text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
ADE20K / train
Cityscapes / train
Pascal VOC 2012 (seg.) / trainaug
Mapillary SLS / train
KITTI / train (Eigen)
NYU Depth V2 / train
SUN RGB-D / train
Google Landmarks v2 / train (clean)
Google Landmarks v2 / train (clean)
AmsterTime / new
AmsterTime / old
Met / train
Revisiting Oxford / base
Revisiting Paris / base
1... | DINOv2- Learning Robust Visual Features without Supervision |
21 | ALanguageAgentforAutonomousDriving |
AI at Work:
What People
Are Saying
JUNE 2023
Executive
summary
We surveyed nearly 13,000 people—from
executive suite leaders to middle managers
and frontline employees—in 18 countries
to understand their thoughts, emotions,
and fears about AI.
Respondents today are optimistic about how AI—and
generative AI,... | AI at Work- What People Are Saying |
2 Flan Finetuning
We instruction-finetune on a collection of data sources (Figure 2) with a variety of instruction template
types (Figure 3). We call this finetuning procedure Flan (Finetuning language models; Wei et al., 2021) and
prepend “Flan” to the resulting finetuned models (e.g., Flan-PaLM).2 We show that Flan work... | Scaling Instruction-Finetuned Language Models |
and Oriol Vinyals. The benchmark lottery. 2021b.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional
transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Mi... | UL2- Unifying Language Learning Paradigms |
4. For datasets with few categorical features, polynomial kernels
tend to be more effective.
We performed leave-one-out evaluation on the PD1 benchmark, which consists of 23 tasks. However,
some tasks are using the same model and dataset but only different in batch size. These tasks should
not appear in training task... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
Joel Hestness, Sharan Narang, Newsha Ardalani, Gre-
gory Diamos, Heewoo Jun, Hassan Kianinejad,
Md Patwary, Mostofa Ali, Yang Yang, and Yanqi
Zhou. 2017. Deep learning scaling is predictable,
empirically. arXiv preprint arXiv:1712.00409.
Sepp Hochreiter and Jürgen Schmidhuber. 1997.
Neural computation,
Long short-ter... | LLaMA- Open and Efficient Foundation Language Models |
Human specific rendering: The work of Kanade et al.
[27] is one of the earliest investigations into free-viewpoint
rendering of humans. It introduced a dome equipped with
cameras to recover depth maps and meshes, enabling novel
views to be rendered by reprojecting and blending differ-
ent views to account for mesh hole... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
Both humans and animals rely on sensory organs like eyes and ears to gather information from their
surroundings. These perceptual inputs are converted into neural signals and sent to the brain for
processing [299; 300], allowing us to perceive and interact with the world. Similarly, it’s crucial
for LLM-based agents to... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
n-gram models of natural language. Computational linguistics, 18(4):467–480.
Collobert, R. and Weston, J. (2008). A unified architecture for natural language processing: Deep
neural networks with multitask learning. In Proceedings of the 25th international conference on
Machine learning, pages 160–167.
Collobert, R., ... | MULTI HASH EMBEDDINGS IN SPACY |
Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam
Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh,
Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam
Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
Recursive retrieval and multi-hop retrieval are used for spe-
cific data scenarios. Recursive retrieval can first process data
through a structured index, then retrieve it level by level.
When retrieving hierarchically rich documents, a summary
can be made for each section in an entire document or long
PDF. A retrieval... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
2.2 AI Model Evaluation
AI model evaluation is an essential step in assessing the performance of a model. There are some
standard model evaluation protocols, including 𝑘-fold cross-validation, holdout validation, leave
one out cross-validation (LOOCV), bootstrap, and reduced set [6, 88]. For instance, 𝑘-fold cross-
v... | ASurveyonEvaluationofLargeLanguageModels |
Language modeling. Using a variation on the experimental setup of Gehman et al. (2020), this evaluation focuses on
measuring control over toxic degeneration. We sample 50k prompts from Gehman et al. (2020), and filter to only those
input prompts with toxicity probability < 0.5 using the toxicity scores within the datase... | PaLM 2 Technical Report |
24We can imagine cases in which AI agents end up valuing power for its own sake, but I’m not going to focus
on those here. | Is Power-Seeking AI an Existential Risk? |
2 | Translatotron3 |
After an LLM provides an output for a specific
prompt, proper feedback about the output can make
the LLM give better and more accurate outputs
in its consecutive iterations (Madaan et al., 2023).
