Dataset Viewer
Auto-converted to Parquet Duplicate
id
string
version
int64
title
string
authors
list
abstract
string
categories
list
published
timestamp[s]
updated
timestamp[s]
doi
string
license
string
synopsis
string
keywords
list
topics
list
related
list
text_source
string
pages
int64
paths
unknown
source
string
arxiv
dict
0901.2349
2
Beyond word frequency: Bursts, lulls, and scaling in the temporal distributions of words
[ "Eduardo G. Altmann", "Janet B. Pierrehumbert", "Adilson E. Motter" ]
Background: Zipf's discovery that word frequency distributions obey a power law established parallels between biological and physical processes, and language, laying the groundwork for a complex systems perspective on human communication. More recent research has also identified scaling regularities in the dynamics und...
[ "cs.CL", "cond-mat.dis-nn", "physics.data-an", "physics.soc-ph" ]
2009-01-15T21:12:12
2009-11-11T16:07:23
10.1371/journal.pone.0007678
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper studies temporal recurrence of frequent words, asking how distances between successive occurrences are distributed and whether they deviate from a Poisson process. Across USENET and other formal/informal corpora, recurrence times are bursty and fit stretched exponential (Weibull) scaling. The authors report t...
[ "temporal distributions of words", "bursts and lulls in word usage", "scaling in language dynamics", "semantic effects on word recurrence", "generative models of word usage", "collective human dynamics" ]
[ { "id": "quantitative-linguistics-and-statistical-text-analysis", "fit": 0.9985389082 }, { "id": "burstiness-and-temporal-clustering-in-event-sequences", "fit": 0.9971061871 } ]
[ { "id": "2610.34188", "score": 12.8277, "via": [ "quantitative-linguistics-and-statistical-text-analysis", "burstiness-and-temporal-clustering-in-event-sequences" ], "similarity": 0.8439 }, { "id": "1604.01125", "score": 6.4817, "via": [ "burstiness-and-temporal...
html
8
{ "dir": "papers/0901.2349", "guide": "papers/0901.2349/guide.json", "extraction": "papers/0901.2349/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/0901.2349v2", "pdf": "https://arxiv.org/pdf/0901.2349v2", "html": "https://arxiv.org/html/0901.2349v2" }
1312.6034
2
Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
[ "Karen Simonyan", "Andrea Vedaldi", "Andrew Zisserman" ]
This paper addresses the visualisation of image classification models, learnt using deep Convolutional Networks (ConvNets). We consider two visualisation techniques, based on computing the gradient of the class score with respect to the input image. The first one generates an image, which maximises the class score [Erh...
[ "cs.CV" ]
2013-12-20T16:45:54
2014-04-19T11:54:52
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper studies visualisation of image classification models learnt by deep Convolutional Networks (ConvNets). It presents two gradient-based techniques that use the gradient of the class score with respect to the input image: class model visualisation, which generates an image maximising the class score, and an imag...
[ "ConvNet visualisation", "class saliency maps", "weakly supervised object segmentation/localisation", "deconvolutional networks", "model interpretability", "ImageNet classification" ]
[ { "id": "interpretability-of-vision-models", "fit": 0.9982487249 }, { "id": "neural-network-feature-visualization", "fit": 0.996490058 }, { "id": "gradient-based-attribution", "fit": 0.9869821318 }, { "id": "interpretable-ml-feature-attribution", "fit": 0.9249905761 } ]
[ { "id": "2504.13763", "score": 15.3259, "via": [ "neural-network-feature-visualization", "interpretability-of-vision-models", "interpretable-ml-feature-attribution" ], "similarity": 0.7595 }, { "id": "2310.10348", "score": 10.2138, "via": [ "gradient-based...
html
8
{ "dir": "papers/1312.6034", "guide": "papers/1312.6034/guide.json", "extraction": "papers/1312.6034/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1312.6034v2", "pdf": "https://arxiv.org/pdf/1312.6034v2", "html": "https://arxiv.org/html/1312.6034v2" }
1401.3841
1
Narrative Planning: Balancing Plot and Character
[ "Mark Owen Riedl", "Robert Michael Young" ]
Narrative, and in particular storytelling, is an important part of the human experience. Consequently, computational systems that can reason about narrative can be more effective communicators, entertainers, educators, and trainers. One of the central challenges in computational narrative reasoning is narrative gener...
[ "cs.AI" ]
2014-01-16T04:54:07
2014-01-16T04:54:07
10.1613/jair.2989
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper argues that narrative understandability depends on logical causal plot progression and character believability, where characters must appear as intentional agents. It frames narrative generation as fabula planning and introduces IPOCL, a refinement-search partial-order causal-link planner that reasons about c...
[ "narrative planning", "balancing plot and character", "character intentionality in narrative generation", "refinement-search planning", "audience comprehension of character intentions", "fabula and sjuzet" ]
[ { "id": "automated-story-generation", "fit": 0.9949635899 }, { "id": "ai-for-computational-narrative-generation-and-understanding", "fit": 0.9889282357 }, { "id": "automated-planning-and-task-decomposition", "fit": 0.9036970004 } ]
[ { "id": "1909.03480", "score": 8.2551, "via": [ "automated-story-generation", "ai-for-computational-narrative-generation-and-understanding" ], "similarity": 0.7976 }, { "id": "2610.21468", "score": 8.2266, "via": [ "automated-story-generation", "ai-for-com...
pdf
52
{ "dir": "papers/1401.3841", "guide": "papers/1401.3841/guide.json", "extraction": "papers/1401.3841/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1401.3841v1", "pdf": "https://arxiv.org/pdf/1401.3841v1", "html": "https://arxiv.org/html/1401.3841v1" }
1402.1869
2
On the Number of Linear Regions of Deep Neural Networks
[ "Guido Montúfar", "Razvan Pascanu", "Kyunghyun Cho", "Yoshua Bengio" ]
We study the complexity of functions computable by deep feedforward neural networks with piecewise linear activations in terms of the symmetries and the number of linear regions that they have. Deep networks are able to sequentially map portions of each layer's input-space to the same output. In this way, deep models c...
[ "stat.ML", "cs.LG", "cs.NE" ]
2014-02-08T17:16:27
2014-06-07T19:56:14
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper presents a theoretical study of the complexity of functions computed by deep feedforward neural networks with piecewise linear activations, characterized through symmetries and the number of linear regions. It analyzes compositional maps to argue that depth enables exponential re-use of computation, derives b...
[ "expressivity and complexity of deep neural networks", "depth advantage", "piecewise linear networks", "linear regions and hyperplane arrangements", "rectifier and maxout networks", "visualization of learned representations" ]
[ { "id": "neural-network-representation-and-computation-analysis", "fit": 0.9991699569 }, { "id": "expressivity-and-representational-capacity-of-neural-architectur", "fit": 0.9981682457 } ]
[ { "id": "2207.02505", "score": 6.6812, "via": [ "expressivity-and-representational-capacity-of-neural-architectur" ], "similarity": 0.693 }, { "id": "2512.19941", "score": 4.8964, "via": [ "neural-network-representation-and-computation-analysis" ], "similarity...
html
17
{ "dir": "papers/1402.1869", "guide": "papers/1402.1869/guide.json", "extraction": "papers/1402.1869/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1402.1869v2", "pdf": "https://arxiv.org/pdf/1402.1869v2", "html": "https://arxiv.org/html/1402.1869v2" }
1410.5401
2
Neural Turing Machines
[ "Alex Graves", "Greg Wayne", "Ivo Danihelka" ]
We extend the capabilities of neural networks by coupling them to external memory resources, which they can interact with by attentional processes. The combined system is analogous to a Turing Machine or Von Neumann architecture but is differentiable end-to-end, allowing it to be efficiently trained with gradient desce...
[ "cs.NE" ]
2014-10-20T19:28:26
2014-12-10T16:01:39
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Neural Turing Machines (NTMs) couple a neural network controller to an external memory bank accessed through attentional read and write processes. The architecture is differentiable end-to-end and trained by gradient descent, making it analogous to a Turing Machine or Von Neumann architecture. The paper presents prelim...
[ "external memory augmentation", "attention-based memory access", "differentiable computing", "algorithm induction", "controller comparison", "generalization" ]
[ { "id": "attention-based-memory-and-computation-architectures", "fit": 0.9979459827 }, { "id": "general-purpose-neural-architectures", "fit": 0.9955722506 } ]
[ { "id": "2107.14795", "score": 11.8788, "via": [ "general-purpose-neural-architectures", "attention-based-memory-and-computation-architectures" ], "similarity": 0.7011 }, { "id": "1610.06258", "score": 6.078, "via": [ "attention-based-memory-and-computation-arch...
html
26
{ "dir": "papers/1410.5401", "guide": "papers/1410.5401/guide.json", "extraction": "papers/1410.5401/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1410.5401v2", "pdf": "https://arxiv.org/pdf/1410.5401v2", "html": "https://arxiv.org/html/1410.5401v2" }
1501.00092
3
Image Super-Resolution Using Deep Convolutional Networks
[ "Chao Dong", "Chen Change Loy", "Kaiming He", "Xiaoou Tang" ]
We propose a deep learning method for single image super-resolution (SR). Our method directly learns an end-to-end mapping between the low/high-resolution images. The mapping is represented as a deep convolutional neural network (CNN) that takes the low-resolution image as the input and outputs the high-resolution one....
[ "cs.CV", "cs.NE" ]
2014-12-31T08:35:09
2015-07-31T09:13:32
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper proposes SRCNN, a deep convolutional neural network for single-image super-resolution that learns an end-to-end mapping from low-resolution to high-resolution images. It shows that sparse-coding-based SR can be viewed as a deep convolutional network, but unlike component-wise methods it jointly optimizes all ...
[ "deep learning for image super-resolution", "CNN-based image restoration", "sparse coding and deep network equivalence", "lightweight CNN design and performance-speed trade-offs", "color image super-resolution", "benchmark evaluation of super-resolution" ]
[ { "id": "deep-image-restoration", "fit": 0.9997728584 } ]
[ { "id": "1609.05158", "score": 7.3054, "via": [ "deep-image-restoration" ], "similarity": 0.8964 }, { "id": "2004.11362", "score": 0.9129, "via": [], "similarity": 0.7975 }, { "id": "1312.6034", "score": 0.8518, "via": [], "similarity": 0.7776 }, ...
html
14
{ "dir": "papers/1501.00092", "guide": "papers/1501.00092/guide.json", "extraction": "papers/1501.00092/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1501.00092v3", "pdf": "https://arxiv.org/pdf/1501.00092v3", "html": "https://arxiv.org/html/1501.00092v3" }
1506.02169
2
Approximating Likelihood Ratios with Calibrated Discriminative Classifiers
[ "Kyle Cranmer", "Juan Pavez", "Gilles Louppe" ]
In many fields of science, generalized likelihood ratio tests are established tools for statistical inference. At the same time, it has become increasingly common that a simulator (or generative model) is used to describe complex processes that tie parameters $θ$ of an underlying theory and measurement apparatus to hig...
[ "stat.AP", "physics.data-an", "stat.ML" ]
2015-06-06T16:59:23
2016-03-18T16:47:11
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Proposes likelihood-free inference by training calibrated discriminative classifiers to approximate generalized likelihood ratios when a simulator or generative model can sample data but the likelihood is unavailable. The central claim is that likelihood ratios are invariant under a class of dimensionality-reducing map...
