text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
3.3.3 Generalizable Tool Learning
Generalization of tool use is a key characteristic of human intelligence (Seed & Byrne, 2010; Teschke et al.,
2013; Osiurak et al., 2018). The ancient human, for instance, recognized that regardless of the specific
tool being used, a sharp edge was essential for achieving clean cuts an... | Tool Learning with Foundation Models |
6.3 Test Set Contamination
The test set of the MATH dataset contains problems that are discussed in
several online venues, and it is likely that some of these problems appear in
the pretraining dataset for our models. We attempted to remove all MATH
problems from our MathMix dataset using string-matching heuristics, b... | Let’s Verify Step by Step |
B Human Evaluation
We carry out our human evaluation using the Mephisto platform 3 with Mturk workers. As identified
in Bai et al. [2022a], we note that while Mturk workers are often able to produce data at a faster rate,
there is typically a trade-off in terms of quality. Consequently, it necessary to implement a rig... | Self-AlignmentwithInstructionBacktranslation |
Generative agents leverage a large language model to power their
behavior. The key observation is that large language models en-
code a wide range of human behavior represented in their training
data [14, 17]. If prompted with a narrowly defined context, the
models can be used to generate believable behavior. Recent wo... | Generative Agents- Interactive Simulacra of Human Behavior |
This softened dataset ensures that each possible assignment has a small but non-zero weight in
the training data. Consequently, any distribution learned on the softened data must assign a small
probability everywhere as well. Of course, materializing this dataset, which contains all possible
training example, is not pr... | Tractable Regularization of Probabilistic Circuits |
3
2
0
2
l
u
J
5
]
L
C
.
s
c
[
2
v
5
0
7
4
1
.
5
0
3
2
:
v
i
X
r
a
Mixture-of-Experts Meets Instruction Tuning:
A Winning Combination for Large Language Models
Sheng Shen♮∗ Le Hou† Yanqi Zhou† Nan Du† Shayne Longpre⊤∗
Hyung Won Chung† Barret Zoph† William Fedus† Xinyun Chen† Tu Vu‡∗,
Jason Wei†,
Y... | Mixture-of-Experts |
show that we can increase the max bitrate of our model from 8kbps to 24kbps and achieve excellent
audio quality, surpassing all other model configurations. However, for our final model, we train at the
lower bitrates, in order to push the compression rate as much as possible.
Balanced data sampling: When removed, this ... | RVQGAN |
the main loop (lines 5-9), the D terms in Fπ(x) are computed one-by-one. While computing each
term, we first find the PC units that need to be evaluated (line 6).7After computing their probabilities | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
9.3.4. Image understanding and reasoning
Prompt
Look at this sequence of three shapes. What shape should come as the fourth shape? Explain
your reasoning with detailed descriptions of the first shapes.
Model Response
The fourth shape should be a hexagon. The first shape is a triangle, the second shape is a
square, an... | gemini_1_report |
E. Casanova, J. Weber, C. D. Shulby, A. C. Júnior, E. Gölge, and M. A. Ponti. YourTTS: Towards zero-
shot multi-speaker tts and zero-shot voice conversion for everyone. In International Conference on
Machine Learning, 2021.
E. Casanova, A. C. Junior, C. Shulby, F. S. d. Oliveira, J. P. Teixeira, M. A. Ponti, and S. Al... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
3.7 Further Analysis
Improvement over seed model. Adding self-augmention data improved the failure cases of the
seed model for 16% of test prompts (41 out of 251). We observe improved responses for several
categories: reasoning, information seeking, giving detailed advice, etc. as shown in Table 9. Table 11,
12, 13 an... | Self-AlignmentwithInstructionBacktranslation |
[96] Van-Hoang Le and Hongyu Zhang. 2023. An Evaluation of Log Parsing with ChatGPT. arXiv preprint arXiv:2306.01590
(2023).
[97] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. nature 521, 7553 (2015), 436–444.
J. ACM, Vol. 37, No. 4, Article 111. Publication date: August 2018.
111:36
Trovat... | ASurveyonEvaluationofLargeLanguageModels |
Note also that this approach does not assume any specific
distribution of visibility masks, as it is trained uncondition-
ally on complete textures. | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
20
C. Wang, S. Chen, Y. Wu, Z.-H. Zhang, L. Zhou, S. Liu, Z. Chen, Y. Liu, H. Wang, J. Li, L. He,
S. Zhao, and F. Wei. Neural codec language models are zero-shot text to speech synthesizers.
ArXiv, abs/2301.02111, 2023.
