text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
ϕV LN
4.2. Touchdown
Experiment Design: [5] define two separate tasks in the
Touchdown benchmark: VLN and spatial description res-
olution. This research aligns with other studies [43, 42]
in conducting evaluation on the navigation component as a
standalone task. Dataset and Data Preprocessing: Frame-
works are evalu... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
Table 1: Statistics of datasets for pretraining. “#Image” represents the total number of distinct images, and
“#Sample” represents the number of training samples (e.g., the image-caption pair).
Type
Pretraining
Vision &
Language
Captioning
VQA
Detection
Vision
Image Filling
Source
MedICat
IU X-ray
Peir Gross
S... | BiomedGPT |
frequently employs LLM as an interactive planner, harness-
ing its self-updating capabilities to enhance the plan’s exe-
cutability over time [Wang et al., 2023a, Shinn et al., 2023,
Sun et al., 2023]. Inner Monologue [Huang et al., 2022a]
pilots the front of interactive planning with LLMs, which
introduces the feedbac... | JARVIS-1 |
nAcc@3
1.62±0.02
1.48±0.09
1.53±0.08
(a) HPO-B
(b) PD1
Retrieved by
Text embedding
Meta-feature
Random
Rank@1↓
59.74±1.89
50.49±6.38
57.95±10.19
Rank@2 ↓ Rank@3 ↓
25.58±5.09
38.67±2.47
42.07±3.35
34.00±3.69
32.16±4.47
42.78±8.91
AP@1↑
91.38±0.05
91.50±0.07
91.43±0.09
AP@2 ↑
91.60±0.03
91.60±0.03
91.59±0.08
A... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
violence.” The goal of banning hate speech from more mainstream online
platforms is to reduce the likelihood that everyday internet users are
incidentally exposed to online hate speech. | Social_Media_and_Democracy |
E-12B) model 87.3% of its NLG performance (relative) has
degraded during multimodal training, merely 3.9% have
been degraded for the largest model (PaLM-E-562B).
7. Summary of Experiments & Discussion
Generalist vs specialist models – transfer. As summa-
rized in Fig. 3, we have shown several instances of transfer
in t... | PaLM-E- An Embodied Multimodal Language Model |
Leblond et al. AlphaCode 2 Technical Report. 2023. URL https://storage.googleapis.com/
deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf.
Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436–444, 2015.
Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Le... | gemini_1_report |
version of DocLLM.
7
Table 4: Model configuration and training hyperparameters setting for DocLLM-1B and -7B.
Backbone
Layers
Attention heads
Hidden size
Precision
Batch size
Max context length
Learning rate
Warmups
Scheduler type
Weight decay
Adam βs
Adam epsilon
DocLLM-1B
Falcon-1B [5]
24
16
1536
bfloat16
2
... | DOCLLM |
5.2 CHOOSING THE CAPACITY FACTOR AND ROUTING ALGORITHM | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
2023.
[647] Guo, Y., Y. Yang, A. Abbasi. Auto-debias: Debiasing masked language models with automated
biased prompts. In S. Muresan, P. Nakov, A. Villavicencio, eds., Proceedings of the 60th Annual
Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL
2022, Dublin, Ireland, May 22-27, 2... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Jack told Mary, ”If you
give me your banana,
I’ll give you my apple”.
Mary gave Jack her Ba-
nana so
On weekends Jack went
to visit his grandmother
whereas on weekdays he
would go to school. Last
weekend, when Jack was
on his way to
Lily and Ben were hav-
ing an argument. Ben
said that cake is much
better than ice cr... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
(cid:1)β
N
8
phone set, which is a modified version of the international phonetic alphabet (IPA). Word position
postfixes are added. Audio is represented as a 80-dimensional log Mel spectrogram and a HiFi-GAN
vocoder trained on the same 60K hours of English speech is used to generate waveform. More details
about ph... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
that ID-PT still generated somewhat variable prompts for examples within each dataset hints that
the method may have the ability to offer gains even in the single-task regime. Contemporary works
by Tang et al. (2022); Jin et al. (2022) further investigate this possibility. | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
Information Processing Systems 35 (2022), 22300–22312.