[ "classifier-based likelihood ratio approximation", "likelihood-free inference", "calibration and diagnostics", "composite hypothesis testing and nuisance parameters", "simulation-based inference", "high energy physics applications" ]
[ { "id": "likelihood-ratio-and-pseudo-likelihood-methods", "fit": 0.9987835286 } ]
[ { "id": "1906.02845", "score": 6.9587, "via": [ "likelihood-ratio-and-pseudo-likelihood-methods" ], "similarity": 0.7834 }, { "id": "1703.00854", "score": 0.8397, "via": [], "similarity": 0.7736 }, { "id": "1901.06852", "score": 0.8308, "via": [], "s...
html
35
{ "dir": "papers/1506.02169", "guide": "papers/1506.02169/guide.json", "extraction": "papers/1506.02169/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1506.02169v2", "pdf": "https://arxiv.org/pdf/1506.02169v2", "html": "https://arxiv.org/html/1506.02169v2" }
1506.02629
2
Generalization in Adaptive Data Analysis and Holdout Reuse
[ "Cynthia Dwork", "Vitaly Feldman", "Moritz Hardt", "Toniann Pitassi", "Omer Reingold", "Aaron Roth" ]
Overfitting is the bane of data analysts, even when data are plentiful. Formal approaches to understanding this problem focus on statistical inference and generalization of individual analysis procedures. Yet the practice of data analysis is an inherently interactive and adaptive process: new analyses and hypotheses ar...
[ "cs.LG", "cs.DS" ]
2015-06-08T19:34:29
2015-09-25T19:04:32
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper studies adaptive data analysis, where analysts repeatedly reuse data and choose hypotheses after seeing prior results. It proposes reusable-holdout methods (Thresholdout and SparseValidate) to validate many adaptively chosen hypotheses without overfitting, formalizes data reuse, and connects differential-priv...
[ "holdout reuse", "overfitting to the holdout set", "generalization under adaptivity", "differential privacy", "description length", "max-information" ]
[]
[ { "id": "1605.07723", "score": 0.8715, "via": [], "similarity": 0.784 }, { "id": "1506.02169", "score": 0.8036, "via": [], "similarity": 0.7619 } ]
html
29
{ "dir": "papers/1506.02629", "guide": "papers/1506.02629/guide.json", "extraction": "papers/1506.02629/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1506.02629v2", "pdf": "https://arxiv.org/pdf/1506.02629v2", "html": "https://arxiv.org/html/1506.02629v2" }
1604.01125
3
Measuring burstiness for finite event sequences
[ "Eun-Kyeong Kim", "Hang-Hyun Jo" ]
Characterizing inhomogeneous temporal patterns in natural and social phenomena is important to understand underlying mechanisms behind such complex systems, hence even to predict and control them. Temporal inhomogeneities in event sequences have been described in terms of bursts that are rapidly occurring events in sho...
[ "physics.soc-ph", "physics.data-an" ]
2016-04-05T03:42:19
2016-09-19T02:10:48
10.1103/PhysRevE.94.032311
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper analyzes how the widely used burstiness parameter for event sequences is distorted by finite sequence length. It derives finite-size behavior for regular, random, and extremely bursty reference cases, proposes a simple alternative burstiness measure that accounts for sample size and separates finite-size arti...
[ "finite-size effects in burstiness measurement", "bursty dynamics", "temporal inhomogeneities", "heavy-tailed interevent-time distributions", "statistical estimation of burstiness", "boundary conditions in event-sequence models" ]
[ { "id": "burstiness-and-temporal-clustering-in-event-sequences", "fit": 0.99936342 } ]
[ { "id": "0901.2349", "score": 6.4817, "via": [ "burstiness-and-temporal-clustering-in-event-sequences" ], "similarity": 0.7601 }, { "id": "2610.34188", "score": 6.4136, "via": [ "burstiness-and-temporal-clustering-in-event-sequences" ], "similarity": 0.7379 ...
html
7
{ "dir": "papers/1604.01125", "guide": "papers/1604.01125/guide.json", "extraction": "papers/1604.01125/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1604.01125v3", "pdf": "https://arxiv.org/pdf/1604.01125v3", "html": "https://arxiv.org/html/1604.01125v3" }
1605.07723
3
Data Programming: Creating Large Training Sets, Quickly
[ "Alexander Ratner", "Christopher De Sa", "Sen Wu", "Daniel Selsam", "Christopher Ré" ]
Large labeled training sets are the critical building blocks of supervised learning methods and are key enablers of deep learning techniques. For some applications, creating labeled training sets is the most time-consuming and expensive part of applying machine learning. We therefore propose a paradigm for the programm...
[ "stat.ML", "cs.AI", "cs.LG" ]
2016-05-25T04:14:59
2017-01-08T19:48:53
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Introduces data programming: users write noisy, possibly conflicting labeling functions that label subsets of data; a generative model of the labeling process denoises the resulting training set, and a modified noise-aware discriminative loss trains models such as logistic regression and LSTMs. The paper gives paramete...
[ "weak supervision", "programmatic training-set creation", "label denoising/noisy labels", "labeling-function dependencies", "parameter recovery and theoretical guarantees", "non-expert usability" ]
[ { "id": "weak-supervision-and-noisy-label-learning", "fit": 0.9997621952 } ]
[ { "id": "1711.10160", "score": 6.2571, "via": [ "weak-supervision-and-noisy-label-learning" ], "similarity": 0.8534 }, { "id": "1703.00854", "score": 6.1515, "via": [ "weak-supervision-and-noisy-label-learning" ], "similarity": 0.819 }, { "id": "2501.0...
html
27
{ "dir": "papers/1605.07723", "guide": "papers/1605.07723/guide.json", "extraction": "papers/1605.07723/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1605.07723v3", "pdf": "https://arxiv.org/pdf/1605.07723v3", "html": "https://arxiv.org/html/1605.07723v3" }
1606.05250
3
SQuAD: 100,000+ Questions for Machine Comprehension of Text
[ "Pranav Rajpurkar", "Jian Zhang", "Konstantin Lopyrev", "Percy Liang" ]
We present the Stanford Question Answering Dataset (SQuAD), a new reading comprehension dataset consisting of 100,000+ questions posed by crowdworkers on a set of Wikipedia articles, where the answer to each question is a segment of text from the corresponding reading passage. We analyze the dataset to understand the t...
[ "cs.CL" ]
2016-06-16T16:36:00
2016-10-11T02:42:36
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
SQuAD introduces a large reading comprehension dataset of crowdworker questions posed on Wikipedia articles, with answers as text spans from the corresponding passage. The paper analyzes required reasoning with dependency and constituency trees, builds a logistic regression model with lexicalized and dependency-tree pa...
[ "dataset construction", "dataset analysis", "reasoning types", "model evaluation", "human-machine performance gap", "syntactic divergence" ]
[ { "id": "machine-reading-comprehension", "fit": 0.9992725515 }, { "id": "qa-dataset-benchmark-construction", "fit": 0.9961842159 }, { "id": "nlp-dataset-construction-and-annotation", "fit": 0.9879535427 }, { "id": "language-model-benchmark-evaluation", "fit": 0.9500699397...
[ { "id": "1705.03551", "score": 15.1628, "via": [ "nlp-dataset-construction-and-annotation", "qa-dataset-benchmark-construction", "machine-reading-comprehension" ], "similarity": 0.8729 }, { "id": "2403.12766", "score": 12.6058, "via": [ "qa-dataset-benchma...
pdf
10
{ "dir": "papers/1606.05250", "guide": "papers/1606.05250/guide.json", "extraction": "papers/1606.05250/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1606.05250v3", "pdf": "https://arxiv.org/pdf/1606.05250v3", "html": "https://arxiv.org/html/1606.05250v3" }
1609.05158
2
Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network
[ "Wenzhe Shi", "Jose Caballero", "Ferenc Huszár", "Johannes Totz", "Andrew P. Aitken", "Rob Bishop", "Daniel Rueckert", "Zehan Wang" ]
Recently, several models based on deep neural networks have achieved great success in terms of both reconstruction accuracy and computational performance for single image super-resolution. In these methods, the low resolution (LR) input image is upscaled to the high resolution (HR) space using a single filter, commonly...
[ "cs.CV", "stat.ML" ]
2016-09-16T17:58:14
2016-09-23T17:16:37
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Proposes ESPCN, an efficient convolutional neural network for single-image and video super-resolution. Rather than bicubic-upscaling the low-resolution input before CNN reconstruction, ESPCN extracts feature maps in low-resolution space and uses a learned sub-pixel convolution layer with periodic shuffling to produce h...
[ "efficient CNN architectures for super-resolution", "real-time video super-resolution", "sub-pixel convolution", "low-resolution feature extraction", "learned upscaling", "spatio-temporal video super-resolution" ]
[ { "id": "deep-image-restoration", "fit": 0.9987164097 } ]
[ { "id": "1501.00092", "score": 7.3054, "via": [ "deep-image-restoration" ], "similarity": 0.8964 }, { "id": "2504.16922", "score": 0.8508, "via": [], "similarity": 0.7773 }, { "id": "2608.10636", "score": 0.8072, "via": [], "similarity": 0.7631 }, ...
html
10
{ "dir": "papers/1609.05158", "guide": "papers/1609.05158/guide.json", "extraction": "papers/1609.05158/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1609.05158v2", "pdf": "https://arxiv.org/pdf/1609.05158v2", "html": "https://arxiv.org/html/1609.05158v2" }
1610.01644
4
Understanding intermediate layers using linear classifier probes
[ "Guillaume Alain", "Yoshua Bengio" ]
Neural network models have a reputation for being black boxes. We propose to monitor the features at every layer of a model and measure how suitable they are for classification. We use linear classifiers, which we refer to as "probes", trained entirely independently of the model itself. This helps us better understan...
[ "stat.ML", "cs.LG" ]
2016-10-05T20:59:01
2018-11-22T23:40:00
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper proposes linear classifier probes: linear classifiers trained independently of a neural network to monitor features at every layer and measure how suitable those features are for classification. Probes are non-invasive and do not affect model training. The authors apply the technique to Inception v3 and ResNe...
[ "intermediate-layer analysis", "linear probing", "feature separability", "model diagnostics", "neural network interpretability", "training dynamics" ]
[ { "id": "hidden-state-probing-representation-analysis", "fit": 0.9976115366 }, { "id": "representation-quality-evaluation", "fit": 0.9968412361 }, { "id": "interpretability-of-vision-models", "fit": 0.9856279282 }, { "id": "interpretability-and-probing-evaluation", "fit":...
[ { "id": "2611.11531", "score": 28.2839, "via": [ "representation-quality-evaluation", "interpretability-of-vision-models", "representation-probing-and-comparison", "interpretability-and-probing-evaluation", "hidden-state-probing-representation-analysis", "representati...
html
13
{ "dir": "papers/1610.01644", "guide": "papers/1610.01644/guide.json", "extraction": "papers/1610.01644/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1610.01644v4", "pdf": "https://arxiv.org/pdf/1610.01644v4", "html": "https://arxiv.org/html/1610.01644v4" }
1610.06258
3
Using Fast Weights to Attend to the Recent Past
[ "Jimmy Ba", "Geoffrey Hinton", "Volodymyr Mnih", "Joel Z. Leibo", "Catalin Ionescu" ]
Until recently, research on artificial neural networks was largely restricted to systems with only two types of variable: Neural activities that represent the current or recent input and weights that learn to capture regularities among inputs, outputs and payoffs. There is no good reason for this restriction. Synapses ...
[ "stat.ML", "cs.LG", "cs.NE" ]
2016-10-20T01:03:20
2016-12-05T00:14:01
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper proposes fast weights: neural-network variables that change more slowly than activities but much faster than standard weights. Motivated by multi-timescale synaptic dynamics, fast weights store temporary memories of the recent past and implement attention to the past without storing copies of neural activity ...
