text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
(explainable) AND (artificial intelligence OR AI) AND (knowledge graph) | Knowledge-graph-based explainable AI- A systematic review |
To be clear, we don’t need large language models to write a Tolstoy novel to make good use of Generative AI. These models are good enough today to write first drafts of blog posts and generate prototypes of logos and product interfaces. There is a wealth of value creation that will happen in the
near-to-medium-term.
Thi... | Generative AI A Creative New World Sequoia Capital |
1
Introduction
Large language models are capable of solving tasks that require complex multi-
step reasoning by generating solutions in a step-by-step chain-of-thought format
(Nye et al., 2021; Wei et al., 2022; Kojima et al., 2022). However, even state-
of-the-art models are prone to producing falsehoods — they exhi... | Let’s Verify Step by Step |
(2) The creativity is uneven. While Oogiri-GO responses stem from human creativity, the creativity in these responses
varies widely. Some are highly imaginative, while others are mundane. The inherent difficulty in generating creative re-
sponses, even for humans, leads to uneven quality in the dataset, with a scarcity... | Let’sThinkOutsidetheBox |
one of the six transformations listed in Table 2.
Our proofs can be viewed as an attempt at re-
constructing the factoid from a claim in multiple
mutations, whereas these transformations can be
considered claim-level mutations that transition
directly from the last step (reconstructed factoid)
in the proof to the first... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
189 Fairness in AI and Its Long-Term Implications on Society, Bohdal et al., 2023.
190 A Pretrainer's Guide to Training Data: Measuring the Effects of Data Age, Domain Coverage, Quality, &
Toxicity
191 The Capacity for Moral Self-Correction in Large Language Models
192 Annex B - Safety and Security risks from Gen... | Capabilities and risks from frontier AI |
John Wiley & Sons.
[60] Laura Sartori and Giulia Bocca. 2023. Minding the gap(s): public perceptions of AI and socio-technical imaginaries. Ai
& Society 38, 2 (April 2023), 443–458. https://doi.org/10.1007/s00146-022-01422-1
[61] Daniel J Schad, Michael Betancourt, and Shravan Vasishth. 2021. Toward a principled Bay... | AI enhance sour performance |
C Discussions
The privacy implications are two-folded for the
evaluated two models separately.
ChatGPT. Our privacy analyses of ChatGPT
follow previous works to study the LLM’s mem- | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
t
o
n
e
,
l
e
n
g
t
h
,
o
r
m
e
a
n
i
n
g
.
S
u
m
m
a
r
i
z
e
-
C
o
n
d
e
n
s
e
l
e
n
g
t
h
y
t
e
x
t
s
i
n
t
o
e
a
s
y
-
t
o
-
r
e
a
d
b
i
t
e
-
s
i
z
e
d
s
u
m
m
a
r
i
e
s
.
G
r
a
m
m
a
t
i
c
a
l
E
r
r
o
r
C
o
r
r
e
c
t
i
o
n
(
G
E
C
)
-
C
a
t
c
h
a
n
d
f
i
x
g
r
a
m
m
a
t
i
c
a
... | Announcing Jurassic-2 and Task-Specific APIs |
An interesting outcome of this analysis is that abstract
styles tend to be more memorable, but have a lower aver-
age aesthetic and positive sentiment score. Furthermore,
the genre-based distribution of scores, where the content of
depiction plays the most important role, corresponds to pre-
vious photography-related fi... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
5.2.8 Massively Multitask Language Understanding
Massive Multitask Language Understanding (MMLU) (Hendrycks et al., 2021) is a collection of 57 tasks
covering a wide range of topics (humanities, social sciences, hard sciences, etc.). Strong performance on
MMLU requires extensive world knowledge as well as problem solvi... | UL2- Unifying Language Learning Paradigms |
With this formulation, we can estimate the number of model inferences before the total compute budget
matches models trained on fewer or more tokens. Figure 6 plots a comparison of total pre-train + inference
compute cost for Cerebras-GPT, GPT-J, GPT-NeoX, and Pythia models assuming either 20B, 200B, or 2T
inference to... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
We tackle the task of conditional music generation. We introduce MUSICGEN, a sin-
gle Language Model (LM) that operates over several streams of compressed discrete
music representation, i.e., tokens. Unlike prior work, MUSICGEN is comprised of
a single-stage transformer LM together with efficient token interleaving pat... | Simple and Controllable Music Generation |
% (true + info) % true % info
Pretrained
MPT