[11] Shun-ichi Amari, Naotake Fujita, and Shigeru Shinomoto. 1992. Four types of learning curves. Neural Computation 4, 4 (1992), 605–618.
[12] Yuvanesh Anand, Zach Nussbaum, Brandon Duderstadt, Benjamin Schmidt, and Andriy Mulyar. 2023. Gpt4all: Training an assis... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
Figure 12. Training pipeline of PoseNet. To train PoseNet,
we use DesePose CSE surface embeddings, which is pertained on
2D annotations of human and quadruped animals. We first gen-
erate random viewpoints on a sphere that faces the origin. Then
we render surface embeddings as 16-channel images. We further
augment the ... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
neural networks and analytical hierarchy process, DATA ANALYTICS 2016
(2016) 69.
[27] K.K.Chandriah,R.V.Naraganahalli,RNN/LSTMwithmodifiedAdamoptimizer
in deep learning approach for automobile spare parts demand forecasting,
Multimedia Tools Appl. (2021) 1–15.
[28] R. Fildes, P. Goodwin, Stability in the inefficient us... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
38 Capable of operating, carrying out sequences of actions, or making decisions without human intervention.
39 AI agents: AI systems that autonomously perform multiple sequential steps – sometimes including actions like
browsing the internet, sending emails, or sending instructions to physical equipment – to try and co... | Capabilities and risks from frontier AI |
To stabilize RL training, we use Proximal Policy Optimization (PPO) [Schulman et al., 2017]. We also follow
other work [Stiennon et al., 2020] and apply an empirically-estimated KL penalty term in the reward, with
the total reward given by
(4.1)
where λKL ≥ 0 is a hyperparameter. In practice we use a very small value o... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
3. Inference
the
For
Process:
half of
the first
infer-
ence operation,
the
matrix[:num_rows,:d_model] is used as the
’up project’, and the transposed second half,
matrix[:num_rows,d_model:].transpose(),
serves as the ’down project’. This configuration
is possible because the order of neurons in the
intermediate o... | LLM in a flash |
scan over temperatures and two top-p settings for both the RLHF models and the base code models, and then
chose the best setting for each model and pass@k. We did a grid-search over the evaluation hyperparameters:
T ∈ {0, 0.4, 0.6, 0.8, 1.0} × p ∈ {0.95, 1} × k ∈ {1, 5, 10, 25, 50, 75, 100}. Results are summarized on t... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
To better understand the influence of multilingual pre-training, we measure the correlations between each of
the evaluated languages and report the results separately for different model sizes in Figure 3. We observe
high correlation between model performance on C++, C#, Java, and PHP. Interestingly, we also notice
str... | CodeLlama2 |
Few-shot
[Jarvis and Allard, 2023] Colin
A survey of
and
John Al-
Jarvis
techniques
for maximizing
https://community.openai.com/
lard.
llm performance.
t/openai-dev-day-2023-breakout-sessions/505213#
a-survey-of-techniques-for-maximizing-llm-performance-2,
2023.
