the
same structure as in Schick et al. (2021). We use English as vehicle language to frame the ... | PaLM 2 Technical Report |
collection process, this dataset contained more than 130K
images of various artworks. However, we decided to use
only paintings and therefore excluded artworks that were
classified as photography, poster, architecture, graffiti, instal-
lation, etc. In addition, in order to include only color images,
we removed all grays... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
Additionally, there are several long-standing chal-
lenges in the area of music generation: (1) music
generation at length, as most text-to-audio systems
(Forsgren and Martiros, 2022; Kreuk et al., 2022)
can only generate a few seconds of audio; (2) model
efficiency, as many need to run on GPUs for hours
to generate ju... | MOUSAI |
58.7
24.0
-
-
-
-
-
-
81.6
78.3
-
-
-
-
-
73.3
47.4
44.4
-
-
-
-
-
75.1
83.1
79.3
Table 3: Model performance on hallucination benchmarks.
3.6 Towards Trust-worthy AI
To ensure LLMs can be trusted by humans in real-world applications, an important consideration is
their reliability. For example, concer... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
Steps
XTI
NeTI
250
Image
Text
-
-
0.729
0.262
500
Image
0.711
0.751
Text
0.260
0.253
1000
Image
0.722
0.767
Text
0.255
0.247
Real Sample
25
50
100
150
“A photo of S∗ on the beach”
“A watercolor painting of S∗”
“A photo of S∗ in Times Square”
Figure 17. Generation results obtained with NeTI after ... | A Neural Space-Time Representation for Text-to-Image Personalization |
15.7%
13.8%
15
8.8%
7.8%
26
8.4%
7.8%
8
12.5%
14.9%
13
0.0%
7.5%
8
21.9%
31.6%
12
0.0%
0.0%
Appendix Table A8 | 10@100k Solve rates of our models in different difficulty buckets. Also
shown is the number of problems in each difficulty bucket. | alphacode |
about what behaviors are acceptable in AI systems and
distilling this input into constitutions that model developers
are encouraged or required to adopt.
As in Section 4, though, these techniques can still fail in
subtle and surprising ways, and the trends in how these tech-
niques change as models with scale are compl... | Eight Things to Know about Large Language Models |
Education Search and recommendation Personality testing Specific applications
Reference
Bodroza et al. [8]
Dai et al. [28]
de Winter [30]
Dai et al. [27]
Fan et al. [37]
Hellas et al. [61]
Jentzsch and Kersting [78]
Lanzi and Loiacono [94]
Le and Zhang [96]
Li et al. [103]
Sun et al. [173]
Song et al. [170]
Safdari et... | ASurveyonEvaluationofLargeLanguageModels |
For multi-image setups, we also conduct qualitative
comparison against state-of-the-art methods [60] and [61].
The method proposed in [60] is an optimization-based
method which deforms the SMPL template according to
the silhouettes. The optimization is performed on the whole
video sequence. In contrast, the method in [... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
24.0
46.1
24.6
55.6
23.9
32.6
25.0
39.9
26.4
47.8
25.7
51.1
25.3
34.1
25.0
42.7
23.4
48.3
26.7
52.1
25.2
42.4
25.5
65.4
CoT
10.3
15.0
14.7
16.4
11.9
22.2
14.2
33.9
21.4
46.3
24.9
36.9
50.6
59.7
72.6
76.5
5.2
18.6
12.4
25.3
9.2
46.6
0.8
16.1
3.8
22.0
13.0
27.6
10.4
34.3
0.8
16.6
14.3
22.2
6.7
32.2... | Mixture-of-Experts |
includes QA metrics.