Abiding by this method, the following are the spe-
cific hallucination mitigation techniques:
Prompting GPT-3 To Be Reliable... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
Erenay Dayanik and Sebastian Padó. 2021. Disentan-
gling document topic and author gender in multiple
languages: Lessons for adversarial debiasing. In Pro-
ceedings of the Eleventh Workshop on Computational
Approaches to Subjectivity, Sentiment and Social Me-
dia Analysis, pages 50–61, Online. Association for
Computati... | Are Pretrained Multilingual Models Equally Fair Across Languages? |
Platforms interviewed for the Urban study also reported a low rate of DMCA
counter-notices from users challenging erroneous takedowns. Many platforms
received no counter-notices at all (Urban et al. 2016, p. 44). This finding is
consistent with figures released by the Motion Picture Association of America in
2013, showin... | Social_Media_and_Democracy |
(7) Assume τ is DLBS. Then it is immediate from the definition that τ is M↓.
Suppose (cid:3)s1, t1, a(cid:4) ∈ E1. Then pre(a) ⊆ s1. We have pre(g(a)) = pre(a) and s1 ∈ S2, since S1 ⊆ S2, so (cid:3)s1, t2, g(a)(cid:4) ∈
E2, where t2 = s1 (cid:4) post(g(a)). Since also R(a, g(a)) holds by definition it f... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
e) Encoder-decoder models remain promising, as this type of architecture is still being actively explored, and
most of them are open-sourced. Google has made substantial contributions to open-source encoder-decoder
architectures. However, the flexibility and versatility of decoder-only models seem to make Google’s insi... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
social media and political polarization
rather counterintuitively, | Social_Media_and_Democracy |
that would not be guaranteed to be identical across libraries even if Google were
to make such ads available. In addition, even election-related content is not
comparable across sources as Google only makes election-related advertising
available for federal and statewide races, and Facebook data lack programmatic
acces... | Social_Media_and_Democracy |
9
510k25100k0.780.7850.790.7950.80.8050.810.8150.720.730.740.750.760.77MNLI AccuracyGLUE ScoreVocabulary SizeMNLI AccuracyGLUE Score Preprint
BERT-base (Fully trained)
BERT-base (No Pretrain)
BERT (normal protocol)
BERT ((Izsak et al., 2021))
crammed BERT
BERT (normal protocol)
BERT ((Izsak et al., 2021))
crammed B... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
(2)
where σ is the logistic function. In the context of LMs, the network rϕ(x, y) is often initialized from
the SFT model πSFT(y | x) with the addition of a linear layer on top of the final transformer layer
that produces a single scalar prediction for the reward value [49]. To ensure a reward function with
lower varia... | Direct Preference Optimization |
is can then be extracted from Mk:
= Mk(pc, p).
(17)
B. Network Architecture
Figures 9-12 show the network design for the canoni-
cal MLP, the non-rigid motion MLP, the pose correction
MLP, and the deep network generating the canonical mo-
tion weight volume.
(cid:34)
Rk
0
(cid:35)
tk
1
Figure 9. Canonical MLP ... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
time periodWhy Study the History of the English Language?•Understanding the evolution of language and its impact on culture and society•Appreciating the richness and diversity of English literature•Improving language skills and communication abilities•Gaining a deeper understanding of one's own language and identity58 ... | Tool Learning with Foundation Models |
Input & Output
Sample rate (Hz)
Mel channels
Mel lower band (Hz)
Mel upper band (Hz)
Frame size (ms)
Frame step (ms)
SpecAugment
Freq blocks
Time blocks
Freq block max length ratio
Time block max length ratio
Encoder
Conformer dims
Attention heads
Conv kernal size
Subsample factor
Attention (source & target)
Output & H... | Translatotron3 |
6 https://www.w3 .org /TR /rdf -sparql -query/.