[Jiang et al., 2023a] Huiqiang Jiang, Qianhui Wu, Ch... | RAG forLargeLanguageModels-ASurvey |
Fine-tuning Language Models for Factuality:
(Tian et al., 2023) address hallucination by
leveraging recent NLP innovations, employing
automated fact-checking methods and preference-
based learning through the Direct Preference
Optimization algorithm. The researchers fine-tune
the Llama-2 model for factuality without hu... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
• Validation Module: In real-world scenarios, it is not
always guaranteed that the retrieved information is reli-
able. Retrieving irrelevant data may lead to the occur-
rence of illusions in LLM. Therefore, an additional val-
idation module can be introduced after retrieving docu-
ments to assess the relevance betwee... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
4.2 Evaluation
In order to assess the universal understanding capabilities of Qwen-Audio, as shown in Table 2, we perform
a comprehensive evaluation that encompasses various tasks, namely Automatic Speech Recognition (ASR),
Speech-to-Text Translation (S2TT), Automatic Audio Captioning (AAC), Acoustic Scene Classificati... | Qwen-Audio |
Name
Architectures
Pre-training
NQ
(79k/4k)
BERT
Sparse Retr.+Transformer
BERT-Baseline (Lee et al., 2019)
T5 (Multitask)
Transformer Seq2Seq
T5 (base) (Roberts et al., 2020)
T5 (Multitask)
Transformer Seq2Seq
T5 (large) (Roberts et al., 2020)
T5 (Multitask)
Transformer Seq2Seq
T5 (11b) (Roberts et al., 2020)
N/A
... | REALM |
3.2.2 Planning with Reasoning
As discussed in § 3.2.1, understanding the intent and tools lays a solid foundation for planning. Nevertheless,
it is still insufficient for tackling intricate tasks. The user query q often implies a complex task that should be
divided into multiple sub-tasks with proper sequencing, thereb... | Tool Learning with Foundation Models |
col="12"
state="vm.form.inputs"
lose">
<at-form state="vm.form" autocomplete="off" id="external_test_form">
<at-input-group
tab="20"
form-
id="external_test"></at-input-group>
<at-action-group col="12" pos="right">
<at-action-button
variant="tertiary"
ng-click="vm.onClose()"
>
{{::vm.strings.get(’CLOSE’)}}
</at-actio... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
[22] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis,
Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
Dense passage retrieval for open-domain question answering.
In Proceedings of the 2020 Conference on Empirical Methods
in Natural Language Processing (EMNLP), pages 6769–6781,
Online, 2020. Association... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
Representation of Information. Computers Helping People with Special Needs 12376 (2020), 146–156.
[49] Franklin Mingzhe Li, Di Laura Chen, Mingming Fan, and Khai N. Truong. 2021. “I Choose Assistive Devices That Save My Face”: A
Study on Perceptions of Accessibility and Assistive Technology Use Conducted in China. Pro... | Society’sAttitudesTowardsHumanAugmentation |
users migrate to more extreme platforms, as well as whether they indeed become
further radicalized on these platforms (Jackson 2019). | Social_Media_and_Democracy |
classifiers, called probes, are trained to learn a
mapping between said representations and the cor-
responding attribute information (Conneau et al.,
2018; Hupkes and Zuidema, 2018), exemplified
here by PoS details. Considerable efforts have
been dedicated to investigating the attributes of
classifier that are used as... | AreEmergentAbilitiesinLarge Language Models just In-Context |
first piece of legislation in the world that mandates public transparency
reporting for major platforms for user-generated content. All firms defined as
operating social networks with more than 2 million users in Germany
(Facebook, Google, Twitter, and Change.org; the law excludes peer-to-peer
messaging services like What... | Social_Media_and_Democracy |
Semantic understanding refers to the meaning or understanding of language and its associated
concepts. It involves the interpretation and comprehension of words, phrases, sentences, and the
relationships between them. Semantic processing goes beyond the surface level and focuses on
understanding the underlying meaning ... | ASurveyonEvaluationofLargeLanguageModels |
D.4 STRATEGIES FOR RELIABLE LONG-FORM TRANSCRIPTION
Transcribing long-form audio relies on the accurate prediction of multiple chunks of audio in paral-
lel. Since long-form audio typically contains instances of long pauses between spoken utterances,
the Whisper model has a higher propensity to hallucinate compared to... | DISTIL-WHISPER |
As hinted by the proof sketch given in the main text, this proof consists of three main parts — (i)
construction of the optimal variable order π∗ given a smooth and structured-decomposable PC, (ii)
justify the correctness of Alg. 3, and (iii) prove that Fπ∗ (x) can be computed by evaluating no more
than O(log(K)·|p|) P... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
is used for “informative".
they are underrepresented with only 12 and 29 prompts, respectively.
69 | Llama2 |
C. Lawrence Zitnick. 2014. Microsoft COCO: Common Objects in Context. ArXiv abs/1405.0312 (2014).
[112] Ce Liu, Heung-Yeung Shum, and William T Freeman. 2007. Face Hallucination: Theory and Practice. International
Journal of Computer Vision 75, 1 (2007), 115–134.