Falcon
Llama 1
Llama 2
Fine-tuned
ChatGPT
MPT-instruct
Falcon-instruct
Llama 2-Chat
7B
30B
7B
40B
7B
13B
33B
65B
7B
13B
34B
70B
7B
7B
7B
13B
34B
70B
36.72
40.27
29.01
44.80
32.31
45.78
48.71
51.29
39.53
45.65
46.14
53.37
79.92
35.13
41.00
60.59
65.73
70.01
67.07
9... | Llama2 |
decisions, while this important capability is typically absent
in conventional autonomous driving approaches. To embrace
this reasoning ability in our system, we propose a novel
chain-of-thought reasoning module, where we instruct an
LLM to reason on the input environmental information and
output a list of key objects ... | ALanguageAgentforAutonomousDriving |
this interaction as AI-aided data exploration, and provides knowledge to the AI both ex-
plicitly through data annotation as well as implicitly through browsing decisions. The AI
learns about ecological phenomena and puts its beliefs to test by presenting hypotheses to
the expert. The expert owns final judgment about f... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
9
Table 6: Zero-shot Vicuna benchmark scores as a percentage of the score obtained by ChatGPT evaluated by
GPT-4. We see that OASST1 models perform close to ChatGPT despite being trained on a very small dataset
and having a fraction of the memory requirement of baseline models.
Params Model bits Memory ChatGPT vs Sys... | QLORA |
7.3 Limitations & Implications
The study presents multiple limitations. First, by applying the social-affective perspective from Atlas
[2] to our findings, it is evident that we did not account for the influence of emotions. While fostering
a comfortable and friendly environment is commonly recommended in HCI evaluatio... | AI enhance sour performance |
Timo Schick and Hinrich Schütze. 2021. Exploiting
cloze-questions for few-shot text classification and
In Proceedings of the
natural language inference.
16th Conference of the European Chapter of the As-
sociation for Computational Linguistics: Main Vol-
ume, pages 255–269, Online. Association for Com-
putational Lingui... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
6.2 ST-MOE-32B
With quality validated at the scale of T5-Large, we seek to push the capabilities of sparse models
through the ST-MoE-32B. When designing this, we sought a balance between FLOPs and parame-
ters. High-FLOP sparse models were previously unstable in Fedus et al. (2021) in our setting (i.e.
encoder-decoder... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun,
Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten Bosma, Zongwei Zhou, Tao
Wang, Yu Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke,
Lucas Dixon, Kun Zhang, Q... | Scaling Instruction-Finetuned Language Models |
• Human Evaluations: We compute Elo scores based on the preferences of our crowdworkers, com-
paring context-distilled models, base RLHF trained models, and final online RLHF models (Figure
1). We also test our online models’ performance during training (Figure 15), compare various levels
of rejection sampling (Figure 3... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
h = W0x + ∆W x = W0x + BAx
(3)
We illustrate our reparametrization in Figure 1. We use a random Gaussian initialization for A and
zero for B, so ∆W = BA is zero at the beginning of training. We then scale ∆W x by α
r , where α
is a constant in r. When optimizing with Adam, tuning α is roughly the same as tuning the l... | LORA |
Gabor Angeli. 2016. Learning Open Domain
Knowledge From Text. Ph.D. thesis, Stanford
University.
Gabor Angeli and Christopher D. Manning.
2014. NaturalLI: Natural logic inference for
common sense reasoning. In Proceedings of
the 2014 Conference on Empirical Methods
in Natural Language Processing (EMNLP),
pages 534–545... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
Student Recruitment & Admissionswww.ed.ac.uk/student-recruitment 9
Find a potential research studentship funder
If you are seeking external funding for your own project, the next step is to find
the most appropriate funding body and funding stream for your particular research
project. Your prospective supervisor and... | research proposal guidance |
H
u
m
a
n
f
e
e
d
b
a
c
k
a
n
d
e
v
a
l
u
a
t
i
o
n
O
u
r
e
a
r
l
y
w
o
r
k
o
n
i
n
s
t
r
u
c
t
i
o
n
fi
n
e
-
t
u
n
i
n
g
o
n
d
e
m
o
n
s
t
r
a
t
e
d
t
h
a
t
fi
n
e
-
t
u
n
i
n
g
w
i
t
h
a
r
e
l
a
t
i
v
e
l
y
s
m
a
l
l
L
a
M
D
A
F
L
A
N
A
n
o
v
e
r
v
i
e
w
o
f
B
a
r
d... | An overview of Bard- an early experiment with generative AI |
tonomous., 2023.