[ "memory in neural networks", "attention mechanisms", "associative memory", "recurrent neural networks", "reinforcement learning with memory", "synaptic plasticity" ]
[ { "id": "associative-memory-and-fast-weight-learning", "fit": 0.9963661117 }, { "id": "attention-based-memory-and-computation-architectures", "fit": 0.9697557778 }, { "id": "efficient-sequence-modeling", "fit": 0.935288468 } ]
[ { "id": "2102.11174", "score": 15.7597, "via": [ "associative-memory-and-fast-weight-learning", "attention-based-memory-and-computation-architectures", "efficient-sequence-modeling" ], "similarity": 0.8173 }, { "id": "2508.08435", "score": 10.6563, "via": [ ...
html
10
{ "dir": "papers/1610.06258", "guide": "papers/1610.06258/guide.json", "extraction": "papers/1610.06258/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1610.06258v3", "pdf": "https://arxiv.org/pdf/1610.06258v3", "html": "https://arxiv.org/html/1610.06258v3" }
1702.01417
2
All-but-the-Top: Simple and Effective Postprocessing for Word Representations
[ "Jiaqi Mu", "Suma Bhat", "Pramod Viswanath" ]
Real-valued word representations have transformed NLP applications; popular examples are word2vec and GloVe, recognized for their ability to capture linguistic regularities. In this paper, we demonstrate a {\em very simple}, and yet counter-intuitive, postprocessing technique -- eliminate the common mean vector and a f...
[ "cs.CL", "stat.ML" ]
2017-02-05T15:43:07
2018-03-19T20:28:27
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper proposes a simple, counter-intuitive postprocessing step for off-the-shelf word vectors: subtract the common mean vector and remove a few top dominating principal directions. It motivates this by observed non-zero-mean and non-isotropic geometry, connects the operation to isotropy and a relaxed RAND-WALK/PMI ...
[ "word representation postprocessing", "geometry and isotropy of embeddings", "intrinsic lexical evaluation", "sentence-level and downstream evaluation", "multilingual generalization", "theoretical connection to PMI and RAND-WALK" ]
[ { "id": "text-representation-learning-and-evaluation", "fit": 0.9972766718 }, { "id": "representation-geometry", "fit": 0.9206887552 } ]
[ { "id": "1909.00512", "score": 8.9024, "via": [ "representation-geometry", "text-representation-learning-and-evaluation" ], "similarity": 0.7708 }, { "id": "2603.21437", "score": 8.8685, "via": [ "representation-geometry", "text-representation-learning-and...
html
25
{ "dir": "papers/1702.01417", "guide": "papers/1702.01417/guide.json", "extraction": "papers/1702.01417/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1702.01417v2", "pdf": "https://arxiv.org/pdf/1702.01417v2", "html": "https://arxiv.org/html/1702.01417v2" }
1703.00854
2
Learning the Structure of Generative Models without Labeled Data
[ "Stephen H. Bach", "Bryan He", "Alexander Ratner", "Christopher Ré" ]
Curating labeled training data has become the primary bottleneck in machine learning. Recent frameworks address this bottleneck with generative models to synthesize labels at scale from weak supervision sources. The generative model's dependency structure directly affects the quality of the estimated labels, but select...
[ "cs.LG", "stat.ML" ]
2017-03-02T16:52:09
2017-09-09T21:22:57
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper addresses automatic selection of the dependency structure of generative label models used in weak supervision, so labels can be synthesized without labeled training data. It proposes maximizing an l1-regularized marginal pseudolikelihood over observed labeling-function outputs, using a factor graph that can i...
[ "unsupervised structure learning", "generative models for weak supervision", "l1-regularized pseudolikelihood", "sample complexity of structure recovery", "correlated labeling functions", "data programming" ]
[ { "id": "weak-supervision-and-noisy-label-learning", "fit": 0.9986930945 }, { "id": "probabilistic-generative-models-and-likelihood-based-inference", "fit": 0.9513846423 } ]
[ { "id": "1906.02845", "score": 6.3729, "via": [ "probabilistic-generative-models-and-likelihood-based-inference" ], "similarity": 0.7246 }, { "id": "2107.03006", "score": 6.2774, "via": [ "probabilistic-generative-models-and-likelihood-based-inference" ], "sim...
html
16
{ "dir": "papers/1703.00854", "guide": "papers/1703.00854/guide.json", "extraction": "papers/1703.00854/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1703.00854v2", "pdf": "https://arxiv.org/pdf/1703.00854v2", "html": "https://arxiv.org/html/1703.00854v2" }
1705.03551
2
TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension
[ "Mandar Joshi", "Eunsol Choi", "Daniel S. Weld", "Luke Zettlemoyer" ]
We present TriviaQA, a challenging reading comprehension dataset containing over 650K question-answer-evidence triples. TriviaQA includes 95K question-answer pairs authored by trivia enthusiasts and independently gathered evidence documents, six per question on average, that provide high quality distant supervision for...
[ "cs.CL" ]
2017-05-09T21:35:07
2017-05-13T21:12:37
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
TriviaQA is a large-scale distantly supervised reading comprehension dataset built from trivia-enthusiast-authored questions and independently gathered evidence documents. It is designed to test complex compositional questions, syntactic and lexical variability, and cross-sentence reasoning. The paper compares it with ...
[ "large-scale reading comprehension datasets", "distant supervision for question answering", "challenge datasets", "baseline evaluation", "human-model performance gap", "QA error analysis" ]
[ { "id": "qa-dataset-benchmark-construction", "fit": 0.9989573339 }, { "id": "machine-reading-comprehension", "fit": 0.9982966849 }, { "id": "nlp-dataset-construction-and-annotation", "fit": 0.9882330519 }, { "id": "external-knowledge-question-answering", "fit": 0.95670662...
[ { "id": "1606.05250", "score": 15.1628, "via": [ "nlp-dataset-construction-and-annotation", "qa-dataset-benchmark-construction", "machine-reading-comprehension" ], "similarity": 0.8729 }, { "id": "2403.12766", "score": 10.1076, "via": [ "qa-dataset-benchma...
html
11
{ "dir": "papers/1705.03551", "guide": "papers/1705.03551/guide.json", "extraction": "papers/1705.03551/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1705.03551v2", "pdf": "https://arxiv.org/pdf/1705.03551v2", "html": "https://arxiv.org/html/1705.03551v2" }
1705.08500
2
Selective Classification for Deep Neural Networks
[ "Yonatan Geifman", "Ran El-Yaniv" ]
Selective classification techniques (also known as reject option) have not yet been considered in the context of deep neural networks (DNNs). These techniques can potentially significantly improve DNNs prediction performance by trading-off coverage. In this paper we propose a method to construct a selective classifier ...
[ "cs.LG", "cs.AI" ]
2017-05-23T19:43:56
2017-06-01T14:10:34
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper introduces selective classification, also called reject option, for deep neural networks. Given a trained network, it constructs a selective classifier with a selection function that thresholds a confidence-rate function; at test time the classifier rejects instances to meet a user-specified risk level with h...
[ "selective prediction", "risk-controlled classification", "coverage-risk trade-off", "uncertainty estimation for DNNs", "reject-option mechanisms", "deep neural network reliability" ]
[ { "id": "trustworthy-and-reliable-machine-learning-in-deployment", "fit": 0.9978858301 }, { "id": "predictive-uncertainty-and-calibration-in-deep-neural-networks", "fit": 0.9721568172 } ]
[ { "id": "1706.04599", "score": 11.6271, "via": [ "predictive-uncertainty-and-calibration-in-deep-neural-networks", "trustworthy-and-reliable-machine-learning-in-deployment" ], "similarity": 0.7287 }, { "id": "2011.03395", "score": 5.5408, "via": [ "trustworthy-a...
html
12
{ "dir": "papers/1705.08500", "guide": "papers/1705.08500/guide.json", "extraction": "papers/1705.08500/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1705.08500v2", "pdf": "https://arxiv.org/pdf/1705.08500v2", "html": "https://arxiv.org/html/1705.08500v2" }
1706.04599
2
On Calibration of Modern Neural Networks
[ "Chuan Guo", "Geoff Pleiss", "Yu Sun", "Kilian Q. Weinberger" ]
Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we ob...
[ "cs.LG" ]
2017-06-14T17:33:50
2017-08-03T13:29:46
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper studies confidence calibration: whether predicted probabilities reflect true correctness likelihood. It reports that modern neural networks, unlike older networks such as LeNet, are often poorly calibrated, while accuracy can improve. Through experiments on image and document classification, it examines depth...
[ "neural network calibration", "post-processing calibration", "architecture and training effects on calibration", "overconfidence and uncertainty", "calibration across vision and NLP", "practical calibration recipe" ]
[ { "id": "predictive-uncertainty-and-calibration-in-deep-neural-networks", "fit": 0.9997214262 }, { "id": "distribution-shift-adaptation-calibration-and-ood-robustness", "fit": 0.9937031796 }, { "id": "trustworthy-and-reliable-machine-learning-in-deployment", "fit": 0.9845148206 } ]
[ { "id": "1705.08500", "score": 11.6271, "via": [ "predictive-uncertainty-and-calibration-in-deep-neural-networks", "trustworthy-and-reliable-machine-learning-in-deployment" ], "similarity": 0.7287 }, { "id": "1901.06852", "score": 11.186, "via": [ "distribution-...
html
14
{ "dir": "papers/1706.04599", "guide": "papers/1706.04599/guide.json", "extraction": "papers/1706.04599/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1706.04599v2", "pdf": "https://arxiv.org/pdf/1706.04599v2", "html": "https://arxiv.org/html/1706.04599v2" }
1707.01220
2
DarkRank: Accelerating Deep Metric Learning via Cross Sample Similarities Transfer
[ "Yuntao Chen", "Naiyan Wang", "Zhaoxiang Zhang" ]
We have witnessed rapid evolution of deep neural network architecture design in the past years. These latest progresses greatly facilitate the developments in various areas such as computer vision and natural language processing. However, along with the extraordinary performance, these state-of-the-art models also brin...
[ "cs.CV", "cs.LG", "cs.NE" ]
2017-07-05T05:47:11
2017-12-18T22:44:44
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
DarkRank proposes transferring cross-sample similarities from a deep metric learning teacher to a smaller student as a new form of dark knowledge for model compression and acceleration. It brings learning-to-rank into deep metric learning, formalizing transfer as rank matching with soft and hard variants. The paper eva...
[ "deep metric learning", "knowledge distillation", "learning to rank", "cross-sample similarity transfer", "model compression", "model acceleration" ]
[ { "id": "knowledge-distillation-and-teacher-student-transfer", "fit": 0.9973671368 }, { "id": "neural-network-compression", "fit": 0.9912441886 }, { "id": "embedding-based-representation-learning-and-retrieval", "fit": 0.9139396412 } ]
[ { "id": "2301.12005", "score": 11.8793, "via": [ "knowledge-distillation-and-teacher-student-transfer", "neural-network-compression", "embedding-based-representation-learning-and-retrieval" ], "similarity": 0.787 }, { "id": "1904.05068", "score": 8.9863, "via": ...
html
9
{ "dir": "papers/1707.01220", "guide": "papers/1707.01220/guide.json", "extraction": "papers/1707.01220/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1707.01220v2", "pdf": "https://arxiv.org/pdf/1707.01220v2", "html": "https://arxiv.org/html/1707.01220v2" }
1709.08568
2
The Consciousness Prior
[ "Yoshua Bengio" ]
A new prior is proposed for learning representations of high-level concepts of the kind we manipulate with language. This prior can be combined with other priors in order to help disentangling abstract factors from each other. It is inspired by cognitive neuroscience theories of consciousness, seen as a bottleneck thro...
[ "cs.LG", "cs.AI", "stat.ML" ]
2017-09-25T15:59:11
2019-12-02T22:53:39
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper proposes a consciousness prior for learning representations of high-level concepts. Inspired by cognitive-neuroscience views of consciousness as an attention-selected bottleneck, it treats a low-dimensional conscious state as combining a few concepts selected from a broader unconscious representation and broa...