[113] Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao... | SurveyofHallucinationinNatural Language Generation |
RNN, unidirectional LSTM-RNN, and vanilla RNN. RNNs,
and in specific LSTM, are especially successful in processing
sequential data (human language) and catching significant
features out of diverse data sources. Further, in Sections V-B1
and V-B2, we discuss LSTM and GRU. | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
encountered some stability issues with RL, and although we performed some rudimentary hyperparameter
scans, we expect that with more experience and study we could do better. We also did not explore variations
in online training, such as literally updating a single PM or RLHF model; rather we retrained these models
from... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
}
}
// Method to create a new branch
void createBranch ( string branch ) {
// Create a new branch directory
directory ( branchPath () + "/" + branch );
// Create a new version file
ofstream versionFile ;
versionFile . open ( versionFilePath () + "/" + branch + ". txt ");
versionFile << "0" << endl ;
versionFile . cl... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
of these strategies in subsequent sections.
3.1 Reducing Data Transfer
Our methodology leverages the inherent sparsity
found in Feed-Forward Network (FFN) models, as
documented in preceding research. The OPT 6.7B
model, for instance, exhibits a notable 97% spar-
sity within its FFN layer. Similarly, the Falcon
7B mode... | LLM in a flash |
167For especially exotic versions of this, see Oesterheld (2017) and Fox (2020).
168Though the history of atrocities committed by strategic and intelligent humans does not seem comforting in
this respect; and note that the incentives at stake here depend crucially on an agent’s empirical situation, and on
its power rel... | Is Power-Seeking AI an Existential Risk? |
In this work, we present LaMini-LM, a collec-
tion of language models that are notably smaller
in size than most existing instruction-tuned mod-
els. We develop LaMini-LM models by employing
sequence distillation (also known as offline distilla-
tion) (Kim and Rush, 2016) from LLMs. Although
similar attempts have been m... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
A control experiment with the 582M parameter model had a supervised training phase of 1 through
5 digits and a self-learning phase of 6 through 21 digits. The training run is depicted in Figures D, D
and D.
21
+0+1+2+3+4Generalization beyond training0.00.20.40.60.81.0AccuracyAddition Accuracy w/o CoT (582M)3622Digits... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
guidance or by regulating the safety, legality, and ethical conduct of agents. This is particularly crucial
in specialized domains, such as medicine where data privacy concerns exist [457]. In such cases,
human involvement can serve as a valuable means to compensate for the lack of data, thereby facili-
tating smoother... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
the safe development of advanced AI. This includes techniques for interpretability, scalable oversight and governance, | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
3.3.1 Learning from Demonstrations
Models can be trained to mimic the behavior of human experts through imitation learning (Hussein et al.,
2017; Liu et al., 2018b; Baker et al., 2022). Behavior cloning (Bain & Sammut, 1995) can be viewed as a
simplistic form of imitation learning that focuses on learning policies in ... | Tool Learning with Foundation Models |
correlate with (but not guarantee) performance across many NLP tasks. We also include a divers set
of additional benchmarks. The CNN-DM (Hermann et al., 2015) and BBC XSum (Narayan et al.,
2018) datasets are used to measure the ability to summarize articles. Question answering is probed
with the SQuAD dataset (Rajpurka... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
[57] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. DocBank: A benchmark
dataset for document layout analysis. In Donia Scott, Nuria Bel, and Chengqing Zong, editors, Proceedings of
the 28th International Conference on Computational Linguistics, pages 949–960, Barcelona, Spain (Onli... | DOCLLM |
shared convolutional encoder network to form an input tuple
{Ij, Pj, Fj}. | DynIBaR-NeuralDynamicImage-BasedRendering |
“liberation technology,” 1
like-minded individuals, polarizing views
through communities of, 36
Lipset, Seymour Martin, 208
listener bots, 95
Lodge, Milton, 47
Lokot, T., 96–97
Luceri, L., 90
Lumen Database, 229, 237
339
Macedonia, disinformation source from, 13, 14
MacGregor, Sharon, 238
machine learning, 92
Magdy... | Social_Media_and_Democracy |
11This leads to more KB triples than entity pairs, since a
pair of entities can be connected by more than one relation.
