那一抹天空蓝

机器学习顶级论文及实现（附地址及简介）

2017/12/22 Machine-Learning

State-of-the-art result for all Machine Learning Problems

LAST UPDATE: 17th November 2017

NEWS: I am looking for a Collaborator esp who does research in NLP, Computer Vision and Reinforcement learning. If you are not a researcher, but you are willing, contact me. Email me: redditsota@gmail.com

This repository provides state-of-the-art (SoTA) results for all machine learning problems. We do our best to keep this repository up to date. If you do find a problem’s SoTA result is out of date or missing, please raise this as an issue (with this information: research paper name, dataset, metric, source code and year). We will fix it immediately.

You can also submit this Google Form if you are new to Github.

This is an attempt to make one stop for all types of machine learning problems state of the art result. I can not do this alone. I need help from everyone. Please submit the Google form/raise an issue if you find SOTA result for a dataset. Please share this on Twitter, Facebook, and other social media.

This summary is categorized into:

Supervised Learning

NLP

1. Language Modelling

Research Paper Datasets Metric Source Code Year
BREAKING THE SOFTMAX BOTTLENECK: A HIGH-RANK RNN LANGUAGE MODEL
  • PTB
  • WikiText-2
  • Perplexity: 47.69
  • Perplexity: 40.68
Pytorch 2017
DYNAMIC EVALUATION OF NEURAL SEQUENCE MODELS
  • PTB
  • WikiText-2
  • Perplexity: 51.1
  • Perplexity: 44.3
Pytorch 2017
Averaged Stochastic Gradient Descent
with Weight Dropped LSTM or QRNN
  • PTB
  • WikiText-2
  • Perplexity: 52.8
  • Perplexity: 52.0
Pytorch 2017
FRATERNAL DROPOUT
  • PTB
  • WikiText-2
  • Perplexity: 56.8
  • Perplexity: 64.1
Pytorch 2017
Factorization tricks for LSTM networks One Billion Word Benchmark Perplexity: 23.36 Tensorflow 2017