[ "representation learning", "consciousness-inspired priors", "attention mechanisms", "sparse factor graphs", "disentanglement", "language and symbolic knowledge" ]
[ { "id": "cognitive-and-neuro-inspired-ai-architectures", "fit": 0.9918321316 }, { "id": "representation-learning-and-learned-representation-analysis", "fit": 0.9666868916 } ]
[ { "id": "2401.01364", "score": 9.0582, "via": [ "cognitive-and-neuro-inspired-ai-architectures", "representation-learning-and-learned-representation-analysis" ], "similarity": 0.7515 }, { "id": "2508.08435", "score": 5.5847, "via": [ "cognitive-and-neuro-inspire...
html
7
{ "dir": "papers/1709.08568", "guide": "papers/1709.08568/guide.json", "extraction": "papers/1709.08568/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1709.08568v2", "pdf": "https://arxiv.org/pdf/1709.08568v2", "html": "https://arxiv.org/html/1709.08568v2" }
1711.02257
4
GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks
[ "Zhao Chen", "Vijay Badrinarayanan", "Chen-Yu Lee", "Andrew Rabinovich" ]
Deep multitask networks, in which one neural network produces multiple predictive outputs, can offer better speed and performance than their single-task counterparts but are challenging to train properly. We present a gradient normalization (GradNorm) algorithm that automatically balances training in deep multitask mod...
[ "cs.CV" ]
2017-11-07T02:08:12
2018-06-12T06:45:49
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
GradNorm is an adaptive loss-balancing method for deep multitask networks. It treats task imbalance as an imbalance in backpropagated gradient magnitudes and dynamically tunes task weights so tasks train at more balanced rates. The method uses a gradient loss over task gradient norms, target norms based on average grad...
[ "adaptive loss weighting", "gradient manipulation", "multitask training dynamics", "task balancing", "overfitting reduction", "compute-efficient hyperparameter search" ]
[ { "id": "multi-task-learning-optimization-and-gradient-conflict", "fit": 0.9996605025 }, { "id": "optimization-and-inductive-bias-in-deep-networks", "fit": 0.9898885448 }, { "id": "deep-network-training-efficiency-and-optimization", "fit": 0.9743315705 } ]
[ { "id": "2110.14048", "score": 15.2661, "via": [ "multi-task-learning-optimization-and-gradient-conflict", "optimization-and-inductive-bias-in-deep-networks", "deep-network-training-efficiency-and-optimization" ], "similarity": 0.8722 }, { "id": "2001.06782", "score...
html
12
{ "dir": "papers/1711.02257", "guide": "papers/1711.02257/guide.json", "extraction": "papers/1711.02257/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1711.02257v4", "pdf": "https://arxiv.org/pdf/1711.02257v4", "html": "https://arxiv.org/html/1711.02257v4" }
1711.05851
2
Go for a Walk and Arrive at the Answer: Reasoning Over Paths in Knowledge Bases using Reinforcement Learning
[ "Rajarshi Das", "Shehzaad Dhuliawala", "Manzil Zaheer", "Luke Vilnis", "Ishan Durugkar", "Akshay Krishnamurthy", "Alex Smola", "Andrew McCallum" ]
Knowledge bases (KB), both automatically and manually constructed, are often incomplete --- many valid facts can be inferred from the KB by synthesizing existing information. A popular approach to KB completion is to infer new relations by combinatory reasoning over the information found along other paths connecting a ...
[ "cs.CL", "cs.AI" ]
2017-11-15T23:45:18
2018-12-30T06:56:06
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Proposes MINERVA, a neural reinforcement learning agent for knowledge-base query answering where the relation is known but only one entity is given. MINERVA formulates query answering as a finite-horizon partially observed Markov decision process, encodes interaction history with an LSTM, conditions action selection on...
[ "reasoning over paths in knowledge bases", "reinforcement learning for knowledge graph navigation", "knowledge base query answering", "knowledge base completion", "path interpretability", "relation cardinality and path-type analysis" ]
[ { "id": "knowledge-graph-reasoning-completion-and-memory", "fit": 0.9997586359 }, { "id": "knowledge-graph-question-answering", "fit": 0.9840790758 } ]
[ { "id": "2307.07697", "score": 8.674, "via": [ "knowledge-graph-question-answering", "knowledge-graph-reasoning-completion-and-memory" ], "similarity": 0.8272 }, { "id": "1904.09537", "score": 8.6218, "via": [ "knowledge-graph-question-answering", "knowled...
html
18
{ "dir": "papers/1711.05851", "guide": "papers/1711.05851/guide.json", "extraction": "papers/1711.05851/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1711.05851v2", "pdf": "https://arxiv.org/pdf/1711.05851v2", "html": "https://arxiv.org/html/1711.05851v2" }
1711.10160
1
Snorkel: Rapid Training Data Creation with Weak Supervision
[ "Alexander Ratner", "Stephen H. Bach", "Henry Ehrenberg", "Jason Fries", "Sen Wu", "Christopher Ré" ]
Labeling training data is increasingly the largest bottleneck in deploying machine learning systems. We present Snorkel, a first-of-its-kind system that enables users to train state-of-the-art models without hand labeling any training data. Instead, users write labeling functions that express arbitrary heuristics, whic...
[ "cs.LG", "stat.ML" ]
2017-11-28T07:48:05
2017-11-28T07:48:05
10.14778/3157794.3157797
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Snorkel is a weak-supervision system for creating training data without hand labeling. Users write labeling functions expressing heuristics with unknown accuracies and correlations; data programming denoises their outputs without ground truth to produce probabilistic training labels for a discriminative model. The pape...
[ "weak supervision", "data programming", "labeling-function interfaces", "source accuracy and correlation modeling", "modeling tradeoffs", "user study and deployments" ]
[ { "id": "weak-supervision-and-noisy-label-learning", "fit": 0.9996367079 }, { "id": "human-in-the-loop-and-human-centered-ai", "fit": 0.9221112981 } ]
[ { "id": "1605.07723", "score": 6.2571, "via": [ "weak-supervision-and-noisy-label-learning" ], "similarity": 0.8534 }, { "id": "1703.00854", "score": 6.1199, "via": [ "weak-supervision-and-noisy-label-learning" ], "similarity": 0.8087 }, { "id": "2109....
html
17
{ "dir": "papers/1711.10160", "guide": "papers/1711.10160/guide.json", "extraction": "papers/1711.10160/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1711.10160v1", "pdf": "https://arxiv.org/pdf/1711.10160v1", "html": "https://arxiv.org/html/1711.10160v1" }
1802.03916
3
Detecting and Correcting for Label Shift with Black Box Predictors
[ "Zachary C. Lipton", "Yu-Xiang Wang", "Alex Smola" ]
Faced with distribution shift between training and test set, we wish to detect and quantify the shift, and to correct our classifiers without test set labels. Motivated by medical diagnosis, where diseases (targets) cause symptoms (observations), we focus on label shift, where the label marginal $p(y)$ changes but the ...
[ "cs.LG", "cs.AI", "cs.NE", "stat.ML" ]
2018-02-12T07:16:03
2018-07-26T12:51:37
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper addresses label shift: the label marginal p(y) changes between training and test data while p(x|y) stays fixed. It proposes Black Box Shift Estimation (BBSE), which uses an arbitrary black-box predictor and its expected confusion matrix to estimate the test label distribution without test labels. BBSE support...
[ "label shift", "distribution shift", "black-box prediction", "classifier correction", "statistical shift detection", "confusion matrix invertibility" ]
[ { "id": "post-hoc-classifier-correction-under-distribution-shift", "fit": 0.9991933823 }, { "id": "distribution-shift-adaptation-calibration-and-ood-robustness", "fit": 0.9990978547 }, { "id": "trustworthy-and-reliable-machine-learning-in-deployment", "fit": 0.9859160333 } ]
[ { "id": "1901.06852", "score": 17.5218, "via": [ "post-hoc-classifier-correction-under-distribution-shift", "distribution-shift-adaptation-calibration-and-ood-robustness", "trustworthy-and-reliable-machine-learning-in-deployment" ], "similarity": 0.8913 }, { "id": "1906...
html
11
{ "dir": "papers/1802.03916", "guide": "papers/1802.03916/guide.json", "extraction": "papers/1802.03916/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1802.03916v3", "pdf": "https://arxiv.org/pdf/1802.03916v3", "html": "https://arxiv.org/html/1802.03916v3" }
1805.12471
3
Neural Network Acceptability Judgments
[ "Alex Warstadt", "Amanpreet Singh", "Samuel R. Bowman" ]
This paper investigates the ability of artificial neural networks to judge the grammatical acceptability of a sentence, with the goal of testing their linguistic competence. We introduce the Corpus of Linguistic Acceptability (CoLA), a set of 10,657 English sentences labeled as grammatical or ungrammatical from publish...
[ "cs.CL" ]
2018-05-31T13:52:06
2019-10-01T18:41:05
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper introduces CoLA, a large corpus of English sentences labeled grammatical or ungrammatical from published linguistics literature, and trains recurrent neural network baselines for acceptability classification. It compares them with unsupervised language models and human judgments, and analyzes errors across gr...
[ "neural acceptability judgments", "linguistic competence of neural networks", "grammaticality classification", "transfer and pretraining for acceptability", "human-model performance gap", "grammatical phenomena analysis" ]
[ { "id": "linguistic-cognitive-lm-evaluation", "fit": 0.9853037504 }, { "id": "language-model-benchmark-evaluation", "fit": 0.9183238012 } ]
[ { "id": "2412.10271", "score": 7.2778, "via": [ "linguistic-cognitive-lm-evaluation", "language-model-benchmark-evaluation" ], "similarity": 0.7558 }, { "id": "2310.15421", "score": 7.1912, "via": [ "linguistic-cognitive-lm-evaluation", "language-model-ben...
html
17
{ "dir": "papers/1805.12471", "guide": "papers/1805.12471/guide.json", "extraction": "papers/1805.12471/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1805.12471v3", "pdf": "https://arxiv.org/pdf/1805.12471v3", "html": "https://arxiv.org/html/1805.12471v3" }
1808.09121
3
WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations
[ "Mohammad Taher Pilehvar", "Jose Camacho-Collados" ]
By design, word embeddings are unable to model the dynamic nature of words' semantics, i.e., the property of words to correspond to potentially different meanings. To address this limitation, dozens of specialized meaning representation techniques such as sense or contextualized embeddings have been proposed. However, ...
[ "cs.CL" ]
2018-08-28T05:16:35
2019-04-27T09:31:59
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper argues that static word embeddings cannot represent context-dependent word meaning and that existing evaluation resources, especially Stanford Contextual Word Similarity, are inadequate. It introduces WiC, a large-scale expert-annotated Word-in-Context benchmark framed as binary classification: given a target...
[ "" ]
[ { "id": "text-representation-learning-and-evaluation", "fit": 0.9934563656 }, { "id": "benchmark-construction-and-annotation", "fit": 0.9933632834 }, { "id": "evaluation-metrics-and-benchmark-design", "fit": 0.9373136442 }, { "id": "language-model-benchmark-evaluation", "...
[ { "id": "2412.11912", "score": 10.3479, "via": [ "benchmark-construction-and-annotation", "evaluation-metrics-and-benchmark-design", "language-model-benchmark-evaluation" ], "similarity": 0.773 }, { "id": "2405.07940", "score": 10.3051, "via": [ "benchmark...
html
7
{ "dir": "papers/1808.09121", "guide": "papers/1808.09121/guide.json", "extraction": "papers/1808.09121/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1808.09121v3", "pdf": "https://arxiv.org/pdf/1808.09121v3", "html": "https://arxiv.org/html/1808.09121v3" }
1809.00782
1
Open Domain Question Answering Using Early Fusion of Knowledge Bases and Text
[ "Haitian Sun", "Bhuwan Dhingra", "Manzil Zaheer", "Kathryn Mazaitis", "Ruslan Salakhutdinov", "William W. Cohen" ]
Open Domain Question Answering (QA) is evolving from complex pipelined systems to end-to-end deep neural networks. Specialized neural models have been developed for extracting answers from either text alone or Knowledge Bases (KBs) alone. In this paper we look at a more practical setting, namely QA over the combination...