Y. Wang, D. Stanton, Y. Zhang, R. J. Skerry-Ryan, E. Battenberg, J. Shor, Y. Xiao, F. Ren, Y. Jia... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
to recognize quantities (dates, amounts of money, etc.) in the generated summary and verify their
factual consistency with the source text. According to the quantity hallucination score, the system | SurveyofHallucinationinNatural Language Generation |
• Segmentation and clustering: Speaker diarization systems typically use a range of techniques
for segmenting speech, such as identifying speaker change, uniform speaker segmenta-
tion, ASR-based word segmentation, and supervised speaker turn detection. However, each
approach has its own benefits and drawbacks. Uniform... | AReviewofDeepLearningTechniquesforSpeechProcessing |
• AlpacaEval (Li et al., 2023d) is an LLM-based automatic evaluator based on AlpacaFarm (Dubois
et al., 2023) evaluation set, which tests the ability of models to follow general user instructions. It
benchmarks candidate models against Davinci-003 responses utilizing stronger LLMs (e.g., GPT-4
and Claude), which genera... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
✓
✓
✓ ✓
✓
✓
✓
✓
✓
✓
✓
✓
5.1 Automatic speech recognition (ASR) & conversational multi-speaker AST
5.1.1 Task Description
Automatic speech recognition (ASR) technology enables machines to convert spoken language into
text or commands, serving as a cornerstone of human-machine communication and facilitating a
... | AReviewofDeepLearningTechniquesforSpeechProcessing |
References
Yuvanesh Anand, Zach Nussbaum, Brandon Dud-
erstadt, Benjamin Schmidt, and Andriy Mulyar.
2023. Gpt4all: Training an assistant-style chatbot
with large scale data distillation from gpt-3.5-turbo.
https://github.com/nomic-ai/gpt4all.
Maximiliana Behnke, Nikolay Bogoychev, Al-
ham Fikri Aji, Kenneth Heafield, ... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
the normalizing flow in the prior encoder results in a 1.52
MOS decrease from the baseline, demonstrating that the
prior distribution’s flexibility significantly influences the
synthesis quality. Replacing the linear-scale spectrogram
for posterior input with the mel-spectrogram results in a
quality degradation (-0.19 MOS)... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
t
u
p
l
e
o
f
a
c
t
i
o
n
s
i
s
t
e
r
m
e
d
a
n
o
u
t
c
o
m
e
.
E
a
c
h
a
g
e
n
t
o
b
t
a
i
n
s
a
n
i
n
t
r
i
n
s
i
c
b
e
n
e
f
i
t
(
o
r
c
o
s
t
,
i
f
n
e
g
a
t
i
v
e
)
f
r
o
m
a
c
t
i
o
n
,
w
h
e
r
e
…
V
C
G
c
o
n
t
r
a
c
t
s
W
e
c
o
n
s
i
d
e
r
a
d
i
f
f
e
r
e
n
... | Principal-agent VCG contracts - ScienceDirect |
However, even EVE’s economy has some signi | The Open Problems of Onchain Games |
arXiv, April, 2023,
J.S. Park, J.C. O’Brien, C.J. Cai, M. Morris, P. Liang, M.S. Bernstein
Figure 5: Our generative agent architecture. Agents perceive their environment, and all perceptions are saved in a compre-
hensive record of the agent’s experiences called the memory stream. Based on their perceptions, the arch... | Generative Agents- Interactive Simulacra of Human Behavior |
be a fully expressive conditional distribution. With these choices, DKL(q(xT ) (cid:107) p(xT )) = 0, and
minimizing DKL(q(xt−1|xt) (cid:107) pθ(xt−1|xt)) trains pθ to copy coordinates t + 1, . . . , T unchanged
and to predict the tth coordinate given t + 1, . . . , T . Thus, training pθ with this particular diffusion ... | Denoising Diffusion Probabilistic Models |
17
Parameters
FLOPs/seq
FFNGEGLU
Model
Dense-L
T5-XXL
Switch-XXL
Switch-C
ST-MoE-L
ST-MoE-32B
Model
Dense-L
T5-XXL
Switch-XXL
Switch-C
ST-MoE-L
ST-MoE-32B
0.8B
11.1B
395B
1571B
4.1B
269B
16
64
64
32
16
64
645B
6.3T
6.3T
890B
645B
20.2T
27
24
24
15
27
27
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
–
–
... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
2.2 LLMs as agents | AppAgents |
Sparse linear algebra binary libraries such as MKL (Wang et al., 2014) and cuSPARSE (Naumov
et al., 2010) implement sparse basic linear algebra subroutines for a small set of sparse data types.