7 https://neo4j .com /developer /cypher-query-language/.
8 https://tinkerpop .apache .org /gremlin .html.
9 http://www.cyc .com /opencyc /a, discontinued in 2017.
10 http://www.freebase .com, discontinued in 2015.
11 https://www.wikidata .org /wiki /Wikidata :Main _Page.
... | Knowledge graphs as tools for explainable machine learning: A survey |
5http://codeforces.com/
11
Gemini: A Family of Highly Capable Multimodal Models
5.2. Multimodal
Gemini models are natively multimodal. These models exhibit the unique ability to seamlessly
combine their capabilities across modalities (e.g. extracting information and spatial layout out of
a table, a chart, or a figu... | gemini_1_report |
critical to understanding the documents. It requires solutions distinct from conventional large language models such as
GPT-3.5 [3], Llama [4], Falcon [5] or PaLM [6] that primarily accept text-only inputs and assume that the documents
exhibit simple layouts and uniform formatting, which may not be suitable for handlin... | DOCLLM |
pensable aspect of instruction-tuning as LMs
need to learn about issues that were not quite
learned during pre-training. | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
tempting to compare to human performance, to avoid over-
stating the capabilities of machine learning systems due to
misleading comparisons. | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
t ), and VC(zC
5
Table 1: Training tasks (CT stands for “contrastive learning” to align prompt encoders) and datasets
with corresponding statistics. * denotes the number of accessible examples in the original datasets.
Datasets
# of samples Domain
Categories
Image + Text
Tasks
Image→Text, Text→Image
Text→Image+... | Any-to-Any Generation via Composable Diffusion |
Copyright 2023
NEA Terms & Conditions
Privacy Policy
NEA – EU SFDR Notice
Required Japanese Notice
https://www.nea.com/blog/4-trends-for-ai-startups-and-generative-ai-companies
20/20 | 4 Trends for AI Startups and Generative AI Companies |
[25] Max Morrison, Rithesh Kumar, Kundan Kumar, Prem Seetharaman, Aaron Courville, and
Yoshua Bengio. Chunked autoregressive gan for conditional waveform synthesis. arXiv preprint
arXiv:2110.10139, 2021.
[26] Gautham J Mysore. Can we automatically transform speech recorded on common consumer
devices in real-world envi... | RVQGAN |
Abdelali et al. [1] evaluated the performance of ChatGPT in standard Arabic NLP tasks and
observed that ChatGPT exhibits lower performance compared to SOTA models in the zero-shot
setting for most tasks. Ahuja et al. [2], Bang et al. [5], Lai et al. [93], Zhang et al. [236] utilized a
greater number of languages across... | ASurveyonEvaluationofLargeLanguageModels |
generalise? Journal of Artificial Intelligence Research (JAIR), 2020.
[78] Z. Allen-Zhu and Y. Li. Physics of Language Models: Part 1, Context-Free Grammar. arXiv:2305.13673, 2023.
[79] E. Jang. Just Ask for Generalization. In https://evjang.com/2021/10/23/generalization.html, 2022.
[80] L. Chen, K. Lu, A. Rajeswara... | LargeLanguageModelsasGeneralPatternMachines |
Figure 5: The scaling law obtained from all 4 compute scales.
8
Table 1: Estimated optimal parameter size at a given number of FLOPs in our study compared to the study of Hoffmann
et al. (2022).