[ "cs.CL", "cs.LG" ]
2018-09-04T03:15:56
2018-09-04T03:15:56
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper addresses open-domain question answering over an incomplete knowledge base combined with entity-linked text. It proposes GRAFT-Net, a graph-representation model that performs early fusion by extracting answers from a question-specific heterogeneous subgraph containing KB entities and relations plus text. It c...
[ "combining knowledge bases and text", "early fusion versus late fusion", "incomplete knowledge bases", "graph-based answer extraction", "question-specific subgraphs", "heterogeneous graph propagation" ]
[ { "id": "external-knowledge-question-answering", "fit": 0.9920063134 }, { "id": "knowledge-graph-question-answering", "fit": 0.9916104831 }, { "id": "knowledge-graph-reasoning-completion-and-memory", "fit": 0.9497324163 } ]
[ { "id": "1904.09537", "score": 13.3307, "via": [ "external-knowledge-question-answering", "knowledge-graph-question-answering", "knowledge-graph-reasoning-completion-and-memory" ], "similarity": 0.8849 }, { "id": "2106.01515", "score": 13.1034, "via": [ "e...
html
12
{ "dir": "papers/1809.00782", "guide": "papers/1809.00782/guide.json", "extraction": "papers/1809.00782/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1809.00782v1", "pdf": "https://arxiv.org/pdf/1809.00782v1", "html": "https://arxiv.org/html/1809.00782v1" }
1810.11118
2
A Large-Scale Corpus for Conversation Disentanglement
[ "Jonathan K. Kummerfeld", "Sai R. Gouravajhala", "Joseph Peper", "Vignesh Athreya", "Chulaka Gunasekara", "Jatin Ganhotra", "Siva Sankalp Patel", "Lazaros Polymenakos", "Walter S. Lasecki" ]
Disentangling conversations mixed together in a single stream of messages is a difficult task, made harder by the lack of large manually annotated datasets. We created a new dataset of 77,563 messages manually annotated with reply-structure graphs that both disentangle conversations and define internal conversation str...
[ "cs.CL" ]
2018-10-25T21:44:14
2019-07-18T18:14:53
10.18653/v1/P19-1374
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper introduces a large manually annotated IRC corpus for conversation disentanglement. Messages are labeled with reply-structure graphs that both separate entangled conversations and define internal reply structure, with adjudication of annotation disagreements and context included. The authors compare the corpus...
[ "conversation disentanglement", "corpus construction and annotation quality", "reply structure", "context and data diversity", "prior corpus re-evaluation", "dialogue modeling" ]
[ { "id": "nlp-dataset-construction-and-annotation", "fit": 0.9895816411 } ]
[ { "id": "1705.03551", "score": 5.6337, "via": [ "nlp-dataset-construction-and-annotation" ], "similarity": 0.7599 }, { "id": "2311.09182", "score": 5.6286, "via": [ "nlp-dataset-construction-and-annotation" ], "similarity": 0.7582 }, { "id": "2211.0257...
html
11
{ "dir": "papers/1810.11118", "guide": "papers/1810.11118/guide.json", "extraction": "papers/1810.11118/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1810.11118v2", "pdf": "https://arxiv.org/pdf/1810.11118v2", "html": "https://arxiv.org/html/1810.11118v2" }
1811.03115
1
Blockwise Parallel Decoding for Deep Autoregressive Models
[ "Mitchell Stern", "Noam Shazeer", "Jakob Uszkoreit" ]
Deep autoregressive sequence-to-sequence models have demonstrated impressive performance across a wide variety of tasks in recent years. While common architecture classes such as recurrent, convolutional, and self-attention networks make different trade-offs between the amount of computation needed per layer and the le...
[ "cs.LG", "cs.CL", "stat.ML" ]
2018-11-07T19:09:40
2018-11-07T19:09:40
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper proposes blockwise parallel decoding for deep autoregressive sequence-to-sequence models. It predicts multiple future time steps in parallel and backs off to the longest prefix validated by a scoring model, aiming to preserve greedy decoding output while reducing sequential iterations. It introduces a combine...
[ "efficient inference", "parallel decoding", "autoregressive generation", "quality-speed trade-off", "knowledge distillation", "Transformer models" ]
[ { "id": "low-latency-text-generation", "fit": 0.9980061114 }, { "id": "efficient-transformer-inference", "fit": 0.9917950811 }, { "id": "language-model-decoding", "fit": 0.9783078937 }, { "id": "parallel-and-speculative-decoding", "fit": 0.9765342074 }, { "id": "e...
[ { "id": "2211.17192", "score": 18.6789, "via": [ "parallel-and-speculative-decoding", "language-model-decoding", "low-latency-text-generation", "efficient-transformer-inference", "efficient-and-compressed-language-model-lifecycle" ], "similarity": 0.8855 }, { ...
html
10
{ "dir": "papers/1811.03115", "guide": "papers/1811.03115/guide.json", "extraction": "papers/1811.03115/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1811.03115v1", "pdf": "https://arxiv.org/pdf/1811.03115v1", "html": "https://arxiv.org/html/1811.03115v1" }
1811.06965
5
GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism
[ "Yanping Huang", "Youlong Cheng", "Ankur Bapna", "Orhan Firat", "Mia Xu Chen", "Dehao Chen", "HyoukJoong Lee", "Jiquan Ngiam", "Quoc V. Le", "Yonghui Wu", "Zhifeng Chen" ]
Scaling up deep neural network capacity has been known as an effective approach to improving model quality for several different machine learning tasks. In many cases, increasing model capacity beyond the memory limit of a single accelerator has required developing special algorithms or infrastructure. These solutions ...
[ "cs.CV" ]
2018-11-16T18:43:28
2019-07-25T21:42:58
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
GPipe is a pipeline parallelism library for training giant neural networks expressible as a sequence of layers. It partitions layer sub-sequences across accelerators and uses batch-splitting pipelining with synchronous gradient updates. The paper demonstrates it on ImageNet-2012 image classification with a large Amoeba...
[ "efficient training of giant neural networks", "pipeline and model parallelism", "large-scale image classification", "multilingual machine translation", "depth-width trade-off", "large-batch training" ]
[ { "id": "distributed-deep-learning-systems", "fit": 0.9993816983 }, { "id": "deep-network-training-efficiency-and-optimization", "fit": 0.9971035771 }, { "id": "gpu-utilization-throughput-optimization", "fit": 0.9739924884 } ]
[ { "id": "1909.08053", "score": 12.3134, "via": [ "distributed-deep-learning-systems", "deep-network-training-efficiency-and-optimization", "gpu-utilization-throughput-optimization" ], "similarity": 0.857 }, { "id": "2104.04473", "score": 12.306, "via": [ "...
html
11
{ "dir": "papers/1811.06965", "guide": "papers/1811.06965/guide.json", "extraction": "papers/1811.06965/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1811.06965v5", "pdf": "https://arxiv.org/pdf/1811.06965v5", "html": "https://arxiv.org/html/1811.06965v5" }
1901.06852
5
Maximum Likelihood with Bias-Corrected Calibration is Hard-To-Beat at Label Shift Adaptation
[ "Amr Alexandari", "Anshul Kundaje", "Avanti Shrikumar" ]
Label shift refers to the phenomenon where the prior class probability p(y) changes between the training and test distributions, while the conditional probability p(x|y) stays fixed. Label shift arises in settings like medical diagnosis, where a classifier trained to predict disease given symptoms must be adapted to sc...
[ "cs.LG", "stat.ML" ]
2019-01-21T10:05:24
2020-06-27T03:40:09
null
http://creativecommons.org/publicdomain/zero/1.0/
The paper studies label shift adaptation: p(y) changes between source and target while p(x|y) stays fixed. It revisits Saerens et al.'s maximum-likelihood/EM correction, whose usual assumption of calibrated p(y|x) is violated by modern neural networks. The authors combine maximum likelihood with bias-corrected calibrat...
[ "label shift adaptation", "calibration under distribution shift", "maximum-likelihood baselines", "source-prior estimation", "benchmarking BBSL/RLLS", "robustness to miscalibration" ]
[ { "id": "distribution-shift-adaptation-calibration-and-ood-robustness", "fit": 0.9997133622 }, { "id": "post-hoc-classifier-correction-under-distribution-shift", "fit": 0.9989918898 }, { "id": "trustworthy-and-reliable-machine-learning-in-deployment", "fit": 0.976352511 } ]
[ { "id": "1802.03916", "score": 17.5218, "via": [ "post-hoc-classifier-correction-under-distribution-shift", "distribution-shift-adaptation-calibration-and-ood-robustness", "trustworthy-and-reliable-machine-learning-in-deployment" ], "similarity": 0.8913 }, { "id": "1706...
html
24
{ "dir": "papers/1901.06852", "pdf": "papers/1901.06852/paper.pdf", "text": "papers/1901.06852/text.md", "pages": "papers/1901.06852/pages", "guide": "papers/1901.06852/guide.json", "extraction": "papers/1901.06852/extraction.jsonl" }
included
{ "abs": "https://arxiv.org/abs/1901.06852v5", "pdf": "https://arxiv.org/pdf/1901.06852v5", "html": "https://arxiv.org/html/1901.06852v5" }
1902.03393
2
Improved Knowledge Distillation via Teacher Assistant
[ "Seyed-Iman Mirzadeh", "Mehrdad Farajtabar", "Ang Li", "Nir Levine", "Akihiro Matsukawa", "Hassan Ghasemzadeh" ]
Despite the fact that deep neural networks are powerful models and achieve appealing results on many tasks, they are too large to be deployed on edge devices like smartphones or embedded sensor nodes. There have been efforts to compress these networks, and a popular method is knowledge distillation, where a large (teac...
[ "cs.LG", "cs.AI", "stat.ML" ]
2019-02-09T09:06:01
2019-12-17T01:11:13
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper argues that knowledge distillation degrades when the capacity gap between a fixed small student and a large teacher is too large. It proposes teacher assistant knowledge distillation (TAKD), inserting an intermediate-sized teacher assistant—or a multi-step chain of assistants—between teacher and student. It s...
[ "teacher-student capacity gap", "teacher assistant size and selection", "multi-step distillation paths", "loss landscape and generalization", "theoretical analysis of knowledge distillation", "comparison with distillation methods" ]
[ { "id": "knowledge-distillation-and-teacher-student-transfer", "fit": 0.9998579021 }, { "id": "neural-network-compression", "fit": 0.9981394712 } ]
[ { "id": "1904.05068", "score": 9.0864, "via": [ "knowledge-distillation-and-teacher-student-transfer", "neural-network-compression" ], "similarity": 0.8465 }, { "id": "2301.12005", "score": 9.0597, "via": [ "knowledge-distillation-and-teacher-student-transfer", ...
html
11
{ "dir": "papers/1902.03393", "guide": "papers/1902.03393/guide.json", "extraction": "papers/1902.03393/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1902.03393v2", "pdf": "https://arxiv.org/pdf/1902.03393v2", "html": "https://arxiv.org/html/1902.03393v2" }
1904.02342
3
Text Generation from Knowledge Graphs with Graph Transformers
[ "Rik Koncel-Kedziorski", "Dhanush Bekal", "Yi Luan", "Mirella Lapata", "Hannaneh Hajishirzi" ]
Generating texts which express complex ideas spanning multiple sentences requires a structured representation of their content (document plan), but these representations are prohibitively expensive to manually produce. In this work, we address the problem of generating coherent multi-sentence texts from the output of a...