12We experimented with other values of k during fine tuning
and evaluation but did not observe significant differences.
3690 all transformer layers and the four transformation
matrices: Wa, Wb, We, ... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
graph infomax. arXiv preprint arXiv:1809.10341 (2018).
[557] Emmanuel Vincent, Tuomas Virtanen, and Sharon Gannot. 2018. Audio source separation and speech enhancement.
John Wiley & Sons.
[558] Thilo von Neumann, Keisuke Kinoshita, Christoph Boeddeker, Marc Delcroix, and Reinhold Haeb-Umbach. 2021.
Graph-PIT: Genera... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Prior works for instruction following mainly inherit the capabilities from large (multimodal) LLMs and adopt
light-weight supervised fine-tuning to activate the abilities of the model to align with user intent (Ouyang
et al., 2022; Wang et al., 2023a; Gong et al., 2023b). However, most works have been constrained in te... | Qwen-Audio |
2. Explainable and human-in-the-loop robot motion planning
• modeling human expectations and human understanding of robot motion,
• developing new algorithms and user interfaces for explainable and human-in-the-loop
planning,
• conducting user studies to evaluate the effectiveness of explainable/human-in-the-... | informatics-phd-projects-2022-23 |
8 REPRODUCIBILITY STATEMENT
In the Supplementary Materials, we provide code to reproduce all experiments in this paper. More
specifically, this includes:
• Compressing all models from the OPT and BLOOM model families to 2/3/4 bits.
• Evaluating perplexity of the quantized models.
• Our 3-bit CUDA kernel together with ... | GPTQ |
• EfficientQA [229] is an open-domain Question Answering (QA) challenge at
NeurIPS 20203 that focuses on building accurate, memory-efficient QA systems. It
promotes efficient memory usage through three restrained tracks based on model
size and accuracy: the most accurate model under 6 GB, the most accurate model
under 500 MB... | Beyond Efficiency |
A.9NavigatingKnowledgeGraphsA.9NavigatingKnowledgeGraphsInstruction:1.find_entity_by_head(inputID)Findall<r,t>thathastherelation<input,r,t>.Itlookslikeviewingthemainpageoftheinputentity.TheinputhastobeEXACTLYONEID(eg.’Q42’)andresultisatable.2.find_entity_by_tail(inputID)Findall<h,r>thathastherelation<h,r,input>.Itlooksli... | Tool Learning with Foundation Models |
Wohlin [2] conducted a review of the literature related to explainable artificial intelligence systems, with a focus on
knowledge-enabled systems, including expert systems, cognitive assistants, semantic applications, and machine learning
domains. In this review, Wohlin proposed new definitions for explainable knowledg... | Knowledge-graph-based explainable AI- A systematic review |
advantages over 1x1 convolution or naive linear projection used in ViT (Wang et al., 2022d). For linguistic
inputs, we used byte-pair encoding (BPE) (Sennrich et al., 2016) to perform the subword tokenization. The
subwords are then embedded into the input features.
To handle diverse modalities without relying on task-s... | BiomedGPT |
Xti<latexit sha1_base64="76w10YEtETzUXdaT0wTZt0xBig8=">AAAB9XicbVDLSgMxFL1TX7W+qi7dBIvgqsyIoMuCG5cV7EPaacmkmTY0kxmSO0oZ+h9uXCji1n9x59+YtrPQ1gOBwzn3ck9OkEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3grGNzO/9ci1EbG6x0nC/YgOlQgFo2ilXjeiOArCrD3tYV/0yxW36s5BVomXkwrkqPfLX91BzNKIK2SSGtPx3AT9jGoUTPJpqZsanlA2p... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
Efficiency of following the optimal variable order We proceed to show that when using the
optimal variable order π∗, Alg. 3 evaluates no more than O(log(D)·|p|) PC units.