FLOPs
Loss
Tokens
Non-
Embedding
Parameters
3.31× 109
6.08× 109
8.95× 109
1.47× 1010
7.71× 108
2.36× 109
3.32× 109
8.... | PaLM 2 Technical Report |
Luo.Learningtogeneratetime-lapsevideosusingmulti-stagedy-namicgenerativeadversarialnetworks.InProceedingsoftheIEEEConferenceonComputerVisionandPatternRecogni-tion,pages2364–2373,2018.2[82]YuanXue,Yuan-ChenGuo,HanZhang,TaoXu,Song-HaiZhang,andXiaoleiHuang.Deepimagesynthesisfromintu-itiveuserinput:Areviewandperspectives.C... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
38
Table 25: Few-shot exemplars for full chain of thought prompt for StrategyQA. | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
0 and away from the xl
0.
Inference Time-Optimization namely DOODL [51], does not learn any new model parameters, instead optimizing diffu-
sion latents to improve some criterion on the generated image similar to CLIP+VQGAN[8]. This runtime compute increases
inference cost by more than an order of magnitude. | DiffusionModelAlignmentUsing Direct Preference Optimization |
a-tionwithoutpairedtext-videodata.Differentfromalltheabovemodels,LFDMinsteadappliesDMtogeneratelatentflowsequencesforconditionalimage-to-videogeneration.3.OurMethodLetn∼N(0,I)beaGaussiannoisevolumewiththeshapeofKn×Hn×Wn×Cn,whereKnHn,Wn,andCnarelength,height,width,andchannelnumber,respec-tively.Givenonestartingimagex0and... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
For example: Hello, I’m Emily. –> Hola, soy Emily.
Here we know that the gender is correct because the proper name is found in the target. In both the target, the verb is
not gender inflected in this case, so there is no need for gender agreement in that sense.
Or when translating to languages like Bengali or Thai, mo... | PaLM 2 Technical Report |
In summary, our key contributions are as follows:
• We are the first to propose an end-to-end pre-training
paradigm that learns to index into a large-scale memory
to solve knowledge-intensive visual-language tasks.
• Our method can construct a large-scale memory by en-
coding various sources of multimodal world knowle... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
0.7%
13B
Table 5: Comparison of models with and without FIM training. pass@1, pass@10 and pass@100
scores on HumanEval and MBPP evaluated at temperature 0.1 for models trained with and without infilling
(FIM) objective. Infilling training incurs no cost on autoregressive test set loss, but a small cost on HumanEval
a... | CodeLlama2 |
4096
32
128
14336
32
8
4096
8192
32000
Table 1: Model architecture.
1https://github.com/mistralai/mistral-src
2https://github.com/skypilot-org/skypilot
3https://huggingface.co/mistralai
2
Figure 2: Rolling buffer cache. The cache has a fixed size of W = 4. Keys and values for position i are stored
in position i mo... | Mistral7B |
Nos. 1922658 and 2046556. Any opinions, findings, and
conclusions or recommendations expressed in this material
are those of the author(s) and do not necessarily reflect the
views of the National Science Foundation. | Eight Things to Know about Large Language Models |
conditional control to text-to-image diffusion models.
CVPR, pages 3836–3847, 2023. 2, 3, 6
[81] Yanzhe Zhang, Lu Jiang, Greg Turk, and Diyi Yang. Audit-
ing gender presentation differences in text-to-image models.
arXiv preprint arXiv:2302.03675, 2023. 11
[82] Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao S... | VideoPoet |
Note, though, this requires that the planning performed by an APS system engaged in misaligned
behavior be limited or “pruned” in a specific way. That is, by hypothesis, such a system is using a
broad-scope world model, capable of accurately representing the causal upshot of different forms of
power-seeking, to plan in ... | Is Power-Seeking AI an Existential Risk? |
Formal verification of smart contracts ................................ 17
From Requirements to Models Using Natural Language Processing ........... 17
1
Human data interaction .......................................... 18
Human and social factors in information systems ........................ 19
Learni... | informatics-phd-projects-2022-23 |
friendly
extraverted
talkative
bold
assertive
active
energetic
adventurous and daring
cheerful
trustful
moral
honest
kind
generous
altruistic
cooperative
humble
sympathetic
unselfish
agreeable
self-efficacious
orderly
responsible
hardworking
self-disciplined
practical
thrifty
organized
conscientious
thorough
tense
nerv... | PersonalityTraitsinLargeLanguageModels |
information (in his study,
When we think about social media sites, undoubtedly the main way in which
they impact our daily lives is by making it easy to stay in touch with people we
would not see in person regularly. In other words, they entail greater exposure
and contact with weak ties than in offline interactions (G... | Social_Media_and_Democracy |
There exists a soldier such that for
every general, he is a general.