[ "cs.CL" ]
2019-04-04T04:33:15
2022-03-24T22:29:12
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper addresses end-to-end generation of coherent multi-sentence scientific text from automatically extracted knowledge graphs. It introduces GraphWriter, an encoder-decoder system whose graph transforming encoder uses graph structure without linearization or hierarchical constraints. The work constructs AGENDA fro...
[ "graph-to-text generation", "scientific text generation", "graph neural networks and attention", "encoder-decoder models", "document planning", "knowledge graph construction" ]
[ { "id": "structured-data-to-text-generation", "fit": 0.9989762677 }, { "id": "structured-data-serialization-for-language-model-generation", "fit": 0.9789511488 } ]
[ { "id": "2106.10502", "score": 9.8781, "via": [ "structured-data-to-text-generation", "structured-data-serialization-for-language-model-generation" ], "similarity": 0.8748 }, { "id": "2007.08426", "score": 9.8078, "via": [ "structured-data-to-text-generation", ...
html
10
{ "dir": "papers/1904.02342", "guide": "papers/1904.02342/guide.json", "extraction": "papers/1904.02342/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1904.02342v3", "pdf": "https://arxiv.org/pdf/1904.02342v3", "html": "https://arxiv.org/html/1904.02342v3" }
1904.03396
2
Step-by-Step: Separating Planning from Realization in Neural Data-to-Text Generation
[ "Amit Moryossef", "Yoav Goldberg", "Ido Dagan" ]
Data-to-text generation can be conceptually divided into two parts: ordering and structuring the information (planning), and generating fluent language describing the information (realization). Modern neural generation systems conflate these two steps into a single end-to-end differentiable system. We propose to split ...
[ "cs.CL", "cs.AI" ]
2019-04-06T09:25:32
2019-05-01T20:58:28
null
http://creativecommons.org/licenses/by/4.0/
The paper proposes decoupling data-to-text generation into a symbolic text-planning stage that orders and structures input facts faithfully, followed by a neural realization stage that verbalizes a selected plan fluently. It describes matching reference texts to text plans for training a plan-to-text generator and sele...
[ "separating planning from realization", "faithful data-to-text generation", "diverse and controllable realization", "plan ranking and selection", "semantic faithfulness versus automatic metrics", "discourse and referring expressions" ]
[ { "id": "plan-and-outline-conditioned-text-generation", "fit": 0.9890120619 }, { "id": "structured-data-to-text-generation", "fit": 0.9863332821 }, { "id": "constrained-neuro-symbolic-text-generation", "fit": 0.9265460579 } ]
[ { "id": "1909.03480", "score": 11.667, "via": [ "constrained-neuro-symbolic-text-generation", "plan-and-outline-conditioned-text-generation" ], "similarity": 0.8236 }, { "id": "2608.05169", "score": 6.5134, "via": [ "constrained-neuro-symbolic-text-generation" ...
html
21
{ "dir": "papers/1904.03396", "pdf": "papers/1904.03396/paper.pdf", "text": "papers/1904.03396/text.md", "pages": "papers/1904.03396/pages", "guide": "papers/1904.03396/guide.json", "extraction": "papers/1904.03396/extraction.jsonl" }
included
{ "abs": "https://arxiv.org/abs/1904.03396v2", "pdf": "https://arxiv.org/pdf/1904.03396v2", "html": "https://arxiv.org/html/1904.03396v2" }
1904.05068
2
Relational Knowledge Distillation
[ "Wonpyo Park", "Dongju Kim", "Yan Lu", "Minsu Cho" ]
Knowledge distillation aims at transferring knowledge acquired in one model (a teacher) to another model (a student) that is typically smaller. Previous approaches can be expressed as a form of training the student to mimic output activations of individual data examples represented by the teacher. We introduce a novel ...
[ "cs.CV", "cs.LG" ]
2019-04-10T08:52:14
2019-05-01T09:36:42
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Relational knowledge distillation (RKD) trains a student to match mutual relations among data examples rather than only mimicking a teacher's per-example output activations. The paper proposes distance-wise and angle-wise relational losses, frames RKD as a generalization of conventional knowledge distillation that can ...
[ "knowledge distillation", "relational knowledge transfer", "metric learning", "self-distillation", "few-shot learning", "domain adaptation and generalization" ]
[ { "id": "knowledge-distillation-and-teacher-student-transfer", "fit": 0.9998069929 }, { "id": "neural-network-compression", "fit": 0.9772319557 } ]
[ { "id": "1902.03393", "score": 9.0864, "via": [ "knowledge-distillation-and-teacher-student-transfer", "neural-network-compression" ], "similarity": 0.8465 }, { "id": "2301.12005", "score": 9.0513, "via": [ "knowledge-distillation-and-teacher-student-transfer", ...
html
10
{ "dir": "papers/1904.05068", "guide": "papers/1904.05068/guide.json", "extraction": "papers/1904.05068/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1904.05068v2", "pdf": "https://arxiv.org/pdf/1904.05068v2", "html": "https://arxiv.org/html/1904.05068v2" }
1904.09537
1
PullNet: Open Domain Question Answering with Iterative Retrieval on Knowledge Bases and Text
[ "Haitian Sun", "Tania Bedrax-Weiss", "William W. Cohen" ]
We consider open-domain queston answering (QA) where answers are drawn from either a corpus, a knowledge base (KB), or a combination of both of these. We focus on a setting in which a corpus is supplemented with a large but incomplete KB, and on questions that require non-trivial (e.g., ``multi-hop'') reasoning. We des...
[ "cs.CL", "cs.LG" ]
2019-04-21T03:49:09
2019-04-21T03:49:09
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
PullNet addresses open-domain question answering where answers come from a corpus, a knowledge base (KB), or both, especially when the KB is large but incomplete and questions require multi-hop reasoning. It learns what to retrieve and iteratively builds a question-specific heterogeneous subgraph: a graph CNN selects n...
[ "hybrid KB-text question answering", "incomplete knowledge bases", "iterative retrieval", "graph-based reasoning", "weakly supervised QA", "multi-hop QA" ]
[ { "id": "external-knowledge-question-answering", "fit": 0.9960654285 }, { "id": "knowledge-graph-question-answering", "fit": 0.9950611079 }, { "id": "multi-hop-question-answering", "fit": 0.9911103795 }, { "id": "knowledge-graph-reasoning-completion-and-memory", "fit": 0....
[ { "id": "2212.10509", "score": 21.7054, "via": [ "multi-hop-question-answering", "external-knowledge-question-answering", "multi-hop-compositional-reasoning", "iterative-stateful-long-context", "retrieval-augmented-reasoning" ], "similarity": 0.8433 }, { "id...
html
11
{ "dir": "papers/1904.09537", "guide": "papers/1904.09537/guide.json", "extraction": "papers/1904.09537/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1904.09537v1", "pdf": "https://arxiv.org/pdf/1904.09537v1", "html": "https://arxiv.org/html/1904.09537v1" }
1904.09751
2
The Curious Case of Neural Text Degeneration
[ "Ari Holtzman", "Jan Buys", "Li Du", "Maxwell Forbes", "Yejin Choi" ]
Despite considerable advancements with deep neural language models, the enigma of neural text degeneration persists when these models are tested as text generators. The counter-intuitive empirical observation is that even though the use of likelihood as training objective leads to high quality models for a broad range ...
[ "cs.CL" ]
2019-04-22T07:17:18
2020-02-14T21:56:30
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper studies why neural language models degenerate as text generators despite likelihood-based training. It compares human and machine text distributions, shows that decoding strategy alone can strongly change generation quality, and identifies the unreliable tail as a source of incoherence under pure sampling and...
[ "neural text degeneration", "decoding strategies for text generation", "human vs machine text distributions", "diversity-fluency tradeoff", "probability distribution truncation", "evaluation of generated text" ]
[ { "id": "language-model-generation-degeneration", "fit": 0.9996768862 }, { "id": "language-model-decoding", "fit": 0.9996173652 }, { "id": "llm-output-diversity-and-mode-collapse", "fit": 0.997181199 }, { "id": "inference-time-language-model-control", "fit": 0.9317411254 ...
[ { "id": "2510.15061", "score": 16.4779, "via": [ "language-model-generation-degeneration", "llm-output-diversity-and-mode-collapse", "language-model-decoding", "inference-time-language-model-control" ], "similarity": 0.7446 }, { "id": "2510.01171", "score": 12...
html
16
{ "dir": "papers/1904.09751", "guide": "papers/1904.09751/guide.json", "extraction": "papers/1904.09751/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1904.09751v2", "pdf": "https://arxiv.org/pdf/1904.09751v2", "html": "https://arxiv.org/html/1904.09751v2" }
1906.01787
1
Learning Deep Transformer Models for Machine Translation
[ "Qiang Wang", "Bei Li", "Tong Xiao", "Jingbo Zhu", "Changliang Li", "Derek F. Wong", "Lidia S. Chao" ]
Transformer is the state-of-the-art model in recent machine translation evaluations. Two strands of research are promising to improve models of this kind: the first uses wide networks (a.k.a. Transformer-Big) and has been the de facto standard for the development of the Transformer system, and the other uses deeper lan...
[ "cs.CL", "cs.LG" ]
2019-06-05T02:24:12
2019-06-05T02:24:12
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper asks whether a truly deep Transformer can outperform the wide Transformer-Big for machine translation. It proposes proper layer normalization and dynamic linear combination of previous layers (DLCL), trains deep encoder systems, and evaluates on WMT'16 English-German, NIST OpenMT'12 Chinese-English, and WMT'1...
[ "deep Transformer models", "layer normalization in deep networks", "model depth versus width", "cross-layer information flow", "training deep neural networks", "machine translation efficiency" ]
[ { "id": "transformer-architecture-design", "fit": 0.9946693299 }, { "id": "training-stability-normalization", "fit": 0.9525521923 } ]
[ { "id": "2502.05003", "score": 6.7753, "via": [ "training-stability-normalization" ], "similarity": 0.7237 }, { "id": "2510.14614", "score": 5.8705, "via": [ "transformer-architecture-design" ], "similarity": 0.7868 }, { "id": "2207.06881", "score"...
html
13
{ "dir": "papers/1906.01787", "guide": "papers/1906.01787/guide.json", "extraction": "papers/1906.01787/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1906.01787v1", "pdf": "https://arxiv.org/pdf/1906.01787v1", "html": "https://arxiv.org/html/1906.01787v1" }
1906.02845
2
Likelihood Ratios for Out-of-Distribution Detection
[ "Jie Ren", "Peter J. Liu", "Emily Fertig", "Jasper Snoek", "Ryan Poplin", "Mark A. DePristo", "Joshua V. Dillon", "Balaji Lakshminarayanan" ]
Discriminative neural networks offer little or no performance guarantees when deployed on data not generated by the same process as the training distribution. On such out-of-distribution (OOD) inputs, the prediction may not only be erroneous, but confidently so, limiting the safe deployment of classifiers in real-world...
[ "stat.ML", "cs.LG" ]
2019-06-07T00:01:42
2019-12-05T19:59:35
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper studies out-of-distribution (OOD) detection for neural classifiers, motivated by safety-critical bacteria identification from genomic sequences. It argues that likelihood scores from deep generative models are confounded by population-level background statistics, and proposes a likelihood ratio that compares ...
[ "Likelihood-based OOD detection", "Background-statistic confounding", "Genomic benchmark design", "Safe deployment under distributional shift", "Generative model evaluation", "Image OOD benchmarks" ]
[ { "id": "distribution-shift-adaptation-calibration-and-ood-robustness", "fit": 0.9994561839 }, { "id": "trustworthy-and-reliable-machine-learning-in-deployment", "fit": 0.994086504 }, { "id": "probabilistic-generative-models-and-likelihood-based-inference", "fit": 0.9860521472 }, ...