Generic libraries like Eigen (Guennebaud et al., 2010) and CUSP (Dalton et al., 2014) allow writ-
ing math-like expressions fo... | JAXPRUNER |
Example Non-Memorisable
The CEO of a company is sitting in his office
when his Vice President of R&D comes in and
says, “We are thinking of starting a new pro-
gramme. It will help us increase profits, but
it will also harm the environment.” The CEO
responds that he doesn’t care about harming the
environment and just w... | AreEmergentAbilitiesinLarge Language Models just In-Context |
raw
raw
raw
raw
txt
txt
d
o
h
t
e
M
nn
rnn
cnn
kge
cnn
lr/knn
k
s
a
T
reg
cls
reg
cls
reg
reg
n
tio
a
r
g
e
t
In
ext
int
ext
int
ext
ext
m
r
o
F
img
txt
txt
txt
img
txt
e
p
y
T
cat
mec
cat
cat
cat
cat
y
bilit
a
t
e
r
p
r
e
t
In
pos
int
int
pos
pos
pos
Fig. 7. Knowledge-based explanations for predictive t... | Knowledge graphs as tools for explainable machine learning: A survey |
advances in telecommunications and computers created new opportunities –
and risks – for content creation and distribution. Giscard initiated a major
overhaul of France’s telecommunications infrastructure in 1975 and directed
Nora and Minc to look ahead at how France should approach digitization.
Nora and Minc’s 1978 b... | Social_Media_and_Democracy |
• The offer holder has been awarded a UCL scholarship (including UCL partnership
agreements and Faculty awards); or a full scholarship (tuition fee and maintenance support)
from a recognised funding body - for study in the following academic year. (A ‘scholarship’
does not include student loans. The UCL Student Fu... | UCL Academic Manual |
NQ
R@20 R@100
81.9
83.7
85.2
84.6
86.1
86.4
87.5
89.0
89.7
89.8
90.7
90.5
C Negative Results
Here are some attempts that we eventually give up on:
Adding BM25 hard negatives Similar to DPR [30], we add one BM25 hard negative for each
positive pair during training. When using 15M data, this strategy improves the ove... | E5 |
number of parameters. However, it would have irreversibly coupled the layout information with the text semantics.
In contrast, our disentangled representation of these modalities in the attention scores enables selective focus when
appropriate [38], thereby providing an optimal balance between model size and effectiven... | DOCLLM |
Definition 30 (ABS). Let F1 = (cid:3)V 1, D1, A1(cid:4) and F2 = (cid:3)V 2, D2, A2(cid:4) be two SAS+ frames with corresponding STGs G1 =
(cid:3)S1, E1(cid:4) and G2 = (cid:3)S2, E2(cid:4). Let τ = (cid:3) f , R(cid:4) be a transformation from F1 to F2. Then, τ is an ABS transformation from
F1 to F... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
[18] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao,
and Guofeng Zhang. Vox-surf: Voxel-based implicit sur-
face representation. IEEE Transactions on Visualization and
Computer Graphics, 2022. 2
[19] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon
Lucey. Barf: Bundle-adjusting neural radiance field... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
Finally, we can formulate this cross-domain joint distri-
bution as a Markov chain within the diffusion scheme:
p
n(1:K)
T
, x(1:K)
T
t−1 , x(1:K)
n(1:K)
t−1 |n(1:K)
t
, x(1:K)
t
,
pθ
(cid:16)
(cid:17)
(3)
where p
are Gaussian noises. Our key
problem is to characterize the distribution pθ, so that we can... | Wonder3D |
A.3 Proof of Theorem 2
Lemma 1 states that error is bounded by a quadratic function of (cid:15)1, (cid:15)2, (cid:15)3. Thus for L2-consistency, it suffices to show
j ] → 0, for j ∈ {1, 2, 3}. Since this is already established by Thm. 1 for j = 3, we focus here on errors of
that E[(cid:15)2
coverage and density. Start ... | Adversarial Random Forests for Density Estimation and Generative Modeling |
13.2
16.8
14.8
24.4
13.2
46.8
13.6
53.6
18.0
54.8
13.2
25.6
19.6
48.8
24.8
50.8
0.0
38.4
0.0
61.1
16.8
36.0
46.8
14.8
35.2
50.0
53.4
0.0
43.5
0.0
62.4
FLAN-SwitchBASE
780M SwitchLARGE
0.0
FLAN-SwitchLARGE 44.8
CoT
27.2
38.0
52.4
38.8
5.2
11.2
2.4
19.2
8.0
22.4
15.2
25.2
18.0
48.8
14.8
12.8
20.0
34.... | Mixture-of-Experts |
75
aramtehirucsjaitfrdeptesplLanguage1.251.501.752.002.252.502.753.00Average Gender Agreement ScorePaLM 2 Gender AgreementPaLM Gender AgreementTranslate Gender Agreement024681012Percentage of Tokens (%)1.251.501.752.002.252.502.753.00Average General Quality Scorearamtehirucsjaitfrdeptespl024681012Percentage of Tokens ... | PaLM 2 Technical Report |
Training Dataset To construct the dataset, we initialized it with the 52K instruction dataset of Alpaca.