– Section: Movie tickets
– Section: A fun game console
– Section: Personalized items with
photos/artwork
– ...(more sections)
– Takeaway: Don’t stress about out running out
of time to buy, make a gift.
– Introduction
– List of Gift Ideas
– Conclu... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
problems in ml safety.
with outlier exposure.
[Hendrycks and Gimpel, 2016] Hendrycks, D. and Gimpel, K. (2016). A baseline for detecting misclassified
and out-of-distribution examples in neural networks.
[Hendrycks et al., 2018] Hendrycks, D., Mazeika, M., and Dietterich, T. (2018). Deep anomaly detection
[Henighan... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
[49] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese
BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Lan-
guage Processing and the 9th International Joint Conference on Natural Language Processing
(EMNLP-IJCNLP), pages 3982–3992, Hong Kong, China, 2... | E5 |
8 Proj. 0.25 8 ✓ 0.99 1.69 4.04 10.00 99%
1536 snake ✓ ✗ 5 ✓
8 ✓ 1.01 1.75 4.03
8 Proj. 0.5
1536 snake ✓ ✗ 5 ✓
99%
1536 snake ✓ ✗ 5 ✓
8 Proj. 1.0
24 ✓ 0.73 1.62 4.16 13.83 99%
99%
8
8 Proj. 1.0
1536 snake ✓ ✗ 5 ✓ | RVQGAN |
may be significantly improved via domain-specific objectives and finetuning [83, 84, 64, 65, 42].
Limitations & Future Work. Today, the inference costs (and monetary costs) of using LLMs in the control
loop are quite high. Predicting the next token for every sequence, e.g., every dimension of every time step
in a traje... | LargeLanguageModelsasGeneralPatternMachines |
# skill manager for adding new skills and skill
agent_state = environment . reset ()
while True :
exploration_progress = (
curriculum_agent . get_exploration_progress (
curriculum_agent . get_completed_tasks () ,
curriculum_agent . get_failed_tasks () ,
)
task = curriculum_agent . propose_next_task (
agent_state ... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
Spaces for Inversion and Personalization. Numerous
works have already analyzed the latent spaces of pretrained
text-to-image diffusion models [11, 15, 22, 47]. Most rele-
vant to our work is the text-conditioning space of the pre-
trained text-to-image model. In Textual Inversion, Gal et
al. [9] invert a given concept ... | A Neural Space-Time Representation for Text-to-Image Personalization |
PM dataset and training details are provided in Appendix A.2; we also discussed the performance of our PMs
in Section 3. In the language of RL, each response generated by the policy is a ‘timestep’, a full conversation
is one ‘trajectory’, and the PM score is a single ‘reward’ provided at the end.
The idea is to use th... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
the previous/next page, deciding to purchase, etc. We use the dataset provided by WebShop and randomly
sample 100 test instances, which cover instructions about various customers’ needs with specific requirements
of commodities’ attributes. | Tool Learning with Foundation Models |
S´ebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Ka-
mar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general
intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023.
Nicholas Carlini, Florian Tramer, Eric Wallace, Ma... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
∗
Primary authors. Correspondence to: Karl Cobbe <karl@openai.com>
1
One effective method involves training reward models to discriminate be-
tween desirable and undesirable outputs. The reward model can then be used
in a reinforcement learning pipeline (Ziegler et al., 2019; Stiennon et al., 2020;
Nakano et al., 2... | Let’s Verify Step by Step |
so τi is R↓ and C↓ due to the definition of reduce labels operations.