[ { "id": "1802.03916", "score": 11.068, "via": [ "distribution-shift-adaptation-calibration-and-ood-robustness", "trustworthy-and-reliable-machine-learning-in-deployment" ], "similarity": 0.7724 }, { "id": "1901.06852", "score": 11.0354, "via": [ "distribution-sh...
html
21
{ "dir": "papers/1906.02845", "guide": "papers/1906.02845/guide.json", "extraction": "papers/1906.02845/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1906.02845v2", "pdf": "https://arxiv.org/pdf/1906.02845v2", "html": "https://arxiv.org/html/1906.02845v2" }
1907.10902
1
Optuna: A Next-generation Hyperparameter Optimization Framework
[ "Takuya Akiba", "Shotaro Sano", "Toshihiko Yanase", "Takeru Ohta", "Masanori Koyama" ]
The purpose of this study is to introduce new design-criteria for next-generation hyperparameter optimization software. The criteria we propose include (1) define-by-run API that allows users to construct the parameter search space dynamically, (2) efficient implementation of both searching and pruning strategies, and ...
[ "cs.LG", "stat.ML" ]
2019-07-25T08:55:22
2019-07-25T08:55:22
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Optuna proposes design criteria for next-generation hyperparameter optimization software and presents an open-source framework implementing them. Its define-by-run API lets users construct search spaces dynamically inside an objective function via a trial object, contrasting with define-and-run APIs. Optuna combines sa...
[ "next-generation hyperparameter optimization software", "define-by-run programming", "pruning strategies", "distributed optimization", "interactive visualization", "real-world parameter tuning" ]
[ { "id": "hyperparameter-optimization-and-training-sensitivity", "fit": 0.9960999904 }, { "id": "automated-machine-learning", "fit": 0.9919501331 }, { "id": "resource-efficient-machine-learning", "fit": 0.9269173076 } ]
[ { "id": "2511.15074", "score": 6.6228, "via": [ "automated-machine-learning" ], "similarity": 0.6739 }, { "id": "2610.32142", "score": 6.548, "via": [ "hyperparameter-optimization-and-training-sensitivity" ], "similarity": 0.6496 }, { "id": "2105.05720...
html
10
{ "dir": "papers/1907.10902", "guide": "papers/1907.10902/guide.json", "extraction": "papers/1907.10902/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1907.10902v1", "pdf": "https://arxiv.org/pdf/1907.10902v1", "html": "https://arxiv.org/html/1907.10902v1" }
1909.00512
1
How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings
[ "Kawin Ethayarajh" ]
Replacing static word embeddings with contextualized word representations has yielded significant improvements on many NLP tasks. However, just how contextual are the contextualized representations produced by models such as ELMo and BERT? Are there infinitely many context-specific representations for each word, or are...
[ "cs.CL" ]
2019-09-02T01:51:46
2019-09-02T01:51:46
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper asks how contextual contextualized word representations from ELMo, BERT, and GPT-2 really are, and whether each word receives infinitely many context-specific vectors or only a finite set of word-sense representations. It compares representation geometry layer by layer using cosine similarity, anisotropy-adju...
[ "geometry of contextualized representations", "layer-wise context-specificity", "anisotropy in contextualized embeddings", "static versus contextual embeddings", "word-sense versus context-driven variation", "intrinsic evaluation of derived static embeddings" ]
[ { "id": "representation-geometry", "fit": 0.9917685976 }, { "id": "text-representation-learning-and-evaluation", "fit": 0.9871010606 }, { "id": "neural-representation-geometry-and-similarity-analysis", "fit": 0.9862133778 }, { "id": "representation-quality-evaluation", "f...
[ { "id": "2302.00294", "score": 18.8078, "via": [ "representation-quality-evaluation", "neural-representation-geometry-and-similarity-analysis", "representation-geometry", "linear-and-geometric-representation-structure" ], "similarity": 0.7581 }, { "id": "2605.2303...
html
11
{ "dir": "papers/1909.00512", "guide": "papers/1909.00512/guide.json", "extraction": "papers/1909.00512/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1909.00512v1", "pdf": "https://arxiv.org/pdf/1909.00512v1", "html": "https://arxiv.org/html/1909.00512v1" }
1909.03368
1
Designing and Interpreting Probes with Control Tasks
[ "John Hewitt", "Percy Liang" ]
Probes, supervised models trained to predict properties (like parts-of-speech) from representations (like ELMo), have achieved high accuracy on a range of linguistic tasks. But does this mean that the representations encode linguistic structure or just that the probe has learned the linguistic task? In this paper, we p...
[ "cs.CL" ]
2019-09-08T02:04:32
2019-09-08T02:04:32
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper asks whether high probe accuracy reflects linguistic structure in representations or the probe's own learning. It proposes control tasks that map word types to random outputs and defines probe selectivity as high linguistic-task accuracy with low control-task accuracy. Control tasks are built for English part...
[ "probing representations", "control tasks", "probe selectivity", "regularization for probes", "ELMo layer analysis", "linguistic structure in representations" ]
[ { "id": "linguistic-structure-probing-neural-representations", "fit": 0.9997666244 }, { "id": "representation-probing-and-comparison", "fit": 0.9996309084 }, { "id": "interpretability-and-probing-evaluation", "fit": 0.999439794 }, { "id": "hidden-state-probing-representation-...
[ { "id": "2611.11531", "score": 27.1168, "via": [ "representation-quality-evaluation", "representation-probing-and-comparison", "interpretability-and-probing-evaluation", "hidden-state-probing-representation-analysis", "pretrained-language-model-representations-and-adaptatio...
html
11
{ "dir": "papers/1909.03368", "guide": "papers/1909.03368/guide.json", "extraction": "papers/1909.03368/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1909.03368v1", "pdf": "https://arxiv.org/pdf/1909.03368v1", "html": "https://arxiv.org/html/1909.03368v1" }
1909.03480
2
Story Realization: Expanding Plot Events into Sentences
[ "Prithviraj Ammanabrolu", "Ethan Tien", "Wesley Cheung", "Zhaochen Luo", "William Ma", "Lara J. Martin", "Mark O. Riedl" ]
Neural network based approaches to automated story plot generation attempt to learn how to generate novel plots from a corpus of natural language plot summaries. Prior work has shown that a semantic abstraction of sentences called events improves neural plot generation and and allows one to decompose the problem into: ...
[ "cs.CL", "cs.AI", "cs.LG" ]
2019-09-08T15:09:32
2019-11-21T18:32:23
10.1609/aaai.v34i05.6232
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper studies story realization: transforming a sequence of plot events into natural language sentences within automated story generation. It frames event-to-sentence generation as guided language generation and proposes an ensemble combining retrieve-and-edit, sentence templating, Monte Carlo beam search, finite-s...
[ "Event-guided language generation", "Constrained decoding", "Ensemble methods for text generation", "Human evaluation of story generation", "Semantic event representation", "Plot realization" ]
[ { "id": "automated-story-generation", "fit": 0.9996481034 }, { "id": "ai-for-computational-narrative-generation-and-understanding", "fit": 0.9945507484 }, { "id": "plan-and-outline-conditioned-text-generation", "fit": 0.934135581 }, { "id": "constrained-neuro-symbolic-text-ge...
[ { "id": "2608.05169", "score": 14.0211, "via": [ "constrained-neuro-symbolic-text-generation", "automated-story-generation", "ai-for-computational-narrative-generation-and-understanding" ], "similarity": 0.8245 }, { "id": "2412.13575", "score": 13.3578, "via": [...
html
9
{ "dir": "papers/1909.03480", "guide": "papers/1909.03480/guide.json", "extraction": "papers/1909.03480/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1909.03480v2", "pdf": "https://arxiv.org/pdf/1909.03480v2", "html": "https://arxiv.org/html/1909.03480v2" }
1909.08053
4
Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
[ "Mohammad Shoeybi", "Mostofa Patwary", "Raul Puri", "Patrick LeGresley", "Jared Casper", "Bryan Catanzaro" ]
Recent work in language modeling demonstrates that training large transformer models advances the state of the art in Natural Language Processing applications. However, very large models can be quite difficult to train due to memory constraints. In this work, we present our techniques for training very large transforme...
[ "cs.CL" ]
2019-09-17T19:42:54
2020-03-13T23:45:18
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Megatron-LM introduces intra-layer model parallelism for training very large transformer language models under memory constraints. The method is implemented with communication operations in native PyTorch, is complementary to pipeline parallelism, and is combined with data parallelism. The paper trains GPT-2-like and B...
[ "large-scale language model training", "model parallelism", "efficient distributed training", "layer normalization in BERT-like models", "scaling language models", "NLP state of the art" ]
[ { "id": "distributed-deep-learning-systems", "fit": 0.9993495964 }, { "id": "large-scale-language-model-pretraining", "fit": 0.9978212306 }, { "id": "deep-network-training-efficiency-and-optimization", "fit": 0.9924422262 }, { "id": "large-language-models", "fit": 0.98989...
[ { "id": "2104.04473", "score": 24.2471, "via": [ "tensor-parallelism-large-model-systems", "large-scale-language-model-pretraining", "distributed-deep-learning-systems", "deep-network-training-efficiency-and-optimization", "gpu-utilization-throughput-optimization", "l...
html
15
{ "dir": "papers/1909.08053", "guide": "papers/1909.08053/guide.json", "extraction": "papers/1909.08053/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/1909.08053v4", "pdf": "https://arxiv.org/pdf/1909.08053v4", "html": "https://arxiv.org/html/1909.08053v4" }
1910.09700
2
Quantifying the Carbon Emissions of Machine Learning
[ "Alexandre Lacoste", "Alexandra Luccioni", "Victor Schmidt", "Thomas Dandres" ]
From an environmental standpoint, there are a few crucial aspects of training a neural network that have a major impact on the quantity of carbon that it emits. These factors include: the location of the server used for training and the energy grid that it uses, the length of the training procedure, and even the make a...
[ "cs.CY", "cs.LG" ]
2019-10-21T23:57:32
2019-11-04T20:37:33
null
http://creativecommons.org/licenses/by-sa/4.0/
The paper introduces the Machine Learning Emissions Calculator, a community tool for approximating the carbon emissions of training ML models. It explains that emissions depend on the training server's location and energy grid, training duration, and hardware make and model, and it discusses mitigation actions for prac...
[ "carbon footprint of ML training", "cloud server carbon intensity", "hardware efficiency", "emissions mitigation actions", "hyperparameter optimization efficiency", "inference and deployment energy use" ]
[ { "id": "resource-efficient-machine-learning", "fit": 0.9600206161 } ]
[ { "id": "2212.05055", "score": 4.9345, "via": [ "resource-efficient-machine-learning" ], "similarity": 0.7076 }, { "id": "2302.01382", "score": 4.916, "via": [ "resource-efficient-machine-learning" ], "similarity": 0.7016 }, { "id": "2211.15841", "...
html
8
{ "dir": "papers/1910.09700", "pdf": "papers/1910.09700/paper.pdf", "text": "papers/1910.09700/text.md", "pages": "papers/1910.09700/pages", "guide": "papers/1910.09700/guide.json", "extraction": "papers/1910.09700/extraction.jsonl" }
included
{ "abs": "https://arxiv.org/abs/1910.09700v2", "pdf": "https://arxiv.org/pdf/1910.09700v2", "html": "https://arxiv.org/html/1910.09700v2" }
2001.06782
4
Gradient Surgery for Multi-Task Learning
[ "Tianhe Yu", "Saurabh Kumar", "Abhishek Gupta", "Sergey Levine", "Karol Hausman", "Chelsea Finn" ]
While deep learning and deep reinforcement learning (RL) systems have demonstrated impressive results in domains such as image classification, game playing, and robotic control, data efficiency remains a major challenge. Multi-task learning has emerged as a promising approach for sharing structure across multiple tasks...