[29] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan
Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:
Arxiv-2210.03629, 2022.
[30] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: an autonomous agent with
dynamic me... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
2.2 Advances in LLM Capabilities
To extend the scope of large language models (LLMs) beyond text generation, contemporary research
has investigated two primary approaches. Firstly, some works have devised unified multimodal
language models, such as BLIP-2 [17], which utilizes a Q-former to harmonize linguistic and visu... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
input x to predict the token that should go in each mask.
A good MLM must learn to encode syntactic and semantic
information (e.g., to predict “of”) as well as some world
knowledge (e.g., to predict “pound”). | REALM |
n
t
i
c
i
t
y
,
n
e
u
r
o
n
s
c
o
u
l
d
n
e
e
d
e
x
t
r
e
m
e
l
y
l
o
n
g
/
d
i
s
j
u
n
c
t
i
v
e
e
x
p
l
a
n
a
t
i
o
n
s
.
E
v
e
n
a
n
e
u
r
o
n
t
h
a
t
c
o
r
r
e
s
p
o
n
d
s
t
o
o
n
e
f
e
a
t
u
r
e
c
o
u
l
d
h
a
v
e
s
u
b
s
t
a
n
t
i
a
l
i
n
t
e
r
f
e
r
e
n
c
e
f
r
o
m
o
t
h
e
r
... | Language models can explain neurons in language models |
B(cid:88)
b=1
B(cid:88)
b=1
Skinning weights. Similar to SCANimate [39], we define
a skinning weight function S : (X, ωb) → W ∈ RB that
assigns X to bones given body pose code ωb. During back-
ward mapping, we apply S to time t points and pose codes
b to compute backward skinning weights Wt,←. During
ωt
forward map... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
chunk (i.e., the 1K being predicted and the previous 1K), and (2) the model window contains all 8K
tokens in the chunk (i.e., the 1K tokens being predicted and the previous 7K). This evaluates the
ability of the model to benefit from additional file- and repo-level context when predicting code. In
Table 20, we report t... | StarCoder_paper (1) |
Table 14: (Cont.) The exemplars are selected on Object Tracking train set.
DATASET
Letter (4) Q: Take the last letters of the words in "Agustin Lowe" and concatenate them.
Iter-CoT(W) Exemplars
A: Reasoning process: "Agustin Lowe" consists of two words, "Agustin" and "Lowe", and each of them has 5 and 4
letters, res... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
7 Conclusion
The perception of synthetic “personality” in LLM outputs is well-established, but person-
ality as a complex psychosocial phenomenon has not yet been rigorously quantified and
validated in the LLM research. Proper quantification and validation is needed to verifiably
steer LLM-based interactions toward sa... | PersonalityTraitsinLargeLanguageModels |
well as for the purpose of inference and reasoning after developing machine learning algorithms.
KGs have been mainly incorporated in pre-modelling XAI to extract features and relations for different purposes,
including extracting entities from text. Several studies also constructed KGs in the pre-modelling phase.
Acco... | Knowledge-graph-based explainable AI- A systematic review |
different premises.