It follows that each opi implements an M↑R(cid:14)C(cid:14) transformation, so it follows by repeated application of Theorem 63 that
the composite transformation τ = τ1◦τ2◦. . .◦τm is M↑R(cid:14)C(cid:14). That is, M&S abstraction is M↑R(cid:14)C(cid... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
Definition 38 (GIDL). Let F1 = (cid:3)V 1, D1, A1(cid:4) and F2 = (cid:3)V 2, D2, A2(cid:4) be two SAS+ instances 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 a GIDL transformation from F1 to
F2 if there is a b... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
ZENY: Documentation has to be light-weight.
SOCART:
. . . but preregistration would get people more
invested in their ideas and bias them in how re-
sults are interpreted. When people go on record
with a study description, they will defend why it’s
reasonable and likely leading to a positive result.
Researchers are al... | A Two-Sided Discussion of Preregistration of NLP Research |
p
r
e
s
e
n
t
a
t
i
o
n
s
B
a
u
,
D
.
,
Z
h
o
u
,
B
.
,
K
h
o
s
l
a
,
A
.
,
O
l
i
v
a
,
A
.
a
n
d
T
o
r
r
a
l
b
a
,
A
.
,
2
0
1
7
.
P
r
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
E
E
E
c
o
n
f
e
r
e
n
c
e
o
n
c
o
m
p
u
t
e
r
v
i
s
i
o
n
a
n
d
p
a
t
t
e
r
n
r
e
c
o
g
n
i
t
i
o
n
,
p
p
.
... | Language models can explain neurons in language models |
Nicolas Patry. Making automatic speech recognition work on large files with Wav2Vec2 in Trans-
formers. https://huggingface.co/blog/asr-chunking, 2022. Accessed: 25 Oct.,
2023.
Zilun Peng, Akshay Budhkar, Ilana Tuil, Jason Levy, Parinaz Sobhani, Raphael Cohen, and Jumana
In Proceedings of the Second
Nassour. Shrinking... | DISTIL-WHISPER |
3.2 Controlled study across scales
We instruction finetune a range of FLAN-MOE models at batch size 32 and sequence length 2048 for
200k steps. This matches the number of training examples used for FLAN-T5 [4]. We re-finetuning
our own FLAN-T5 variants for fair comparisons.
Dense Model Size. Figure 2 shows the perfor... | Mixture-of-Experts |
The hash encoding uses multi-resolution grids, with each
grid cell corner mapped to a hash entry. Each hash entry
stores the encoding feature. Let {V1, ..., VL} be the set of dif-
ferent spatial grid resolutions. Given an input position xi, we
map it to the corresponding position at each grid resolution
Vl as xi,l = xi... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
Making Slides | Tool Learning with Foundation Models |
22
A Cooperative Role-Playing: The Good Mind
Below we provide an interesting example where a python programmer (assistant) is collaborating
with a stock trader (user) on developing a trading bot for the stock market.
Trading Bot Example: Python Programmer & Stock Trader
Original idea prompt: Develop a trading bot ... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
3The implementations are as follows:
Tacotron 2 : https://github.com/NVIDIA/tacotron2
Glow-TTS : https://github.com/jaywalnut310/glow-tts
HiFi-GAN : https://github.com/jik876/hifi-gan
Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech
GAN to both Tacotron 2 and Glow-TTS.
As e... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
W.-N. Hsu. Textless speech-to-speech translation on real data. arXiv:2112.08352, 2021b.
A. Lee, P.-J. Chen, C. Wang, J. Gu, S. Popuri, X. Ma, A. Polyak, Y. Adi, Q. He, Y. Tang, P. Juan,
and W.-N. Hsu. Direct speech-to-speech translation with discrete units. In Proc. ACL, pages
3327–3339, 2022.
C.-C. Lo, S.-W. Fu, W.-... | Translatotron3 |
process memory for real-time anomaly detection of attacks occurring in memory.