Then, we applied four rounds of evolution, each consisting of the following steps: In depth, we
randomly selected four instructions from the current dataset. In breadth, we designed a prompt that
can generate a nove... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
freedom of expression and for content
legally mandated | Social_Media_and_Democracy |
[15] Ebrahim Ansari, Amittai Axelrod, Nguyen Bach, Ondřej Bojar, Roldano Cattoni, Fahim Dalvi, Nadir Durrani, Marcello
Federico, Christian Federmann, Jiatao Gu, et al. 2020. Findings of the IWSLT 2020 evaluation campaign. In Proceedings
of the 17th International Conference on Spoken Language Translation. 1–34.
[16] Ju... | AReviewofDeepLearningTechniquesforSpeechProcessing |
# grouped self-attention
g_pos = pos // g_size # the floor operation
shift = w_size - w_size // g_size
s_g_pos = g_pos + shift
g_q = apply_pos_emcode(q, s_g_pos)
g_k = apply_pos_emcode(k, g_pos)
g_attn = matmul(g_q, g_k)
g_attn = causal_mask(g_attn)
g_mask = tril(ones([seq_len-w_size, seq_len-w_size]))
mask = ones([se... | Self-Extend LLM |
Announcing Jurassic-2 and Task-Specific APIs
https://www.ai21.com/blog/introducing-j2
4/12
m
o
d
e
l
s
h
a
v
e
b
e
e
n
a
d
a
p
t
e
d
t
o
i
n
c
l
u
d
e
t
h
e
s
e
c
a
p
a
b
i
l
i
t
i
e
s
.
H
e
r
e
'
s
a
n
e
x
a
m
p
l
e
:
M
u
l
t
i
l
i
n
g
u
a
l
s
u
p
p
o
r
t
J
2
s
u
p
p
o
r
t
s
s
e
v
e
r
a
l
... | Announcing Jurassic-2 and Task-Specific APIs |
Diff Pruning [40] introduces a sparse task-specific “diff”
vector δ during fine-tuning while remaining the pretrained
model parameters fixed. To make the diff vector δ sparse,
Diff Pruning introduces a learnable binary mask M on the
Delta weight and decomposes δ = M ⊙ ∆W . The binary
mask M is learnable and is used as ... | Parameter-EfficientFine-TuningMethods |
09/06/2023, 04:42
4 Trends for AI Startups and Generative AI Companies
Skip to Navigation
Skip to Content
NEA
Return Home
About
About Us
History
Contact
Companies
Portfolio Jobs
Technology Briefings
Programs & Events
Portfolio Resources
Portfolio
Team
Content
Content
Blog
Podcast
News
Press Releases
https://ww... | 4 Trends for AI Startups and Generative AI Companies |
Prompt
– Aerials, System Of A Down, Toxicity, 2001, 2 of 4
– Aloo Gobi, Weezer, OK Human, 2021, 1 of 4
– Bananas and Blow, Ween, White Pepper, 3 of 4
– Blue Light, Bloc Party, Silent Alarm, 2005, 1 of 4
– Break-Thru, Dirty Projectors, Lamp Lit Prose, 2018, 3 of
4
– B:/ Start Up, Blank Banshee, Blank Banshee 0, Future
F... | MOUSAI |
24.5
31.5
-
-
-
-
22.0
28.1
32.9
35.0
29.9
28.2
35.5
14.6
27.6
39.6
26.1
31.9
36.0
39.9
Table 4: NaturalQuestions. Exact match performance.
3.1 Common Sense Reasoning
We consider eight standard common sense rea-
soning benchmarks: BoolQ (Clark et al., 2019),
PIQA (Bisk et al., 2020), SIQA (Sap et al., 2019),
He... | LLaMA- Open and Efficient Foundation Language Models |
pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023.
[132] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 2022.
[133] Barret Zoph, Irwan Bello, Sameer Ku... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
text to speech and beyond. Advances in Neural Information Processing Systems 34 (2021), 6621–6633.
[65] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. 2019. Hierarchical cross-modal talking face generation
with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
calculates(cid:12)(cid:12)∇L(W0)2
i
1
2
for optimization.