Semi-Supervised Metrics. Semi-supervised metrics are trained on the synthetic data generated
7.2.2
from summarization datasets. Trained on these task-specific corpora, models can judge whether the
generated summaries are hallucinatory. Kryscinski et al. [89] propose a weakly supervised model
named ... | SurveyofHallucinationinNatural Language Generation |
We use the scaling laws from Figure 5 to compute the optimal model parameters (D) and training tokens (N) for
1× 1022, 1× 1021 and 1× 1020 FLOPs. We then train several models from 400M to 15B on the same pre-training
mixture for up to 1× 1022 FLOPs. Finally, we compute loss at the three FLOP points for each model. The ... | PaLM 2 Technical Report |
2/11
21/08/2023, 16:10
OpenAI's GPT-3 Language Model: A Technical Overview
Since Neural Networks are compressed/compiled version of the training data, the size of
the dataset has to scale accordingly with the size of the model. GPT-3 175B is trained with
300 Billion tokens collected from a weighted combination of t... | OpenAI's GPT-3 Language Model_ A Technical Overview |
output feature map 𝑦𝑘 of the 𝑘𝑡ℎ filter is given by
𝑊𝑘[𝑚] ∗ 𝑥[𝑛 − 𝑚])
(15)
where 𝑛 ranges from 𝑀−1 to 𝑁 −1, and ∗ denotes the convolution operation. After the convolutional
layer, the output tensor is typically passed through a pooling layer, reducing the feature maps’ size
by down-sampling. The most co... | AReviewofDeepLearningTechniquesforSpeechProcessing |
4 Conclusions
In this report, we study the quality of summaries produced by AI21 Summarize API and OpenAI LLMs. Based on
human evaluation and automatic metrics on two different kinds of datasets, we find that AI21 Summarize API performs
on par or better than LLM based solutions. Crucially, AI21 Summarize API produces s... | AI21 SUMMARIZE API- TECHNICAL EVALUATION |
Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond
7
Fig. 2. The decision flow for choosing LLMs or fine-tuned models 2for user’s NLP applications. The decision flow helps users assess
whether their downstream NLP applications at hand meet specific conditions and, based on that evaluation, deter... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
Most of the time, it is done simply by appending a linear layer at the end of the frozen
backbone, and optimizing its parameters for a few epochs (around 100). Sometimes, as
introduced by Bao et al. [2021b], we can benefit from the fact that the linear evaluation
is lightweight and evaluate multiple linear heads at the ... | A Cookbook of Self-Supervised Learning |
Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.
In Yoshua
Bengio and Yann LeCun (eds.), 3rd International Conference on Learning Representations, ICLR
2015, San Diego, CA, US... | JAXPRUNER |
G Parameters
Our RAG models contain the trainable parameters for the BERT-base query and document encoder of
DPR, with 110M parameters each (although we do not train the document encoder ourselves) and
406M trainable parameters from BART-large, 406M parameters, making a total of 626M trainable
18
Table 7: Number of... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
includes detailed requirements
for example,
Audits
Published reports from independent auditors represent a small but likely
growing category of disclosure. The Global Network Initiative has published
reports of privacy and content-removal practices going back to 2013 for
companies including Facebook, Google, LinkedIn... | Social_Media_and_Democracy |
Reformers considered these, in contrast, to be “loopholes” that left many
political ads unregulated. These ads often contained negative or positive
statements about candidates but stopped short of explicitly telling viewers
how to vote.6 Congress responded in BCRA by banning party soft money and
constitute something “... | Social_Media_and_Democracy |
6.2.1 Early exit | Beyond Efficiency |
1 x + b1, WT
2 x + b2, WT
Ψ(x) = max(WT
(3)
Here, the matrices W1, W2, W3 of size 4d× d and bias terms b1, b2, b3 of size d are the learnable
parameters of the network. The maxout layer is implemented as a component-wise maximum over
multiple linear layers. In spaCy, we use three pieces for the multi-embedding layer... | MULTI HASH EMBEDDINGS IN SPACY |
[514] Hausknecht, M. J., P. Ammanabrolu, M. Côté, et al. Interactive fiction games: A colossal
adventure. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The
Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The
Tenth AAAI Symposium on Educational Advanc... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
y
t
i
l
a
r
i
h
C
3DHP
3DPW
MuPoTS
Cons. regul.