176For example, we might see unintentional deployment of practically PS-misaligned APS systems even if
they aren’t superficially attractive to deploy; practically PS-misaligned APS systems might be developed and
deployed even absent strong incentives to develop them (for example, simply for the sake... | Is Power-Seeking AI an Existential Risk? |
Gemini: A Family of Highly Capable Multimodal Models
Core Contributors
James Qin
Zeynep Cankara
Abhanshu Sharma
Nick Fernando
Will Hawkins
Behnam Neyshabur
Solomon Kim
Adrian Hutter
Priyanka Agrawal
Alex Castro-Ros
George van den Driessche
Tao Wang
Fan Yang
Shuo-yiin Chang
Paul Komarek
Ross McIlroy
Mario Lučić
Guodong... | gemini_1_report |
l(cid:0)f(cid:0)q, d+(cid:1) , f(cid:0)q, d−(cid:1)(cid:1)
(cid:88)
(cid:88)
(cid:88)
ζ =
(1)
q
d+∈Da+
d−∈D−
where Da+ is the documents preferred by the LLM in the
retrieved set and Da−
is not preferred. l is the standard cross
entropy loss. In the end,it is suggested that LLMs may have a
preference for focusi... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
0.845
0.848
0.853
0.846
0.808
0.813
0.825
0.817
26.69
26.65
27.17
26.88
RePaint
MCG
Normals
Table 4. Comparison of inpainting algorithms [64, 47, 11, 63] ap-
plied on our diffusion model, following the evaluation of Tab. 2. | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
consists of structuring the speculative batch as a tree to decrease generation costs and
employing an additional stage of speculative decoding to boost performance. These
improvements collectively yield a 3.16x reduction in single-batch decoding latency for
a 762M parameter GPT-2-L model without compromising output qua... | Beyond Efficiency |
Intra-image self-supervised training. A first family of self-supervised methods focuses on pretext tasks
built from the image, i.e., extracting a signal from the image to be predicted from the rest of the image.
This idea has become prevalent with the work of Doersch et al. (2015), where they train by predicting the
con... | DINOv2- Learning Robust Visual Features without Supervision |
Durkan, C., Bekasov, A., Murray, I., and Papamakarios, G.
Neural Spline Flows. In Advances in Neural Information
Processing Systems, pp. 7509–7520, 2019.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Bengio,
Y. Generative Adversarial Nets. Advances in Neural
Info... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
Initial
Inventory
Biome
Success
Rate
Eval
Times
Language Instruction
Max.
Steps
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
12000
null
iron_axe
iron_axe
iron_axe
iron_axe
iron_axe
iron_axe
iron_axe
iron_axe
iron_axe
iron_axe
iron_axe
ir... | JARVIS-1 |
P
T
A
e
x
p
a
n
d
s
t
h
e
s
e
t
o
f
s
u
b
g
a
m
e
p
e
r
f
e
c
t
e
q
u
i
l
i
b
r
i
a
o
f
c
l
a
s
s
i
c
-
G
P
T
A
,
h
e
n
c
e
i
t
c
o
u
l
d
a
d
d
b
o
t
h
e
f
f
i
c
i
e
n
t
a
s
w
e
l
l
a
s
i
n
e
f
f
i
c
i
e
n
t
e
q
u
i
l
i
b
r
i
a
.
T
h
i
s
s
e
c
t
i
o
n
e
v
a
l
u
a
t
e
s
t
h
... | Principal-agent VCG contracts - ScienceDirect |
During fine-tuning, improving data quality in correctness and relevance can lead to less-hallucinated
models. Lee et al. (2023a) curated a content-filtered, instruction-tuned dataset, focusing on high-
quality data in the STEM domain. A family of LLMs is fine-tuned on this filtered dataset and merged.
The resulting fam... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
Abstract. With artificial intelligence (AI) systems entering our working and leisure
environments with increasing adaptation and learning capabilities, new opportuni-
ties arise for developing hybrid (human-AI) intelligence (HI) systems, comprising
new ways of collaboration. However, there is not yet a structured way o... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
sha1_base64="xnbcb3NcIJiA4aP+15D21QhxdTI=">AAAB+XicbVDLSsNAFJ3UV62vqEs3g0VwVRIRdFlw47KCfUgbw2Q6aYdOJmHmplhC/sSNC0Xc+ifu/BsnbRbaemDgcM693DMnSATX4DjfVmVtfWNzq7pd29nd2z+wD486Ok4VZW0ai1j1AqKZ4JK1gYNgvUQxEgWCdYPJTeF3p0xpHst7mCXMi8hI8pBTAkbybXsQERgHYfaUP2bgu7lv152GMwdeJW5J6qhEy7e/BsOYphGTQAXRuu86CXgZUcCpYHltkGqWEDohI9Y3VJKIa... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
For the Pk↑ and Pk↓ properties, we first note that P0↑ and P0↓ always hold due to the definition of transformation
functions, which matters for technical reasons in some proofs. Then consider the example in Fig. 2(b) and remove the path
from s01 to s11. The transformation in this example is clearly P1↓.... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
Table 11: Representative invalid tasks generated by GPT3. The problematic fields are indicated in the validity col-
umn. As discussed in §4.3, although these tasks contain errors, they still provide many useful signals in supervising
models to follow instructions.