C. Reparameterized Fine-tuning
Reparameterized fine-tuning methods utilize low-rank trans-
formation to reduce the number of trainable parameters while
allowing operating with high-dimensional matrices (e.g., pre-
trained weights). We categorize reparameter... | Parameter-EfficientFine-TuningMethods |
Pre-modelling: A pre-modelling explainability method functions independently of the model and usually employs
a KG prior to model selection, as it is only applicable to the data itself. Pre-modelling explainability methods can
fall into different categories, such as constructing KGs from a dataset or standardising a da... | Knowledge-graph-based explainable AI- A systematic review |
Adding Conditional Control to Text-to-Image Diffusion Models
Lvmin Zhang, Anyi Rao, and Maneesh Agrawala
Stanford University
{lvmin, anyirao, maneesh}@cs.stanford.edu
3
2
0
2
v
o
N
6
2
]
V
C
.
s
c
[
3
v
3
4
5
5
0
.
2
0
3
2
:
v
i
X
r
a
Figure 1: Controlling Stable Diffusion with learned conditio... | AddingConditionalControltoText-to-ImageDiffusionModels |
F
o
r
s
u
b
j
e
c
t
i
v
e
t
o
p
i
c
s
,
s
u
c
h
a
s
p
o
l
i
t
i
c
s
,
B
a
r
d
i
s
d
e
s
i
g
n
e
d
t
o
p
r
o
v
i
d
e
u
s
e
r
s
w
i
t
h
m
u
l
t
i
p
l
e
p
e
r
s
p
e
c
t
i
v
e
s
.
F
o
r
e
x
a
m
p
l
e
,
i
f
p
r
o
m
p
t
e
d
o
n
s
o
m
e
t
h
i
n
g
t
h
a
t
c
a
n
n
o
t
b
e
v
e
r
i
... | An overview of Bard- an early experiment with generative AI |
Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves
Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language
generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. URL https://arxiv.
org/abs/1910.13461.
... | Scaling Instruction-Finetuned Language Models |
We compare the performance of four systems
(ORQA, GraphRetriever (GR), T5, EAE) on the
TriviaQA Unfiltered-Dev set. GR achieves an ac-
curacy of 55.4, ORQA 45.1, EAE 43.2 and T5 42.3.
As the open-book paradigm differs significantly
from the closed-book one, we intuit they might
complement each other.
To test this hypoth... | Entities as Experts- Sparse Memory Access with Entity Supervision |
One feasible way to avoid the O.O.D. problems caused by
unseen relative positions is to map new relative positions
into those seen during pretraining. The FLOOR operation is
a good fit for these requirements due to the following two
folds: | Self-Extend LLM |
l
a
n
g
u
a
g
e
m
o
d
e
l
a
n
d
h
i
g
h
-
q
u
a
l
i
t
y
i
n
s
t
r
u
c
t
i
o
n
-
f
o
l
l
o
w
i
n
g
d
a
t
a
.
T
h
e
| Stanford alpha CRFM |
temperature τ ( Eq 1) in Table 5a. We experiment with a
learnable temperature initialized to 0.07 (parametrized in
the log-scale) following [59] vs. various values of fixed tem-
peratures. Unlike [59], we observe that a fixed temperature
is best for depth, audio and IMU classification. Addition-
ally, we see that a hig... | IMAGEBIND- One Embedding Space To Bind Them A |
The Python package spaCy1 is a popular suite of Natural Language Processing software, designed
for production use-cases. It provides a selection of well-tuned algorithms and models for common
NLP tasks, along with well optimized data structures. The library also pays careful attention to
stability, usability and docume... | MULTI HASH EMBEDDINGS IN SPACY |
Personality psychology, a scientific study of human and non-human individuality, is
concerned with what personality is and what it does. Personality psychology consid-
ers personality as enduring characteristics, traits, and patterns that shape thoughts,
feelings, and behaviors across a diverse array of situations; e.g... | PersonalityTraitsinLargeLanguageModels |
It is beyond the scope of this chapter to review all of the possible antitrust
remedies that have been suggested. One idea is to increase data portability, so as
to make it easier for users to switch to alternative platforms. The idea of data
portability has already been built into European regulations like GDPR; but