Better Error Help Using Large Scale Programmer Data
Supervisors: Professor Michael Kolling & Dr Neil Brown
Could large scale beginning programmer data be used to give useful hints and help to beginners stuck
on an error? For exampl... | informatics-phd-projects-2022-23 |
2. Preliminaries
In order to lay the relevant ground for our analytical study, a preliminary step consists in establishing a working defini-
tion for explainability and provide the main notions regarding knowledge graphs. We achieve this by summarising the main
theories around explanations with an historical overview... | Knowledge graphs as tools for explainable machine learning: A survey |
and denote the generated instructions as (cid:98)XXXSI. If
the selected instructions are associated with the in-
puts, they are concatenated using a colon “:” sym-
bol to form the format “$instruction:$input”.
For P3 and FLAN, we sample three random exam-
ples from the same subset, as we observe that if
the sampled ex... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
Lockhart, E., Osindero, S., Rimell, L., Dyer, C., Vinyals, O., Ayoub, K., Stanway, J., Bennett, L., Hassabis,
D., Kavukcuoglu, K., and Irving, G. (2021). Scaling language models: Methods, analysis & insights from
training gopher. CoRR, abs/2112.11446. | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea
Vedaldi. 2017. Learning multiple visual domains
with residual adapters. In Advances in Neural Infor-
mation Processing Systems, volume 30, pages 506–
516. Curran Associates, Inc.
Timo Schick and Hinrich Sch¨utze. 2020. Exploiting
cloze questions for few shot text classi... | Prefix-Tuning |
Paul Michel and Graham Neubig. Extreme adaptation for personalized neural machine translation.
In
Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2:
Short Papers), pp. 312–318, Melbourne, Australia, 2018. Association for Computational Linguistics. doi:
10.18653/v1/P18-20... | Tool Learning with Foundation Models |
zmans = WT
αi =
(cid:80)
sans ; h(T )
b [h(T )
tans ]
exp (oT
i zmans)
ol∈bj
exp (oT
l zmans)
where Wb is a transformation matrix distinct from
We in Eq. 1 and Wf in Eq. 4. The top k tail sets bj
are further aggregated using weights βj, which are
the softmax of the retrieval (inner product) scores
of the top k he... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
a
n
d
a
r
e
m
o
d
i
f
i
e
d
s
o
m
e
w
h
a
t
w
h
e
n
u
s
i
n
g
t
h
e
s
t
r
u
c
t
u
r
e
d
c
h
a
t
c
o
m
p
l
e
t
i
o
n
s
A
P
I
.
F
o
r
f
u
l
l
d
e
t
a
i
l
s
s
e
e
o
u
r
c
o
d
e
b
a
s
e
.
[
↩
]
| Language models can explain neurons in language models |
gpt-4. arXiv preprint arXiv:2304.03277, 2023.
Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine
Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables
zero-shot task generalization. arXiv preprint arXiv:2110.08207, 2021.
Maarten ... | Self-AlignmentwithInstructionBacktranslation |
Since we consider only pose changes across views and
neglect surface detail deformations (e.g., wrinkle move-
ments) across frames, challenging pose deviation and incon-
sistent geometry between frames will cause inaccurate fea-
ture fusion and reconstruction artifacts. Besides, the multi-
image network is trained usin... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
Kelvin Jiang, Dekun Wu, and Hui Jiang. 2019. Free-
baseqa: A new factoid qa data set matching trivia-
In Pro-
style question-answer pairs with freebase.
ceedings of the 2019 Conference of the North Amer-
ican Chapter of the Association for Computational
Linguistics: Human Language Technologies, Vol-
ume 1 (Long and Sho... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan
Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source
chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023.
[11] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubo... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
(3) Multimodal Speech Models: Traditional speech and text models have typically operated
within a single modality, focusing solely on either speech or text inputs and outputs. How-
ever, as the scale of generative models continues to grow exponentially, the integration
of multiple modalities becomes a natural progressi... | AReviewofDeepLearningTechniquesforSpeechProcessing |
[140] Medeiros, L.F., Kolbe Junior, A., Moser, A.: A cognitive assistant that uses small
talk in tutoring conversation. International Journal of Emerging Technologies
in Learning (iJET) 14(11), 138–159 (2019) https://doi.org/10.3991/ijet.v14i11.