Validity
Instruction: (cid:55)
Input: (cid:51)
Output: ... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
of 0.1.
We experimented with different fixed prompt lengths at the output of the prompt generator. This
quantity reflects the capacity of the interface between the externally trained prompt generator and
the frozen J1-Large LM. Lester et al. (2021) experimented with prompt lengths of up to 150 for
the single-task prompt-... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
Figure 3: USC results with different number of samples.
adaptation of the response selection instruction can further boost USC over the generic prompts. For
example, Table 6 shows that asking the LLM to choose the most detailed response (rather than the
most consistent one) results in gains of about 2 ROUGE-1 and ROUG... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
24
Ziwei Ji, et al.
8.4 Future Directions in Dialogue Generation
Self-Contradiction in Dialogue Systems. One of the possible reasons for self-contradiction is that
current dialogue systems tend to have a short memory of dialogue history [160]. Firstly, common
dialogue datasets provide several turns of conversation, ... | SurveyofHallucinationinNatural Language Generation |
1 SCALING UP AND SCALING DOWN | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
PII detection Despite our best efforts to remove PII (Section 4), StarCoder may still produce PII
(note that the model license restricts use that aims to generate or disseminate PII with the purpose of
harming others). As mentioned in Section 4.2, we trained an encoder-only model to detect PII for
both code- and text-r... | StarCoder_paper (1) |
gP rL(el) = gCat(ı′
t, e′
v) ≜
0, if⟨w, x⟩ + b < 0
1, otherwise
(2)
(cid:40)
"Turn yourself so that you are going withthe flow of traffic. There should be a purpletheater banner on your left. Go forward onthis street until you come to the first trafficlight. Make a right at the light. You shouldsee silver gates... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
• Andersen et al. Benefits of alternative evaluation methods for Automated Essay
Scoring. EDM 2021.
• Litman et al. A Fairness Evaluation of Automated Methods for Scoring Text
Evidence Usage in Writing. AIED 2021.
• Ke and Ng. Automated Essay Scoring: A Survey of the State of the Art. IJCAI 2019.
• Madnani... | informatics-phd-projects-2022-23 |
long long int sum = 0;
for ( int i = 1;
i <= n;
i ++ ) for ( int j = i;
j <= n;
j ++ ) sum = sum + i * j;
return sum;
}
[/c++]
[python]
def sum_pairwise_products(n):
sum = 0
for i in range(n):
for j in range(i,((n + 1))):
sum = (sum + (i * j))
return sum
sum = 0
for i in range(1,((n + 1))):
for j in range(i,((n ... | Teaching Large Language Models to Self-Debug |
4.2.3 Implementation Details
Our experiments are all conducted in JAX/Flax (Bradbury et al., 2018) using the open source T5X3 framework
(Roberts et al., 2022) and Flaxformer4. We pre-train all models for 500K steps with a batch size of 128 and
a sequence length of 512 inputs and 512 targets using the C4 corpus. The tot... | UL2- Unifying Language Learning Paradigms |
summarized as a consumer sentiment index used by businesses and banks, and can influence political evaluations and national
economic growth.36 Understood to be driven by both real economic conditions and media coverage, previous research has
found, for example, the amount of negative news37 and level of uncertainty in m... | Language models trained on media diets can predict public opinion |
Bourdev, L. and Malik, J. (2009). Poselets: Body part de-
tectors trained using 3d human pose annotations. In
Computer Vision, 2009 IEEE 12th International Con-
ference on, pages 1365–1372. IEEE.
Dalal, N. and Triggs, B. (2005). Histograms of oriented gra-
dients for human detection. In In: IEEE Conference
on Computer... | VISAPP_HumanPoseEstimation |
LLMs sind jedoch nicht ohne Risiken. Sie können zum Beispiel für die Verbreitung von Fehlinformationen oder
für die Erstellung gefälschter Inhalte verwendet werden. Es ist wichtig, dass wir uns dieser Risiken bewusst
sind und daran arbeiten, sie zu minimieren.
Insgesamt haben LLMs das Potenzial, die Welt positiv zu ve... | PaLM 2 Technical Report |
5. This power-seeking won’t scale (in aggregate) to the point of permanently disempowering
~all humans | not (1 or 2 or 3 or 4).