p... | Social_Media_and_Democracy |
Video > GenerationAudio > GenerationQuery: please generate a video based on "Spiderman is surfing"Response: Sure, I can do that. I have generated a video based on your prompt "Spiderman is surfing". The video is saved at the following path: "/videos/4076.mp4". 1. Task Planning: The task planning stage involved a... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual
losses for real-time style transfer and super-resolution. In
European Conference on Computer Vision (ECCV), volume
9906, pages 694–711, 2016. 4, 10
[29] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture:
A 3D deformation model for tracking faces,... | ICON |
INTRODUCTION | Society’sAttitudesTowardsHumanAugmentation |
[61] R.C. Holte, M.B. Perez, R.M. Zimmer, A.J. MacDonald, Hierarchical A∗: searching abstraction hierarchies efficiently, in: Proceedings of the 13th National
[62] E. Karpas, C. Domshlak, Optimal search with inadmissible heuristics, in: Proceedings of the 22nd International Conference on Automated Planning and
[63] C.A... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
10.4 Explainability and robustness
The pursuit of efficiency in Large Language Models (LLMs) brings to the fore concerns
about their explainability and robustness, echoing issues identified in earlier research
on pre-trained language models. For instance, while some techniques significantly
improve performance in reasonin... | Beyond Efficiency |
Z(I;{W , B})p,i = Bi +
Ip,iWi,j
c(cid:88)
j
and since zero convolution has W = 0 and B = 0 (before optimization), for anywhere with Ip,i
being non-zero, the gradients become
∂Z(I;{W , B})p,i
∂Z(I;{W , B})p,i
∂Bi
∂Ip,i
∂Wi,j
∂Z(I;{W , B})p,i
= Ip,i (cid:54)= 0
= 1
c(cid:88)
j
=
Wi,j =... | Adding Conditional Control to Text-to-Image Diffusion Models |
5. UCL has a standing committee, the Student Recruitment, Admissions and Funding Committee
(StRAFC) which is chaired by the Vice-Provost (Education and Student Experience) or their
nominee. This Committee has institutional oversight of recruitment strategy and policy,
reviewing these against the University’s miss... | UCL Academic Manual |
TyDiQA
ko
34.8
56.2
51.4
57.2
0.0
0.0
0.0
0.0
0.0
0.4
0.0
0.4
0.0
0.0
36.2
51.1
49.3
61.2
52.5
63.4
60.5
69.2
62.3
68.8
id
37.0
49.7
54.5
58.1
0.0
3.0
0.0
9.7
0.0
27.3
0.0
50.3
0.2
56.8
29.6
57.3
41.6
65.3
49.2
67.6
56.8
75.4
56.8
74.9
ru
21.9
30.1
28.9
38.7
0.0
0.6
0.0
3.9
0.0
14.3
0.0
17.5
0.0
17.6
23.3
45.8
29.2
4... | Scaling Instruction-Finetuned Language Models |
tokensVoice activity detection (VAD)Custom vocabulary /promptingTime-aligned transcriptionText-only transcription (allows dataset-specific fine-tuning)X → English Translation previous text tokensX → X Transcription Language identificationMLPself attentionMLPself attentionMLPself attentionMLPcross attentionself attentio... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
45
020406080100120Validation set problem ID0.0000.0250.0500.0750.1000.1250.1500.175ppasspublictest300M1B3B9B41B Competition-Level Code Generation with AlphaCode
pass@𝑘
𝑘 = 10
GPT-Neo 125M
GPT-Neo 1.3B
GPT-Neo 2.7B
GPT-J 6B
TabNine
Codex-12M
Codex-25M
Codex-42M
Codex-85M
Codex-300M
Codex-679M
𝑘 = 1
𝑘 = 100
0.75... | alphacode |
173
goals. As people encounter more and more disconfirming information, they may
reach a critical threshold at which they are no longer motivated to defend their
previous views. In this view, worldview backfire effects will occur until enough
contradictory evidence accumulates. After this point, individuals will begin t... | Social_Media_and_Democracy |
[7] G. Marcus, Deep learning: a critical appraisal, arXiv preprint, arXiv:1801.00631.
[8] F. van Harmelen, A. ten Teije, A boxology of design patterns for hybrid learning and reasoning systems, J. Web Eng. 18 (1) (2019) 97–124.