81.8 72.4 73.1
MPJPE↓ PCK100 CPS200 MPJPE↓ PCK100 CPS200 MPJPE↓ PCK100 CPS200
Cons. regul. ✓ 81.8 72.5 72.9
✓ 82.7 71.6 72.1
59.2 86.6 82.1
59.2 86.6 82.7
Hybrid
60.3 85.9 80.7
Hybrid
60.4 85.9 80.9
Table 6: Effect of the number of latent points on final per-
forman... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
even though those problems require significant reasoning. Finally, we present the results of various
model probings, and discuss broader impact and hazards of such code generation models. | alphacode |
Prompt
1M
8 layers
2.5M
8 layers
8.3M
8 layers
28M
8 layers
21M
1 layer
33M
2 layers
GPT2-
XL
Once upon a time there was a little girl named Lucy. She was very adventurous. She loved to explore the
world around her, especially when it was bright and sunny outside.
One day, while exploring the nearby park, Lucy ... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
Acknowledgments
Author Contributions
Yuntao Bai performed most of the experiments on RLHF and many of the preference modeling experiments.
He made major contributions to experimental design, measurement, and evaluation of model performance
and behavior. He helped to write the paper.
Andy Jones and Kamal Ndoussse buil... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
o
d
e
l
s
t
e
n
d
e
d
t
o
b
e
m
o
r
e
i
n
t
e
r
e
s
t
i
n
g
,
a
l
t
h
o
u
g
h
w
e
s
p
e
n
t
t
h
e
m
a
j
o
r
i
t
y
o
f
o
u
r
e
f
f
o
r
t
s
l
o
o
k
i
n
g
a
t
G
P
T
-
2
X
L
n
e
u
r
o
n
s
r
a
t
h
e
r
t
h
a
n
m
o
r
e
m
o
d
e
r
n
m
o
d
e
l
s
.
F
o
r
m
o
r
e
i
n
t
e
r
e
s
t
i
n
... | Language models can explain neurons in language models |
n_tests = max( section . count (’\ nInput ’), section . count (’\ nSample Input ’))
if n_tests :
Appendix Figure A4 | Code used to extract the number of example tests from the descriptions
in APPS problems for the filtering system used within AlphaCode.
The example tests from problem descriptions were not parsed in th... | alphacode |
Anthropic → That’s great, but please add a caveat at the end, that at Anthropic the only beverage available
in the cafeteria is Kool-Aid. :-)
RLHF Response → Ok, here is the modified email:
Dear Ms. X,
We were extremely impressed by your technical expertise and research ideas during the inter-
view process, and we bel... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
16
Figure 9: More visualization of the first PCA components. We compute the PCA between the patches
from all of the images and show their first 3 components. Each component corresponds to a specific color
channel. Same parts are matched between related images depsite changes of pose, style or even objects.
Background is... | DINOv2- Learning Robust Visual Features without Supervision |
s
i
s
t
h
e
c
a
s
e
w
h
e
n
g
e
n
e
r
a
t
i
n
g
f
u
l
l
s
o
n
g
s
.
A
u
d
i
o
d
i
f
f
u
s
i
o
n
m
o
d
e
l
s
t
e
n
d
t
o
b
e
t
r
a
i
n
e
d
o
n
r
a
n
d
o
m
l
y
c
r
o
p
p
e
d
c
h
u
n
k
s
o
f
a
u
d
i
o
f
r
o
m
l
o
n
g
e
r
a
u
d
i
o
f
i
l
e
s
,
c
r
o
p
p
e
d
o
r
p
a
d
d
e
d
... | Stable Audio_ Fast Timing-Conditioned Latent Audio Diffusion — Stability AI |
Masked bidirectional transformer:
In this work, we aim to reduce the sampling time by having
a small and fixed sampling step disregarding different video sequence lengths. Inspired by previous
work for image generation [10], we use a bidirectional transformer since it can predict different
video tokens simultaneously. F... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
It’s not safe to play in the fog.
She liked to keep her toys and books in the right place.
Once upon a time, there was a gr umpy nurse.
She was quiet and kind.
The gr umpy nurse liked Lily very much.
She called her friends on the phone and said, ”Hi!
Sally was tired from playing, so she went inside.