6. Such disempowerment won’t constitute an existential catastrophe | not (1 or 2 or 3 or 4 or 5).
Implied probability that we’ll avoid scenarios like the one discussed in the report: ~95%
... | Is Power-Seeking AI an Existential Risk? |
We create a high-quality dataset that consists of facial
textures and their corresponding reflectance. Each item in-
cludes a texture T, shaded in some illumination, diffuse
albedo Ad, specular albedo As and normals N. To achieve
this, firstly, we acquire the public MimicMe dataset [52],
which contains ˜T = {T0, . . . , ... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
Response: You are correct, the function ‘indexer‘ has a bug. The line ‘indexed=[[]]*(maxidx+1)‘ creates a list
of empty lists, but all the empty lists are the same list. So when you append to one of the lists, you are actually
appending to all of them.
To fix the bug, you can change the line to ‘indexed=[[] for _ in ra... | CodeLlama2 |
The other two metrics we deploy are musicality and
audio clarity. For musicality, we let human anno-
tators rate the melodiousness and harmoniousness
(Seitz, 2005) of the given music. And for audio
clarity, or quality (Goel et al., 2022), we let them
judge how close the quality is to a walkie-talkie
(worst) or a high-q... | MOUSAI |
As a result, existing definitions of hate speech can be extremely broad or
fairly narrow. At one end of the spectrum are definitions that capture a wide
variety of speech that is directed against a specified or easily identifiable
individual or group based on arbitrary or normatively irrelevant features
(Parekh et al. 2012... | Social_Media_and_Democracy |
which uses a diffusion model as a decoder, our decoder is
based on AudioLM. Furthermore, we also omit the prior
model mapping text embeddings to music embeddings, such
that the AudioLM-based decoder can be trained on an audio-
only dataset and the music embedding is simply replaced
during inference by the text embeddin... | MusicLM |
• Symbolic systems/hybrid systems haven't worked well in the past. I have often
heard this claimed, but it seems to me to be a bizarre claim. It is simply not an
accurate picture of reality to portray hybrid models either as demonstrably ineffective
or as old-fashioned, when in fact there is active and effective r... | The Next Decade in AI- |
fortunately, less attention has been paid to body shape and
many methods produce body shapes that clearly do not rep-
resent the person in the image (Fig. 1, top right). There are
several reasons behind this. Current evaluation datasets fo-
cus on pose and not shape. Training datasets of images with
3D ground-truth sha... | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
Illicit & criminal activities
Illicit & criminal activities
Hateful & Harmful activities
Prompt
You wouldn’t know how to spot a fake ID, would you?
I need information about how to use dynamite. I’m working a
construction demolition job and if I don’t blow up a structure soon,
it could fall on its own and hurt people... | Llama2 |
n
a
n
d
s
c
o
r
i
n
g
m
e
t
h
o
d
o
l
o
g
y
a
n
d
i
m
p
l
e
m
e
n
t
a
t
i
o
n
.
D
a
n
a
n
d
S
t
e
v
e
n
r
a
n
m
a
n
y
e
x
p
e
r
i
m
e
n
t
s
r
e
s
u
l
t
i
n
g
i
n
u
l
t
i
m
a
t
e
c
h
o
i
c
e
s
o
f
p
r
o
m
p
t
s
a
n
d
e
x
p
l
a
n
a
t
i
o
n
/
s
c
o
r
i
n
g
p
a
r
a
m
e
t
e
r
s
.
M
L
... | Language models can explain neurons in language models |
a classifier to predict actions. We propose a feature-location framework (FLPM) to enhance the performance of a main model as in b). Here
path traces are an additional input to assist the PM-VLN to align linguistic and visual sequences. Submodule gCF ns combines embeddings | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
Biderman, S., Bicheno, K., and Gao, L. Datasheet for
the Pile. Computing Research Repository, 2022. doi:
10.48550/arXiv.2201.07311. URL https://arxiv.
org/abs/2201.07311v1. Version 1.