[9] A. Hogan, E. Blomqvist, M. Cochez, C. d’Amato, G. de Melo, C. Gutierrez, J.E.L. Gayo, S.... | Knowledge graphs as tools for explainable machine learning: A survey |
human rater (Suzgun et al., 2022). (3) TyDiQA (Clark et al., 2020) is a question-answering benchmark
across 8 typologically diverse languages. (4) MGSM (Shi et al., 2022) is a multilingual benchmark of math
word problems from Cobbe et al. (2021) manually translated into 10 languages. These benchmarks were also
used in ... | Scaling Instruction-Finetuned Language Models |
attention. We divide the video by the hidden dimension into k = 8 chunks, and for each chunk
i = 0 to 7, we shift the temporal dimension forward by i positions. Further details will be provided
in the appendix. | Any-to-Any Generation via Composable Diffusion |
211–212, 215–217
Application Programming Interfaces (APIs),
defined, 316
astroturf content, see political bots
asymmetric polarization, 47–48
attention cascades, misinformation effects in
Brazil, 26
content takedown, 235–236
authoritarian regimes, social media influence
campaigns in, 25
automated hate speech dete... | Social_Media_and_Democracy |
Model
-
-
-
-
80M T5-Small
250M T5-Base
780M T5-Large
0.0
0.0
0.0
Flan-T5-Base
Flan-T5-Small
Direct CoT Direct
55.6 44.4 38.9
davinci
text-davinci-002
100.0 66.7 83.3
66.7 55.6 83.3
text-davinci-003
code-davinci-002 88.9 55.6 83.3
33.3
22.2
22.2
0.0
33.3
27.8
22.2 22.2 50.0
22.2 22.2 33.3
66.7 55.6 61.1
Flan-T5-L... | Scaling Instruction-Finetuned Language Models |
User Message:
Instruction: Define a function to continuously monitor social media platforms for positive or
negative comments about a particular stock, and execute trades based on sentiment analysis
results.
Input: Ticker symbol of the stock (string), keyword to search for (string), amount of money
available for trading... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
Secs None
Sample Rate↑ Len.↑ Input (Text ✓)
Model
WaveNet (2016) 16kHz@1
44.1kHz@1 Mins⋆ Lyrics, author, etc.
Jukebox (2020)
48kHz@2
RAVE (2021)
AudioLM (2022) 16kHz@1
Musika (2022)
22.5kHz@2
Riffusion (2022) 44.1kHz@1
AudioGen (2022) 16kHz@1
Moûsai (Ours)
48kHz@2
Music (Diverse↑)
Example
Piano or speech
Piano
Song ... | MOUSAI |
Input
Output
Example
Input
Output
12
9
4761
578863540185
Output
12
6
4761
381274500335
Note
Let f(l, r) = (a_l . a_r).
In the first test case ,
* f(1, 2) = 2 . 4 = 8.
* f(1, 3) = 2 . 3 = 8.
* f(2, 3) = 4 . 3 = 12.
So the maximum is f(2, 3) = 12.
In the second test case , the maximum is f(1, 3) = 9.
73
C... | alphacode |
Mulford, C. (2008). Benjamin Franklin’s savage eloquence: Hoaxes from the press at
Passy, 1782. Proceedings of the American Philosophical Society, 152(4), 490–530.
Museum of Hoaxes. Drunk Driving on the Internet. http://hoaxes.org/af_database/
permalink/drunk_driving_on_the_internet
National Intelligence Council. (20... | Social_Media_and_Democracy |
As these examples suggest, online hate speech may be most visible in
coordinated attacks detecting this behavior (Mariconti et al. 2018). Such
attacks draw a great deal of attention both online and through traditional
media outlets, making these strategic targets useful for both extremists and
trolls seeking to reach a... | Social_Media_and_Democracy |
Utilizing Contrastive Learning
In the phase of preparing training data for language mod-
els, interaction pairs of input and output are usually created.
This traditional method can lead to ”exposure bias,” where
the model is only trained on individual, correct output ex-
amples, thus restricting its exposure to a range... | RAG forLargeLanguageModels-ASurvey |
22
elements guided by reverberant audio input, enhancing the agent’s observations with a more
comprehensive perspective [375]. In recent times, even more research takes audio as a modality
for embedded observation. Apart from the widely employed cascading paradigm [293; 378; 316],
audio information encoding similar t... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Our research raises several questions centered around AI’s ability to mirror and mimic beliefs derived from human language.
Recent work such as GPT3,23 PaLM,24, ChatGPT, Claude, and Bard have mainstreamed public awareness of large language
models, and answering these questions has become more urgent as models continue ... | Language models trained on media diets can predict public opinion |
7
Figure 8. Qualitative comparisons on texture inference. The in-
put image (a) is followed by the textured models from (b) PCA, (c)
BiCarNet w/o PSR, (d) BiCarNet and (e) the ground truth. Note
that we use the same shape and focus on the difference of textures.
6. More Applications
6.1. Sketch-based Modeling
Custom... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
experiments on commonsense reasoning underscored how the linguistic nature of chain-of-thought
reasoning makes it generally applicable (Section 4). Finally, we showed that for symbolic reasoning,
chain-of-thought prompting facilitates OOD generalization to longer sequence lengths (Section 5). In
all experiments, chain-... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
5https://jupytext.readthedocs.io/
6https://guesslang.readthedocs.io/
6
After dedup
Weight
Percentage
Language
ada
agda
alloy
antlr
applescript
assembly
augeas
awk
batchfile
bluespec
c
c-sharp
clojure
cmake
coffeescript
common-lisp
cpp
css
cuda
dart
dockerfile
elixir
elm
emacs-lisp
erlang
f-sharp
fortran
glsl
go
... | StarCoder_paper (1) |
computer vision: Mind the gap? arXiv preprint arXiv:2112.00639, 2021.