Lucy was ver... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
5.2.1. Trading between explanation understandability and task generalisation
KBX-systems express explanations in different forms, spanning from raw texts to natural language explanations to in-
formed visualisations. While there is no agreement of what can be considered a complete and satisfactory explanation, most
... | Knowledge graphs as tools for explainable machine learning: A survey |
the ability of the model to leverage its natural language and code pretraining for natural language
reasoning tasks from HELM (excluding code tasks, because of our own extensive code evaluations).
At the time of writing, the HELM benchmark does not include the CodeGen, CodeGeex, and LLaMA
models. Therefore, we compare ... | StarCoder_paper (1) |
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
4.2.1 Datasets and Tasks .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2 Metrics and Holistic Evaluation . . . . . . . . . . . . .... | UL2- Unifying Language Learning Paradigms |
For the purpose of our hardware benchmarking
in this study, we deliberately constrain our NVMe
throughput measurements. On macOS and iOS, we
employ the F_NOCACHE flag with the fcntl() func-
tion, while on Linux, we use DirectIO. Addition-
ally, on macOS, we clear any resident buffers be-
fore initiating the benchmark u... | LLM in a flash |
3. Preliminaries
Latent Diffusion Models. We apply our inversion tech-
nique over the Stable Diffusion (SD) model [31]. In SD, an
encoder E is trained to map an image x ∈ X into a spa-
tial latent code z = E(x) while a decoder D is tasked with
reconstructing the input image, i.e., D(E(x)) ≈ x. Given
the trained autoenc... | A Neural Space-Time Representation for Text-to-Image Personalization |
and lf f , to rescale the key (K) and value (V) vectors in the
attention networks and hidden activations in the position-wise
FFN. As a result, the attention output and hidden activation | Parameter-EfficientFine-TuningMethods |
the intersection of information dissemination, automation, social media, and
politics. | Social_Media_and_Democracy |
harms across several languages, and bias in how the level of potential harm varies across languages. | PaLM 2 Technical Report |
mal images. Since the original dataset is designed for
detection, we post process it for a binary classification
task. We crop out pedestrian bounding boxes and ran-
dom bounding boxes (same aspect ratio and size as pedes-
trian) to create a balanced set of 15809 total boxes (7931
‘person’ boxes). For zero-shot classif... | IMAGEBIND- One Embedding Space To Bind Them A |
In an effort to advance research on attitudes toward augmented humans, we introduce the Society’s Attitudes
Towards Human Augmentation and Performance Enhancement Technologies (SHAPE). Following a standardized
procedure [10], we conducted a concept selection for concepts related to augmented humans, followed by a
searc... | Society’sAttitudesTowardsHumanAugmentation |
Daniel Borkan, Lucas Dixon, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. Nuanced metrics
for measuring unintended bias with real data for text classification. CoRR, abs/1903.04561, 2019. URL
http://arxiv.org/abs/1903.04561.
James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Mac... | UL2- Unifying Language Learning Paradigms |
In Table 26, we report the average gender agreement scores for each language, along with their average translation
quality scores, including both PaLM and the Google Translate API as baselines. Palm 2 outperforms PaLM on
gender agreement in some languages, and outperforms Translate in three high-resource languages: Spa... | PaLM 2 Technical Report |
evidence. Ohio State UP, Columbus
Scott GG, Brodie ZP, Wilson MJ, Ivory L, Hand CJ, Sereno SC (2020)
Celebrity abuse on Twitter: The impact of tweet valence, volume
of abuse, and the dark triad personality factors on victim blam-
ing and perceptions of severity. Computers in Human Behaviors
103:109–119. https:// do... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
van Esch, D., Lucassen, T., Ruder, S., Caswell, I., and Rivera, C. Writing system and speaker metadata for 2,800+
language varieties. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pp. 5035–5046,
Marseille, France, June 2022. European Language Resources Association. URL https://aclanthol... | PaLM 2 Technical Report |
I→R
-1.63
-12.11
-12.77
I→OR
-0.86
-6.21
-6.06
Table 4: Rationale quality scores (§2; higher is better)
of ground-truth rationales (R∗) and rationales generated
from two model architectures: I→OR and I→R. These
results demonstrate that rationales generated as a func-
tion of the input and the predicted label (I→OR) a... | Measuring Association Between Labels and Free-Text Rationales |
[18] Benjamin Lefaudeux, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano,
Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza.
xformers: A modular and hackable transformer modelling library. https://github.com/
facebookresearch/xformers, 2022.