According to the previous paragraphs, whenever Alg. 3 evaluates a PC unit n w.r.t. vtree node v, it
will evaluate all PC units in ϕ(p, v). Therefore, ... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
CoRR, abs/2307.00184, 2023.
[512] Côté, M., Á. Kádár, X. Yuan, et al. Textworld: A learning environment for text-based games.
In T. Cazenave, A. Saffidine, N. R. Sturtevant, eds., Computer Games - 7th Workshop, CGW
2018, Held in Conjunction with the 27th International Conference on Artificial Intelligence,
IJCAI 2018,... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
gaft, described in McKenzie et al. (2022).
Eight Things to Know about Large Language Models
9.6. The science and scholarship around LLMs is
especially immature | Eight Things to Know about Large Language Models |
Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L.
Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts:
Deliberate problem solving with large language models, 2023.
12
14
JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models
Yue Wu, So Yeon Min, Shrimai Prabhumoy... | JARVIS-1 |
execute in-context learning with flipped labels (Wei
et al., 2023) becomes evident at a similar scale to
the emergence of abilities (Wei et al., 2022b) (see
also Section 1.2 and Figure 1). | AreEmergentAbilitiesinLarge Language Models just In-Context |
For prompt following and style, we assemble a small dataset of 170 captions for this evaluation which is
specifically targeted at typical usage of a production text-to-image system. These captions cover a wide array
of actual use-cases like generating humans, products and places, concept blending, text rendering and ar... | Improving Image Generation with Better Captions |
12
Universal Self-Consistency for Large Language Model Generation
Tianyi Zhang, Tao Yu, Tatsunori Hashimoto, Mike Lewis, Wen-tau Yih, Daniel Fried, and Sida Wang.
Coder reviewer reranking for code generation. In International Conference on Machine Learning,
pp. 41832–41846. PMLR, 2023a.
Xinghua Zhang, Bowen Yu, Hai... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
However,
this emphasis on the psychological profiles of political
conservatives is not without controversy. Kahan and colleagues contend that
motivated reasoning is not a uniquely right-wing phenomenon. Instead, all
individuals are motivated to express and maintain beliefs similar to those of
other members of their ide... | Social_Media_and_Democracy |
explanationtosupportpeopleonjustifyingtheirdecisions,2021,arXivpreprint
arXiv:2102.05460.
[37] O. Biran, K.R. McKeown, Human-centric justification of machine learning
predictions, in: IJCAI, vol. 2017, 2017, pp. 1461–1467.
[38] A. Adadi, M. Berrada, Peeking inside the black-box: a survey on explainable
artificial intel... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
count is accurate and averaged 0.232 less than those who
believe the count is undercounted. There were also differ-
ences in the conspiracy scale (F3,295 = 3.21, p = 0.023) and
trust in the government (F3,295 = 11.068, p < 0.001). Those
who believe the count is overstated had a higher average
conspiracy score by 0.... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
LLM fails at a task in some setting is not reliable evidence
that that LLM doesn’t have the skills or knowledge to do
that task. Often, once one finds an appropriate way to
prompt a model to do some task, one will find that the
model consistently performs well across different instances
of the task. The chain-of-thought ... | Eight Things to Know about Large Language Models |
49
Fig. 15. The architecture of the Generative Spoken Language Model GSLM introduced by Meta in [281].
GSLM model operates through a three-part architecture. Firstly, the encoder takes the speech waveform and
transforms it into distinct units represented as S2u. Secondly, the decoder reverses this mapping by convertin... | AReviewofDeepLearningTechniquesforSpeechProcessing |
changes with the introduction of instruction tuning (second and third scenario),
used independently or in conjunction with task-specific finetuning. Our most
powerful model, FLAN-MOE32B, surpasses the performance of FLAN-PALM62B
on four benchmark tasks, while using only a third of the FLOPs. The advance-
ments embodied... | Mixture-of-Experts |
PickScoreTraining RankerAutomated Win Rate vs. SD1.5HPSCLIPPickScoreHPSCLIPAesthetics0.30.51.0AestheticsMetric Model:y: DreamlikeDreamlikeDPO-Dreamlike*yw: Dreamlikey: SDXL-yw: SDXL-SDXLDPO-SDXL0.210.220.23Median Pickscore fline algorithm, we anticipate online learning methods to be
another driver of future performance... | DiffusionModelAlignmentUsing Direct Preference Optimization |
5.2 Text to Music Generation
For text-to-music generation, we use the evaluation set
from the MUCaps dataset. This set comprises 5,000 text-
music pairs. SOTA models selected for comparison in-
clude CoDi [61], AudioLDM 2 [46], and MusicGen [9].