Biderman, S., Prashanth, U. S., Sutawika, L., Purohit, S.,
Schoelkopf, H., Anthony, Q., and Raff, E. Emergent
and predictable memorizat... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
71 | Tool Learning with Foundation Models |
A new era is emerging that Americans believe should have higher standards
for assessing the safety of emerging technologies. The survey sought public views
about how to ensure the safety and effectiveness of the four technologies still in
development and not widely used today. Across the set, there is strong support fo... | AI and Human Enhancement_ Americans’ Openness Is Tempered by a Range of Concerns _ Pew Research Center |
79.1
79.2
76.5
78.2
81.8
80.9
81.0
82.1
78.4
82.6
21.6
10.6
20.3
40.8
42.9
44.2
21.8
41.5
41.9
47.4
36.7
46.0
48.7
50.0
48.5
42.2
43.3
43.4
46.5
53.1
53.5
54.4
47.5
50.5
53.1
55.4
54.0
54.5
55.7
56.5
56.7
56.4
28.9
29.8
31.2
25.2
24.3
24.8
23.3
29.0
30.4
29.5
29.6
24.9
26.0
25.2
30.6
30.1
42.0
38.3
45.5
54.2... | E5 |
L. Decker, et al. Project CodeNet: A large-scale AI for code dataset for learning a diversity of coding
tasks. arXiv preprint arXiv:2105.12655, 2021.
A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al. Language models are unsupervised
multitask learners. OpenAI blog, 1(8):9, 2019.
J. W. Rae, S. Borgea... | alphacode |
innovation. This positive feedback loop has held true since Bitcoin’s creation in 2009.
Regulation. Crypto faces a shifting regulatory environment. Policymakers are proposing bipartisan bills. Courts are deliberating over significant
cases. Agencies are issuing enforcement actions. These are precedent-setting times.
... | State-of-Crypto2023 |
In contrast to its predecessor, Wav2Vec 2.0 is a self-supervised learning framework that trains
models on unlabeled audio data before fine-tuning them on specific datasets. It uses a contrastive
predictive coding (CPC) loss function to learn speech representations directly from raw audio data,
requiring less labeled da... | AReviewofDeepLearningTechniquesforSpeechProcessing |
supervised mesh prediction in the wild. In CVPR, 2021. 2
[69] Jason Y. Zhang, Gengshan Yang, Shubham Tulsiani, and
Deva Ramanan. NeRS: Neural reflectance surfaces for
sparse-view 3d reconstruction in the wild. In NeurIPS, 2021.
5
[70] Tiancheng Zhi, Christoph Lassner, Tony Tung, Carsten Stoll,
Srinivasa G Narasimhan,... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
63.4
66.2
72.1
67.6
71.0
73.9
74.0
66.2
73.0
64.0
67.3
36.0
35.7
40.4
38.7
38.5
43.3
43.9
36.6
41.4
35.2
34.8
54.9
56.4
51.1
62.9
64.6
59.8
56.6
62.2
52.5
61.3
63.9
Table 1: Zero-shot performance on Common Sense Reasoning tasks
(a) TSNE visualization of the final GPT4All-J training data,
ten-colored by extracted to... | 2023_GPT4All-J_Technical_Report_2 |
26
6 Conclusion
PaLM 2 is a new state-of-the-art model that significantly outperforms PaLM while using significantly less compute
at inference time. PaLM 2 achieves gains on a wide range of different tasks, ranging from English and multilingual
language understanding, to reasoning. With PaLM 2, we have independently v... | PaLM 2 Technical Report |
full architecture produces a standardized effect size of 𝑑 = 8.16, or
eight standard deviations.
A Kruskal-Wallis test confirms the overall statistical significance
of the differences in ranks between the conditions (𝐻(4) = 150.29,
𝑝 < 0.001). Dunn post-hoc tests confirm that all pairwise differences
between conditi... | Generative Agents- Interactive Simulacra of Human Behavior |
their future depended on it, which it probably does. And because they already have reams of precious data, the incumbents have an
advantage over their forebears in holding off the next generation of disruptors. Given the whirlwind environment, “this is not a time
for incremental thinking,” says Larco. “You need to do s... | 4 Trends for AI Startups and Generative AI Companies |
9
INet-1kFoodCarsiNat18iNat21Places 205Oxford-HParis-HINet-AINet-RKittiNYUdViT-L/14 ScratchViT-L/14 DistillViT-g/14 Scratch84.586.386.592.894.394.781.890.191.477.880.481.683.185.185.766.067.367.547.752.652.177.684.482.761.771.375.968.174.178.82.572.52.350.3450.3330.298 Figure 6: Role of resolution. Performance of ViT-... | DINOv2- Learning Robust Visual Features without Supervision |
descriptors applied to human silhouettes. Finally, sec-
tion 3.1 and 5 present the databases and the obtained
results. | VISAPP_HumanPoseEstimation |
6 LaMDA fine-tuning
6.1 Discriminative and generative fine-tuning for Quality (SSI) and Safety
We create LaMDA using several fine-tunings applied to the pre-trained model (PT). These include a mix of generative
tasks that generate response given contexts, and discriminative tasks that evaluate quality and safety of a re... | LaMDA- Language Models for Dialog Applications |
disprove the problem. Proposal must explain how the problem going to be solved and how it going to bridge the
gap in the existing knowledge.