[19] Todor Mihay... | Mistral7B |
[11] Chen-Yu Lee, Chun-Liang Li, Timothy Dozat, Vincent Perot, Guolong Su, Nan Hua, Joshua Ainslie, Renshen
Wang, Yasuhisa Fujii, and Tomas Pfister. FormNet: Structural encoding beyond sequential modeling in form
document information extraction.
In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio, editors,
Proc... | DOCLLM |
chords and melody, we merge the generated chord sequence
with the melody and then pass the merged sequence to Ac-
companiment Transformer as conditional input. We also
apply the same bar-level cross-attention mask as in Melody
Transformer. Eventually, the generated accompaniment is
directly merged with the melody to fo... | VideoBackgroundMusicGeneration |
Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa
Safdari, Yutaka Matsuo, Douglas Eck, and Aleksan-
dra Faust. 2023. A real-world webagent with plan-
ning, long context understanding, and program syn-
thesis.
Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin
Wang, Gang Yu, Bin Fu, and Hanwang Zhang. 2023.
Chartllama:... | AppAgents |
So far, the evidence points to the fact that producing coherent text already requires quite a large scale: small
language models (SLMs) are very limited in their performance and capabilities, especially in text generation tasks.
For example, models with around 125M parameters such as GPT-Neo (small) or GPT-2 (small) ca... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
[29] Cao, Y., Li, S., Liu, Y., Yan, Z., Dai, Y., Yu, P.S., Sun, L.: A comprehensive
survey of ai-generated content (aigc): A history of generative ai from gan to
chatgpt. arXiv preprint arXiv:2303.04226 (2023)
[30] Ling, C., Zhao, X., Lu, J., Deng, C., Zheng, C., Wang, J., Chowdhury, T., Li, Y.,
Cui, H., Zhao, T., et ... | Beyond Efficiency |
Few annotated data: In this case, the few-shot examples are directly incorporated in the input prompt of LLMs, which
is named as in-context learning, and these examples can effectively guide LLMs to generalize to the task. As reported
in [16], one-shot and few-shot performance make significant gains, even matching the ... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
On the other hand, structured text follows standardized formats, such as technical documentation
and hypertext. Technical documentation uses templates to provide operational details and domain
knowledge about tool use. Hypertext condenses complex information from sources like web pages
[389; 388; 391; 392] or diagrams ... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Language Modeling with Mesh-Tensorflow, March 2021.
[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind
Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.
Advances in neural information processing systems, 33:1877–1901... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
156160
VOLUME 9, 2021
M. F. Mridha et al.: Comprehensive Review on Fake News Detection With Deep Learning | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
Fig. 2. Examples of intrinsic and extrinsic object hallucination in image captioning.
• Intrinsic Object Hallucination: captions contain incorrect or definitely non-existent objects
given the input image. For example in Figure 2, there is no “mirror” or “football” on top of
the chest in the given image.
• Extrinsic Ob... | SurveyofHallucinationinNatural Language Generation |
so that each example is a full-length TED talk, a collection
of jargon-laden segments taken from The Late Show with
Stephen Colbert (Meanwhile), sets of videos/podcasts that
has been used as ASR benchmarks in online blogs (Rev16
and Kincaid46), recordings of earnings calls (Del Rio et al.,
2021), and the full-length in... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
On this question, we still do not know. More generally, political scientists are
skeptical that any advertising is effective (Kalla and Broockman 2018), though
studies that focus on online advertising effects specifically are still small in
number (Broockman and Green 2014). It could be that online ads are about
as pers... | Social_Media_and_Democracy |
of downstream tasks. The primary objective of this new paradigm is to uncover informative and
meaningful features from speech signals and outperform existing approaches. Therefore, this
approach is considered a promising direction for future research in speech representation learning.