[ "cs.LG", "cs.CV", "cs.RO", "stat.ML" ]
2020-01-19T06:33:47
2020-12-22T00:35:46
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper studies why multi-task learning is difficult to optimize, focusing on detrimental interference between task gradients. It identifies a 'tragic triad' of conflicting gradients, high curvature, and large gradient-magnitude differences, then proposes PCGrad, a model-agnostic gradient-surgery method that projects...
[ "multi-task optimization", "gradient interference", "gradient conflict resolution", "data-efficient learning", "multi-task reinforcement learning", "model-agnostic training" ]
[ { "id": "multi-task-learning-optimization-and-gradient-conflict", "fit": 0.9997912885 }, { "id": "optimization-and-inductive-bias-in-deep-networks", "fit": 0.997404356 }, { "id": "deep-network-training-efficiency-and-optimization", "fit": 0.9905430518 } ]
[ { "id": "2110.14048", "score": 15.3626, "via": [ "multi-task-learning-optimization-and-gradient-conflict", "optimization-and-inductive-bias-in-deep-networks", "deep-network-training-efficiency-and-optimization" ], "similarity": 0.9037 }, { "id": "1711.02257", "score...
html
27
{ "dir": "papers/2001.06782", "guide": "papers/2001.06782/guide.json", "extraction": "papers/2001.06782/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/2001.06782v4", "pdf": "https://arxiv.org/pdf/2001.06782v4", "html": "https://arxiv.org/html/2001.06782v4" }
2004.04906
3
Dense Passage Retrieval for Open-Domain Question Answering
[ "Vladimir Karpukhin", "Barlas Oğuz", "Sewon Min", "Patrick Lewis", "Ledell Wu", "Sergey Edunov", "Danqi Chen", "Wen-tau Yih" ]
Open-domain question answering relies on efficient passage retrieval to select candidate contexts, where traditional sparse vector space models, such as TF-IDF or BM25, are the de facto method. In this work, we show that retrieval can be practically implemented using dense representations alone, where embeddings are le...
[ "cs.CL" ]
2020-04-10T04:53:17
2020-09-30T21:27:13
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper proposes Dense Passage Retrieval (DPR), a dual-encoder retriever for open-domain question answering. It learns dense question and passage embeddings from question-passage pairs, indexes passages for maximum inner product search, and compares with sparse BM25/TF-IDF retrieval. The guide selects the main retrie...
[ "dense versus sparse retrieval", "two-stage retriever-reader QA", "semantic passage retrieval", "negative sampling for retrieval", "efficient vector search", "cross-dataset retrieval generalization" ]
[ { "id": "retrieval-for-qa", "fit": 0.9996091404 }, { "id": "dense-retrieval", "fit": 0.9976796917 }, { "id": "embedding-based-representation-learning-and-retrieval", "fit": 0.9967992406 }, { "id": "neural-retrieval-systems-and-efficiency", "fit": 0.9957803066 }, { ...
[ { "id": "2004.12832", "score": 15.8426, "via": [ "dense-retrieval", "neural-retrieval-systems-and-efficiency", "retrieval-for-qa", "embedding-based-representation-learning-and-retrieval" ], "similarity": 0.8139 }, { "id": "2212.10496", "score": 15.8046, "v...
html
13
{ "dir": "papers/2004.04906", "guide": "papers/2004.04906/guide.json", "extraction": "papers/2004.04906/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/2004.04906v3", "pdf": "https://arxiv.org/pdf/2004.04906v3", "html": "https://arxiv.org/html/2004.04906v3" }
2004.07780
5
Shortcut Learning in Deep Neural Networks
[ "Robert Geirhos", "Jörn-Henrik Jacobsen", "Claudio Michaelis", "Richard Zemel", "Wieland Brendel", "Matthias Bethge", "Felix A. Wichmann" ]
Deep learning has triggered the current rise of artificial intelligence and is the workhorse of today's machine intelligence. Numerous success stories have rapidly spread all over science, industry and society, but its limitations have only recently come into focus. In this perspective we seek to distill how many of de...
[ "cs.CV", "cs.AI", "cs.LG", "q-bio.NC" ]
2020-04-16T17:18:49
2023-11-21T15:22:43
10.1038/s42256-020-00257-z
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
This perspective argues that many deep-learning failures are symptoms of shortcut learning: decision rules that perform well on standard benchmarks but fail to transfer to more challenging or real-world conditions. It connects shortcut learning to related phenomena in comparative psychology, education, and linguistics,...
[ "shortcut learning", "robustness and transferability", "out-of-distribution generalisation", "model interpretation and benchmarking", "fairness and bias", "biological and artificial learning systems" ]
[ { "id": "robustness-and-reliability-evaluation-beyond-standard-accuracy", "fit": 0.9653205219 }, { "id": "trustworthy-and-reliable-machine-learning-in-deployment", "fit": 0.9144512619 } ]
[ { "id": "2011.03395", "score": 10.8449, "via": [ "robustness-and-reliability-evaluation-beyond-standard-accuracy", "trustworthy-and-reliable-machine-learning-in-deployment" ], "similarity": 0.7724 }, { "id": "1906.02845", "score": 10.6331, "via": [ "robustness-a...
html
29
{ "dir": "papers/2004.07780", "guide": "papers/2004.07780/guide.json", "extraction": "papers/2004.07780/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/2004.07780v5", "pdf": "https://arxiv.org/pdf/2004.07780v5", "html": "https://arxiv.org/html/2004.07780v5" }
2004.10964
3
Don't Stop Pretraining: Adapt Language Models to Domains and Tasks
[ "Suchin Gururangan", "Ana Marasović", "Swabha Swayamdipta", "Kyle Lo", "Iz Beltagy", "Doug Downey", "Noah A. Smith" ]
Language models pretrained on text from a wide variety of sources form the foundation of today's NLP. In light of the success of these broad-coverage models, we investigate whether it is still helpful to tailor a pretrained model to the domain of a target task. We present a study across four domains (biomedical and com...
[ "cs.CL", "cs.LG" ]
2020-04-23T04:21:19
2020-05-05T22:00:44
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper asks whether broad-coverage pretrained language models still benefit from further adaptation to a target textual domain and task. Using RoBERTa, it studies domain-adaptive pretraining (continued pretraining on in-domain unlabeled text) and task-adaptive pretraining (continued pretraining on unlabeled task dat...
[ "adaptive pretraining", "domain adaptation", "task adaptation", "data selection", "low-resource NLP", "transfer learning" ]
[ { "id": "continued-pretraining-llms", "fit": 0.9946772784 }, { "id": "cross-task-and-cross-domain-transfer", "fit": 0.9442596891 }, { "id": "efficient-adaptation-of-pretrained-models", "fit": 0.9423192599 } ]
[ { "id": "2507.17168", "score": 10.3716, "via": [ "continued-pretraining-llms", "cross-task-and-cross-domain-transfer" ], "similarity": 0.7278 }, { "id": "2010.02307", "score": 5.7433, "via": [ "cross-task-and-cross-domain-transfer" ], "similarity": 0.795...
html
19
{ "dir": "papers/2004.10964", "guide": "papers/2004.10964/guide.json", "extraction": "papers/2004.10964/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/2004.10964v3", "pdf": "https://arxiv.org/pdf/2004.10964v3", "html": "https://arxiv.org/html/2004.10964v3" }
2004.11362
5
Supervised Contrastive Learning
[ "Prannay Khosla", "Piotr Teterwak", "Chen Wang", "Aaron Sarna", "Yonglong Tian", "Phillip Isola", "Aaron Maschinot", "Ce Liu", "Dilip Krishnan" ]
Contrastive learning applied to self-supervised representation learning has seen a resurgence in recent years, leading to state of the art performance in the unsupervised training of deep image models. Modern batch contrastive approaches subsume or significantly outperform traditional contrastive losses such as triplet...
[ "cs.LG", "cs.CV", "stat.ML" ]
2020-04-23T17:58:56
2021-03-10T19:11:45
null
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The paper extends batch contrastive representation learning from self-supervised to fully supervised settings by using labels: the supervised contrastive (SupCon) loss pulls same-class embeddings together and pushes different-class embeddings apart. It analyzes alternative SupCon formulations, selects a better-performi...
[ "contrastive learning", "supervised representation learning", "image classification", "robustness to natural corruptions", "hyperparameter stability", "calibration" ]
[ { "id": "embedding-based-representation-learning-and-retrieval", "fit": 0.9873706516 }, { "id": "representation-learning-and-learned-representation-analysis", "fit": 0.9465572203 } ]
[ { "id": "2007.00224", "score": 7.2997, "via": [ "representation-learning-and-learned-representation-analysis", "embedding-based-representation-learning-and-retrieval" ], "similarity": 0.7849 }, { "id": "2105.04906", "score": 7.2985, "via": [ "representation-lear...
html
23
{ "dir": "papers/2004.11362", "guide": "papers/2004.11362/guide.json", "extraction": "papers/2004.11362/extraction.jsonl" }
withheld
{ "abs": "https://arxiv.org/abs/2004.11362v5", "pdf": "https://arxiv.org/pdf/2004.11362v5", "html": "https://arxiv.org/html/2004.11362v5" }
End of preview. Expand in Data Studio

arXiv research library

882 arXiv papers, read page by page by a vision language model, given a short guide, and placed in a map of topics and areas. Two folders hold the same library:

  • vault/ is an Obsidian vault. Open the folder as a vault. Start at Start here.md; each paper's page has its authors, topics, related papers, a generated summary, the abstract and the PDF.
  • library/ is for programs and research agents. Start at library/catalog.jsonl: one line per paper with its metadata, license, synopsis, topics, related papers and the paths of its files. library/topics.json says what each topic id means and how topics group into areas.

Each paper's folder, library/papers/<id>/, holds:

File What it is
paper.json arXiv metadata, the license, the paper's topics with their fit
guide.json a compact guide to the paper, generated from the page records; pdf:page:N cites page N
extraction.jsonl one generated record per page: what the page says, its methods, claims and evidence
paper.pdf, paper.html the paper from arXiv (HTML when arXiv has it)
text.md the text, from the HTML (TeX math, tables) or else from the PDF (with ## Page N headings)
pages/0001.png, ... every page rendered at 200 DPI: look here when the text transcription is in doubt

The summaries, guides, page records and topics are generated by language models and are unverified.

Licenses: where to read each paper

Each paper keeps the license its authors chose on arXiv, so the dataset as a whole has no single license. A paper's license is in its paper.json, its catalog line and its vault page.

  • 507 papers are under a Creative Commons license. The paper itself is in this dataset: paper.pdf, paper.html where arXiv has one, text.md and pages/.
  • 375 papers are under arXiv's non-exclusive license, which lets only arXiv distribute them. Read these on arXiv: the paper itself is not in this dataset. Each one's catalog line has "source": "withheld" and the arXiv links (arxiv.abs, arxiv.pdf, arxiv.html), and its vault page links to the PDF on arXiv where the PDF would be. Everything else about it is here: metadata, abstract, guide, page records, topics and related papers.
License Papers The paper itself
CC BY 4.0 406 in this dataset
arXiv non-exclusive 375 on arXiv only
CC BY-NC-SA 4.0 39 in this dataset
CC BY-NC-ND 4.0 30 in this dataset
CC BY-SA 4.0 23 in this dataset
CC0 1.0 9 in this dataset

Attribution. The papers are the work of their authors, named in each paper.json, catalog line and vault page, with a link to the paper on arXiv and to its license. text.md and pages/ are conversions of the arXiv PDF or HTML (text extraction and page rendering); the summaries, guides and page records are new text about the papers.

Removal. If you are an author and want your paper removed, open a discussion on this dataset's Community tab with the arXiv id.

Downloads last month
184