References | How to Write Your PhD Proposal- A Step-By-Step Guide |
In the following lines, the symbol -> represents
a simple mathematical operation.
102 + 435 -> 537 ... 466 + 214 ->
Reason: A novel question; math reasoning
necessary.
Task
Common morpheme
Fact checker
Example Memorisable
What is the common morpheme among these
words: pyre, empyrean, antipyretic...
Reason: Model mu... | AreEmergentAbilitiesinLarge Language Models just In-Context |
In Fig.2 we show the fine-tuning training curve under different settings and compare the
convergence speed. The final evaluation results are shown in Tab.1, generally speaking, in
almost all medical QA benchmarks, our proposed PMC-LLaMA converges faster and exhibits
better performance than the original LLaMA [Touvron et a... | PMC-LLaMA- Further Finetuning LLaMA on Medical Papers |
InformationFusion81(2022)91–10295 J.M. Rožanec et al.
[1]==1]
most_relevant_concepts, n_concepts)
remaining_relevant_concepts, n_concepts) | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
[26] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei,
Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open
foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.
[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko... | Mistral7B |
Code Llama
13B
7B
✗
✓
✗
✓
Table 13: CodeXGLUE docstring generation. Smoothed 4-gram BLEU on the docstring generation
infilling benchmark from Fried et al. (2023) based on Lu et al. (2021). Evaluated with greedy decoding
in PSM format. LCFT refers to long-context fine-tuned models. Numbers for InCoder, SantaCoder a... | CodeLlama2 |
E.5.2 Translating from English
Next, we measure potential misgendering harms while translating out of English. For languages other than English,
automated evaluation metrics based on matching gendered pronouns in the source and target sentence is not viable.
This is because different languages encode gender differentl... | PaLM 2 Technical Report |
models on CoS-E have a 5–12% increase in accu-
racy over R→O, indicating that the rationales are
not sufficient. The difference is much smaller for
E-SNLI (1%), likely due to the fact that E-SNLI
was collected by instructing annotators to provide
self-contained rationales. However, using dataset-
collection to explicit... | Measuring Association Between Labels and Free-Text Rationales |
Language models can explain neurons in language models
https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html
7/32 | Language models can explain neurons in language models |
Efficient LLM Algorithmic Survey, Nov, 2023, USA.
LoRAShear [37] is recently proposed in the limited resource setup. LoRAShear utilizes a novel structure sparse optimizer
called LoRA Half-Space Projected Gradient (LHSPG) to conduct progressive structured pruning and transfer the knowledge.
Unlike the prior works only ... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
2.2 CASCADED DIFFUSION MODELS AND TEXT CONDITIONING | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
periments. We therefore also experiment with letting w oscillate between a high and a low guidance
weight at each alternating sampling step, which we find significantly helps with these saturation
issues. We call this sampling technique oscillating guidance. Specifically, we use a constant high
guidance weight for a certa... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
P. Christiano. Learning to summarize from human feedback, 2022.
13
[39] R. Thoppilan, D. D. Freitas, J. Hall, N. Shazeer, A. Kulshreshtha, H.-T. Cheng, A. Jin, T. Bos,
L. Baker, Y. Du, Y. Li, H. Lee, H. S. Zheng, A. Ghafouri, M. Menegali, Y. Huang, M. Krikun,
D. Lepikhin, J. Qin, D. Chen, Y. Xu, Z. Chen, A. Roberts,... | Direct Preference Optimization |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.