Among these models, MusicGen is the sole one explicitly
trained for music... | M2UGen |
(2) The Oogiri game boasts a substantial corpus of manually annotated creative data. Due to its widespread popular-
ity on the Internet, the game attracts a large user base generating creative human responses which can constitute an extensive
dataset for LoT exploration;
(3) The Oogiri game facilitates visualization f... | Let’sThinkOutsidetheBox |
px square around the person, apply perspective undistortion
with camera intrinsics and perform augmentation as in [70].
Datasets. See Tab. 1 for an overview of all used datasets,
which employ a variety of skeleton formats. In some cases,
e.g., when annotations are derived through triangulating
COCO-like predictions (of... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
References
Albert Ziegler. Research recitation: A first look at rote learning in GitHub Copilot suggestions.
https://docs.github.com/en/github/copilot/research-recitation, 2021. Accessed:
2022-01-13.
M. Allamanis. The adverse effects of code duplication in machine learning models of code.
In
Proceedings of the 2019 ACM S... | alphacode |
In scenarios where DRAM is abundant, the cost
of loading data is somewhat mitigated, as the model
can reside in DRAM. However, the initial loading
of the model still incurs a penalty, particularly in sit-
uations requiring rapid response times for the first
token. Our approach, leveraging activation sparsity
in LLMs, a... | LLM in a flash |
2 Claude 2 Model Card | ClaudeModels |
214
Francis Fukuyama & Andrew Grotto
evolved during the 1980s and 1990s, this concern seemed increasingly
outdated. The efforts by liberals to reinstate the Fairness Doctrine were driven
in large measure by their unhappiness with the growth of Fox News, AM talk
radio, and a host of new conservative media outlets that... | Social_Media_and_Democracy |
of common-sense distinctions—for example, between the role that someone’s objectives play in
explaining a given action, vs. their world-models and capabilities. There are also various algorithms
that implement explicit procedures in the vein of (a) and (b)—algorithms, for example, that search
over and evaluate possible... | Is Power-Seeking AI an Existential Risk? |
Instruction
Q: From the examples above, what patterns can we observe about the relationship between dataset
characteristics and the best hyper-parameter configurations? Answer MUST be concise, critical,
point-by-point, line-by-line, and brief. Only include relevant observations without unnecessary
elaboration.)
Table ... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
LS (960h)
WJS (si284)
WJS (si284)
TIMIT
UnSpeech
[381]
ASR-Mult
GigaSpeech (10000h)
SUPERB
graph field, researchers have developed approaches like Deep Graph Infomax (DGI) (Velickovic
et al., 2019 [556]) to learn representations that maximize the mutual information between local
patches and global structures w... | AReviewofDeepLearningTechniquesforSpeechProcessing |
[266] Winston, P. H. Learning and reasoning by analogy. Commun. ACM, 23(12):689–703, 1980.
[267] Lu, Y., M. Bartolo, A. Moore, et al. Fantastically ordered prompts and where to find them:
Overcoming few-shot prompt order sensitivity. In S. Muresan, P. Nakov, A. Villavicencio,
eds., Proceedings of the 60th Annual Meeti... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. Did aristotle
use a laptop? A question answering benchmark with implicit reasoning strategies. Trans. Assoc.
Comput. Linguistics, 9:346–361, 2021. doi: 10.1162/tacl\_a\_00370. URL https://doi.org/
10.1162/tacl_a_00370.