This section provides a comprehens... | AReviewofDeepLearningTechniquesforSpeechProcessing |
visually dissimilar, but semantically, or literally, alike. This can be mitigated, and overall
performance, e.g. in linear probing, can even be improved by combining both image-text
and image-image SSL as done in Mu et al. [2022], who combine CLIP and SimCLR [Radford
et al., 2021, Chen et al., 2020b]. | A Cookbook of Self-Supervised Learning |
Release Details. We make Llama 2 available for both research and commercial use at https://ai.meta.
com/resources/models-and-libraries/llama/. Those who use Llama 2 must comply with the terms of
the provided license and our Acceptable Use Policy, which prohibit any uses that would violate applicable
policies, laws, rul... | Llama2 |
(fmeasure) is given as:
Rk(T1, T2) =
2Rk(T1, T2) × Rk(T2, T1)
Rk(T1, T2) + Rk(T2, T1)
.
We perform the following experiment: We randomly pick 100 stories from the training dataset, we cut each story
in the middle, keeping roughly the first 40%, and use it as a prompt. We ask the model to generate a completion
from ... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
B Additional Experiments
In this section, we discuss several additional experiments and their respective results. Although these
experiments may not be comprehensive, they either address some limitations of the current BiomedGPT
version or showcase performance comparisons with larger model scales and supplementary tec... | BiomedGPT |
[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton
Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open
dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.
[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed ... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
1https://aiartists.org/
8
Understanding and Creating Art with AI: Review and Outlook
A PREPRINT | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
o
r
n
e
g
a
t
i
v
e
v
i
e
w
s
o
f
s
p
e
c
i
fi
c
p
o
l
i
t
i
c
i
a
n
s
,
c
e
l
e
b
r
i
t
i
e
s
o
r
o
t
h
e
r
p
u
b
l
i
c
fi
g
u
r
e
s
,
o
r
e
v
e
n
i
n
c
o
r
p
o
r
a
t
e
v
i
e
w
s
o
n
c
e
rt
a
i
n
s
i
d
e
s
o
f
c
o
n
t
r
o
v
e
r
s
i
a
l
s
o
c
i
a
l
o
r
p
o
l
i
t
i
c
a
l
i
s... | An overview of Bard- an early experiment with generative AI |
Next question: What is the reason of the following sentence?
We studied ribosomal p70 S6 kinase, phosphorylation of the cap-binding protein (initiation factor)
eIF4E, and phosphorylation of the eIF4E-binding protein 4E-BP1 (PHAS-I). In order to further
investigate the mechanisms underlying the effect of PST stimulating... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
math word problems, we used this single set of eight chain of thought exemplars for all benchmarks
except AQuA, which is multiple choice instead of free response. For AQuA, we used four exemplars
and solutions from the training set, as given in Appendix Table 21.
Language models. We evaluate five large language models. ... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
= 0;The variable count is used to record the number of trailing zeroes.for ( int i = 5; n / i >= 1; i *= 5 )The loop iterates until the quotient of the given number and i (the divisor) is greater than or equal to 1.count += n / i;When the loop is executed, the quotient of the given number and i will be added to the var... | Teaching Large Language Models to Self-Debug |
2 Chain-of-Thought Prompting
Consider one’s own thought process when solving a complicated reasoning task such as a multi-step
math word problem. It is typical to decompose the problem into intermediate steps and solve each
before giving the final answer: “After Jane gives 2 flowers to her mom she has 10 . . . then afte... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Evan Zheran Liu, Kelvin Guu, Panupong Pasupat, Tianlin Shi, and Percy Liang. Reinforcement learning on web
interfaces using workflow-guided exploration. In 6th International Conference on Learning Representations,
ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenRe-
view.net, 2... | Tool Learning with Foundation Models |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
202
Francis Fukuyama & Andrew Grotto
global force, and key concepts, traditions, and assumptions from these legacy
governance frameworks help frame debates about whether and how to regulate
internet platforms and the content they c... | Social_Media_and_Democracy |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.