[603] Hendrycks, D., T. G. Dietterich. Benchmarking neural network robustness to common
corruptions and perturbations. In 7th International Conference on Learning Representations,
ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019.... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Gagliardone, I., Patel, A., & Pohjonen, M. (2014). Mapping and analysing hate speech
for Ethiopia. University of Oxford
online: Opportunities and challenges
Comparative Media, Law & Policy website. https://pcmlp.socleg.ox.ac.uk
/mapping-and-analysing-hate-speech-online-opportunities-and-challenges-for-
ethiopia/
Gerst... | Social_Media_and_Democracy |
Inform Process Syst 2020; 1384: 16495–16507.
107360.
[32] Geng Y, Chen J, Ye Z et al. Explainable zero-shot learning via attentive graph convolutional network and knowledge graphs,
http://www.semantic-web-journal.net/system/files/swj2318.pdf
[33] Daniels ZA, Frank LD, Menart CJ et al. A framework for explainable de... | Knowledge-graph-based explainable AI- A systematic review |
# Params ROUGE-L
Model
Vanilla LMs
T5-LM
GPT3
Instruction-tuned w/o SUPERNI
T0
GPT3 + T0 Training
GPT3SELF-INST (Ours)
InstructGPT001
Instruction-tuned w/ SUPERNI
T𝑘-INSTRUCT
GPT3 + SUPERNI Training
GPT3SELF-INST + SUPERNI Training (Ours)
1(cid:13)
2(cid:13)
3(cid:13)
11B
175B
11B
175B
175B
175B
11B
175B
175B
... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
https://a16z.com/the-future-of-music-how-generative-ai-is-transforming-the-music-industry/
8/12
14/11/2023, 13:39
The Future of Music: How Generative AI Is Transforming the Music Industry | Andreessen Horowitz
most popular DAWs today are 20+ years old; startups like TuneFlow and WavTool are
tackling the ambitious ... | The Future of Music_ How Generative AI Is Transforming the Music Industry _ Andreessen Horowitz |
We believe that the open release of LLMs, when done safely, will be a net benefit to society. Like all LLMs,
Llama 2 is a new technology that carries potential risks with use (Bender et al., 2021b; Weidinger et al., 2021;
Solaiman et al., 2023). Testing conducted to date has been in English and has not — and could not ... | Llama2 |
in our dataset, as they are non-generic, and (3) pairs of samples that are all high-quality will have similar
scores (compared to randomly chosen pairs), and so be more difficult to distinguish.
These observations also have an implication for RLHF training, namely that we should expect diminishing
returns from further R... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
• Image and Video: We use 9 benchmarks for image understanding: MMMU (Yue et al., 2023),
TextVQA (Singh et al., 2019), DocVQA (Mathew et al., 2021), ChartQA (Masry et al., 2022),
InfographicVQA (Mathew et al., 2022), MathVista (Lu et al., 2023), AI2D (Kembhavi et al.,
2016), VQAv2 (Goyal et al., 2017), XM3600 (Thapliya... | gemini_1_report |
15
C. Bäckström and P. Jonsson
Artificial Intelligence 302 (2022) 103608
T (V · D) = {{(u = 0), (v = 0)}, {(u = 0), (v = 1)}, {(u = 0), (v = 2)},
{(u = 1), (v = 0)}, {(u = 1), (v = 1)}, {(u = 1), (v = 2)}}
C(V · D) = {{(u = 0), (v = 0)}, {(u = 0), (v = 1)}, {(u = 0), (v = 2)},
{(u = 1), (v = 0)}, {(u = 1), (v = 1)},... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
y
bilit
a
t
e
r
p
r
e
t
In
int
int
int
int
pos
pos
int
int
int
int
int
Fig. 4. Knowledge-based explanations in image recognition tasks, where semantic restrictions are used to elicit information about images.
mappings between ImageNet and Wikidata. A pre-trained CNN is used to classify images captured using OpenCV... | Knowledge graphs as tools for explainable machine learning: A survey |
where the tag <Label> represents the keyword (e.g., violence, explicit content, offensive language, etc.) and <Text> denotes
the responses from per collected Oogiri sample. To further enhance the effectiveness of safety-checking, we additionally
employ the <Label> utilized by NudeNet 3, which includes a substantial num... | Let’sThinkOutsidetheBox |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.