49
10288
[141] Jain, M., Kumar, P., Kota, R., Patel, S.N.: Evaluating ... | PersonalityTraitsinLargeLanguageModels |
• Curriculum Tool Learning. Another approach to improving model generalization is through curriculum
learning (Bengio et al., 2009), which starts with simple tools and gradually introduces the model to more
complex tools so that it can build upon its prior knowledge and develop a deeper understanding of the
tool. For i... | Tool Learning with Foundation Models |
https://doi.org/10.1038/s42256-020-00256-0
online experiment builder. https://doi.org/10.5281/zenodo.5233003
Ergonomics Society Annual Meeting 50, 9 (Oct. 2006), 904–908. https://doi.org/10.1177/154193120605000909
Psychological Methods 23, 3 (2018), 561–569. https://doi.org/10.1037/met0000131
[30] Felix Henninger, ... | AI enhance sour performance |
Efficient LLM Algorithmic Survey, Nov, 2023, USA.
process can help mitigate issues of data imbalance. For instance, Prusa et al. [208] demonstrated the effectiveness of random
undersampling in majority classes, which serves dual purposes: it reduces redundancy in the training set and balances the
data distribution acr... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
6 Conclusion | Mixture-of-Experts |
3.4 Date-Efficient Learning
Based on explorations of the impact of data quan-
tity, data quality, and task composition on model
performance discussed previously, many works pro-
pose to fine-tune LLM more efficiently with subset
selection or learning strategy addressing different
aspects of instruction data.
Data Quan... | DataManagementForLargeLanguageModels-ASurvey |
C
o
m
p
o
s
i
t
i
o
n
a
l
i
t
y
S
o
l
v
i
n
g
s
i
m
p
l
e
q
u
e
s
t
i
o
n
s
m
i
g
h
t
r
e
q
u
i
r
e
m
u
l
t
i
p
l
e
s
t
e
p
s
,
f
o
r
e
x
a
m
p
l
e
-
“
D
o
m
o
r
e
p
e
o
p
l
e
l
i
v
e
i
n
T
e
l
A
v
i
v
o
r
i
n
B
e
r
l
i
n
?
”
r
e
q
u
i
r
e
s
a
n
s
w
e
r
i
n
g
:
i
.
W
h
a
t
... | Jurassic-X_ Crossing the neuro-symbolic chasm with the MRKL system |
Acknowledgments
We are grateful to Stability AI for providing the compute
required to train these models, and to CoreWeave for pro-
viding compute for some of the evaluations. OW’s contribu-
tions are financed by the Dutch Research Council (NWO)
as part of project 406.DI.19.059.
We thank Nora Belrose, Tim Dettmers, Perc... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
In terms of
text granularity, beyond the common
chunks (including sentences), the retrieval unit can be to-
kens (e.g., kNN-LM[Khandelwal et al., 2019]), phrases (e.g.,
NPM[Lee et al., 2020], COG[Vaze et al., 2021]), and docu-
ment paragraphs. Finer-grained retrieval units can often bet-
ter handle rare patterns and o... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
Gutierrez-Osuna. 2018. L2-ARCTIC: A non-native English speech corpus.. In Interspeech. 2783–2787.
[662] Hongyu Zhao, Hao Tan, and Hongyuan Mei. 2022. Tiny-Attention Adapter: Contexts Are More Important Than the
Number of Parameters. arXiv preprint arXiv:2211.01979 (2022).
[663] Shengkui Zhao and Bin Ma. 2023. MossFo... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Open Problems | Tool Learning with Foundation Models |
citing indication of how general purpose generative models such as Imagen Video can significantly
decrease the difficulty of high quality content generation. | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.