2. Machine Translation

</tr> </tbody> </table> #### 3. Text Classification
Research Paper Datasets Metric Source Code Year
WEIGHTED TRANSFORMER NETWORK FOR MACHINE TRANSLATION
  • WMT 2014 English-to-French
  • WMT 2014 English-to-German
  • BLEU: 41.4
  • BLEU: 28.9
2017
Attention Is All You Need
  • WMT 2014 English-to-French
  • WMT 2014 English-to-German
  • BLEU: 41.0
  • BLEU: 28.4
2017
NON-AUTOREGRESSIVE NEURAL MACHINE TRANSLATION
  • WMT16 Ro→En
  • BLEU: 31.44
2017
Improving Neural Machine Translation with Conditional Sequence Generative Adversarial Nets
  • NIST02
  • NIST03
  • NIST04
  • NIST05
  • 38.74
  • 36.01
  • 37.54
  • 33.76
    • </ul     		2017
    </tr> </tr> </tbody> </table> #### 4. Natural Language Inference Leader board: [Stanford Natural Language Inference (SNLI)](https://nlp.stanford.edu/projects/snli/) [MultiNLI](https://www.kaggle.com/c/multinli-matched-open-evaluation/leaderboard)
    Research Paper Datasets Metric Source Code Year
    Learning Structured Text Representations Yelp Accuracy: 68.6 2017
    Attentive Convolution Yelp Accuracy: 67.36 2017
    Research Paper Datasets Metric Source Code Year
    NATURAL LANGUAGE INFERENCE OVER INTERACTION SPACE Stanford Natural Language Inference (SNLI) Accuracy: 88.9 Tensorflow 2017
    #### 5. Question Answering Leader Board [SQuAD](https://rajpurkar.github.io/SQuAD-explorer/)
    Research Paper Datasets Metric Source Code Year
    Interactive AoA Reader+ (ensemble) The Stanford Question Answering Dataset
    • Exact Match: 79.083
    • F1: 86.450
    		NOT FOUND 2017
    #### 6. Named entity recognition
    Research Paper Datasets Metric Source Code Year
    Named Entity Recognition in Twitter using Images and Text Ritter
    • F-measure: 0.59
    		NOT FOUND 2017
    #### 7. Abstractive Summarization Research Paper | Datasets | Metric | Source Code | Year ------------ | ------------- | ------------ | ------------- | ------------- [Cutting-off redundant repeating generations </br> for neural abstractive summarization](https://aclanthology.info/pdf/E/E17/E17-2047.pdf) |
    • DUC-2004
    • Gigaword
    |
    • DUC-2004
      • ROUGE-1: **32.28**
      • ROUGE-2: 10.54
      • ROUGE-L: **27.80**
    • Gigaword
      • ROUGE-1: **36.30**
      • ROUGE-2: 17.31
      • ROUGE-L: **33.88**
    | NOT YET AVAILABLE | 2017 [Convolutional Sequence to Sequence](https://arxiv.org/pdf/1705.03122.pdf) |
    • DUC-2004
    • Gigaword
    |
    • DUC-2004
      • ROUGE-1: 33.44
      • ROUGE-2: **10.84**
      • ROUGE-L: 26.90
    • Gigaword
      • ROUGE-1: 35.88
      • ROUGE-2: 27.48
      • ROUGE-L: 33.29
    | [PyTorch](https://github.com/facebookresearch/fairseq-py) | 2017 #### 8. Dependency Parsing Research Paper | Datasets | Metric | Source Code | Year ------------ | ------------- | ------------ | ------------- | ------------- [Globally Normalized Transition-Based Neural Networks](https://arxiv.org/pdf/1603.06042.pdf) |
    • Final CoNLL ’09 dependency parsing
    |
    • 94.08% UAS accurancy
    • 92.15% LAS accurancy
    |
    • [SyntaxNet](https://github.com/tensorflow/models/tree/master/research/syntaxnet)
    |
    • 2017
    ### Computer Vision #### 1. Classification </tr>       </tr> </tbody> </table> #### 2. Instance Segmentation
    Research Paper Datasets Metric Source Code Year
    Dynamic Routing Between Capsules
    • MNIST
    • Test Error: 0.25±0.005
    		2017
    High-Performance Neural Networks for Visual Object Classification
    • NORB
    • Test Error: 2.53 ± 0.40
    		2011
    ShakeDrop regularization
    • CIFAR-10
    • CIFAR-100
    • Test Error: 2.31%
    • Test Error: 12.19%
    		2017
    Aggregated Residual Transformations for Deep Neural Networks
    • CIFAR-10
    • Test Error: 3.58%
    		2017
    Random Erasing Data Augmentation
    • CIFAR-10
    • CIFAR-100
    • Fashion-MNIST
    • Test Error: 3.08%
    • Test Error: 17.73%
    • Test Error: 3.65%
    		Pytorch </td> 2017
    Dynamic Routing Between Capsules
    • MultiMNIST
    • Test Error: 5%
    		2017
    Learning Transferable Architectures for Scalable Image Recognition
    • ImageNet-1k
    • Top-1 Error:17.3
    		2017
    Squeeze-and-Excitation Networks
    • ImageNet-1k
    • Top-1 Error: 18.68
    		2017
    Aggregated Residual Transformations for Deep Neural Networks
    • ImageNet-1k
    • Top-1 Error: 20.4%
    		2016
    Research Paper Datasets Metric Source Code Year
    Mask R-CNN
    • COCO
    • Average Precision: 37.1%
    		2017
    #### 3. Visual Question Answering
    Research Paper Datasets Metric Source Code Year
    Tips and Tricks for Visual Question Answering: Learnings from the 2017 Challenge
    • VQA
    • Overall score: 69
    		</li></ul> 2017
    #### 4. Person Re-identification       </tr> </tbody> </table> ### Speech [Speech SOTA](https://github.com/syhw/wer_are_we) #### 1. ASR
    Research Paper Datasets Metric Source Code Year
    Random Erasing Data Augmentation
    • Rank-1: 89.13 mAP: 83.93
    • Rank-1: 84.02 mAP: 78.28
    • labeled (Rank-1: 63.93 mAP: 65.05) detected (Rank-1: 64.43 mAP: 64.75)
    		Pytorch </td> 2017
    Research Paper Datasets Metric Source Code Year
    The Microsoft 2017 Conversational Speech Recognition System
    • Switchboard Hub5'00
    • WER: 5.1
    		2017
    ## Semi-supervised Learning #### Computer Vision </tr>       </tr> </tbody> </table> ## Unsupervised Learning #### Computer Vision ##### 1. Generative Model
    Research Paper Datasets Metric Source Code Year
    DISTRIBUTIONAL SMOOTHINGWITH VIRTUAL ADVERSARIAL TRAINING
    • SVHN
    • NORB
    • Test error: 24.63
    • Test error: 9.88
    		Theano 2016
    Virtual Adversarial Training: a Regularization Method for Supervised and Semi-supervised Learning
    • MNIST
    • Test error: 1.27
    		2017
    Unlabeled Samples Generated by GAN Improve the Person Re-identification Baseline in vitro
    • Rank-1: 83.97 mAP: 66.07
    • Rank-1: 84.6 mAP: 87.4
    • Rank-1: 67.68 mAP: 47.13
      •          
    • Test Accuracy: 84.4
    		Matconvnet </td> 2017
    Research Paper Datasets Metric Source Code Year
    PROGRESSIVE GROWING OF GANS FOR IMPROVED QUALITY, STABILITY, AND VARIATION Unsupervised CIFAR 10 Inception score: 8.80 Theano 2017
    ### NLP #### Machine Translation </tr> </tbody> </table> ## Transfer Learning
    Research Paper Datasets Metric Source Code Year
    UNSUPERVISED MACHINE TRANSLATION USING MONOLINGUAL CORPORA ONLY
    • WMT16 (en-fr fr-en de-en en-de)
    • Multi30k-Task1(en-fr fr-en de-en en-de)
    • BLEU:(32.76 32.07 26.26 22.74)
    • BLEU:(15.05 14.31 13.33 9.64)
    		2017
    </tr> </tbody> </table> ## Reinforcement Learning
    Research Paper Datasets Metric Source Code Year
    One Model To Learn Them All
    • WMT EN → DE
    • WMT EN → FR (BLEU)
    • ImageNet (top-5 accuracy)
    • BLEU: 21.2
    • BLEU:30.5
    • 86%
    		2017
    </tr> </tbody> </table> Email: redditsota@gmail.com
    Research Paper Datasets Metric Source Code Year
    Mastering the game of Go without human knowledge the game of Go ElO Rating: 5185
    • <a href=https://github.com/gcp/leela-zero>C++</a></ul>
      • </td>
    		2017

    Search

      Table of Contents