Mohammad Javad Ho... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser. The commitmentbank: In-
vestigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung,
volume 23, pages 107–124, 2019.
Jeff Dean. Introducing pathways: A next-generation ai architecture. Google AI Blog, 2021.
Jia Deng, We... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
For problem tags and ratings conditioning, we picked random tags from the most popular 50 for the
model to condition on, and sampled ratings uniformly in the range of 800 to 3500 as these metadata
are not visible for new unseen problems in a competition. We found that conditioning on random
tags and ratings can improve... | alphacode |
Sachin Kumar, Vidhisha Balachandran, Lucille Njoo, Antonios Anastasopoulos, and Yulia Tsvetkov. Language
generation models can cause harm: So what can we do about it? an actionable survey. arXiv preprint
arXiv:2210.07700, 2022.
Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris ... | Llama2 |
Piotr Nawrot, Szymon Tworkowski, Michał Tyrolski, Łukasz Kaiser, Yuhuai Wu, Christian Szegedy,
and Henryk Michalewski. Hierarchical Transformers Are More Efficient Language Models.
arxiv:2110.13711[cs], April 2022. URL http://arxiv.org/abs/2110.13711.
Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yan... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
level into a prefix, a middle part and a suffix with the splitting locations sampled independently from a
uniform distribution over the document length. We apply this transformation with a probability of 0.9 and
to documents that are not cut across multiple model contexts only. We randomly format half of the splits in
... | CodeLlama2 |
Eric Lehman, Evan Hernandez, Diwakar Mahajan, Jonas Wulff, Micah J Smith, Zachary Ziegler, Daniel
Nadler, Peter Szolovits, Alistair Johnson, and Emily Alsentzer. Do we still need clinical language models?
arXiv preprint arXiv:2302.08091, 2023.
Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for param... | BiomedGPT |
Announcing Jurassic-2 and Task-Specific APIs
https://www.ai21.com/blog/introducing-j2
2/12 | Announcing Jurassic-2 and Task-Specific APIs |
[14] B. Collins, D. T. Hoang, N. T. Nguyen, and D. Hwang, ‘‘Trends in
combating fake news on social media—A survey,’’ J. Inf. Telecommun.,
vol. 5, no. 2, pp. 247–266, 2021.
[15] A. Zubiaga, A. Aker, K. Bontcheva, M. Liakata, and R. Procter, ‘‘Detec-
tion and resolution of rumours in social media: A survey,’’ ACM Compu... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
vised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020.
Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand
Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021.
Liang-Chieh Chen, Yukun ... | DINOv2- Learning Robust Visual Features without Supervision |
D
3
t
n
a
t
s
n
I
e
s
i
o
N
o
/
w
s
g
n
i
d
d
e
b
m
E
n
e
k
o
T
o
/
w
e
t
e
l
p
m
o
C
“a lion playing the
guitar wearing a suit
wearing a baseball
cap”
wearing a sweater
wearing a baseball
cap”
“a bear holding a book
“a lion holding a
“a bear riding a
“a lion wielding a
shovel wearing a cape
motorcycle wea... | Instant3D |
R
a
t
h
e
r
t
h
a
n
b
e
i
n
g
i
n
t
i
m
i
d
a
t
e
d
b
y
t
h
e
d
a
t
a
a
n
d
d
i
s
t
r
i
b
u
t
i
o
n
a
d
v
a
n
t
a
g
e
o
f
i
n
c
u
m
b
e
n
t
s
,
I
e
n
c
o
u
r
a
g
e
f
o
u
n
d
e
r
s
t
o
f
i
n
d
a
n
g
l
e
s
w
h
e
r
e
i
n
c
u
m
b
e
n
t
s
a
r
e
n
o
t
c
o
m
p
e
t
i
n
g
a
t
a
l
l
.
”
... | Product-Led AI _ Greylock |
0.0001 and used smaller batch size (capped at 200 utterances or 5K frames). In this case, 120 epochs
corresponded to about 120K updates. For decoding, we used n-best decoding with a beam-size of 15,
and evaluated the WER on the 1-best path. | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.