/state-of-the-art-result-for-machine-learning-problems

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 If you do find a problem's SoTA result is out of date or missing, please raise this as an issue or submit Google form (with this information: research paper name, dataset, metric, source code and year). We will fix it immediately.

Apache License 2.0Apache-2.0

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

LAST UPDATE: 20th Februray 2019

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: yxt.stoaml@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 If you do find a problem's SoTA result is out of date or missing, please raise this as an issue (with this information: research paper name, dataset, metric, source code and year).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
Language Models are Unsupervised Multitask Learners
  • PTB
  • WikiText-2
  • Perplexity: 35.76
  • Perplexity: 18.34
Tensorflow 2019
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

Research Paper Datasets Metric Source Code Year
Understanding Back-Translation at Scale
  • WMT 2014 English-to-French
  • WMT 2014 English-to-German
  • BLEU: 45.6
  • BLEU: 35.0
2018
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
    		• 2017

    3. Text Classification

    Research Paper Datasets Metric Source Code Year
    Learning Structured Text Representations Yelp Accuracy: 68.6 2017
    Attentive Convolution Yelp Accuracy: 67.36 2017

    4. Natural Language Inference

    Leader board:

    Stanford Natural Language Inference (SNLI)

    MultiNLI

    Research Paper Datasets Metric Source Code Year
    NATURAL LANGUAGE INFERENCE OVER INTERACTION SPACE Stanford Natural Language Inference (SNLI) Accuracy: 88.9 Tensorflow 2017
    BERT-LARGE (ensemble) Multi-Genre Natural Language Inference (MNLI)
    • Matched accuracy: 86.7
    • Mismatched accuracy: 85.9
    		 2018

    5. Question Answering

    Leader Board

    SQuAD

    Research Paper Datasets Metric Source Code Year
    BERT-LARGE (ensemble) The Stanford Question Answering Dataset
    • Exact Match: 87.4
    • F1: 93.2
    		2018

    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
    for neural abstractive summarization
    • 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
    • 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 2017

    8. Dependency Parsing

    Research Paper Datasets Metric Source Code Year
    Globally Normalized Transition-Based Neural Networks
    • Final CoNLL ’09 dependency parsing
    • 94.08% UAS accurancy
    • 92.15% LAS accurancy
    • 2017

    Computer Vision

    1. Classification

               
    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
    Giant AmoebaNet with GPipe
    • CIFAR-10
    • CIFAR-100
    • ImageNet-1k
    • ...
    • Test Error: 1.0%
    • Test Error: 8.7%
    • Top-1 Error 15.7
    • ...
    		 2018
    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 2017
    EraseReLU: A Simple Way to Ease the Training of Deep Convolution Neural Networks
    • CIFAR-10
    • CIFAR-100
    • Test Error: 3.56%
    • Test Error: 16.53%
    		Pytorch 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

    2. Instance Segmentation

    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
    		2017

    4. Person Re-identification

         
    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 2017

    Speech

    Speech SOTA

    1. ASR

    Research Paper Datasets Metric Source Code Year
    The Microsoft 2017 Conversational Speech Recognition System
    • Switchboard Hub5'00
    • WER: 5.1
    		2017
    The CAPIO 2017 Conversational Speech Recognition System
    • Switchboard Hub5'00
    • WER: 5.0
    		2017

    Semi-supervised Learning

    Computer Vision

         
    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
    Few Shot Object Detection
    • VOC2007
    • VOC2012
    • mAP : 41.7
    • mAP : 35.4
    		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 2017

    Unsupervised Learning

    Computer Vision

    1. Generative Model
    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

    Research Paper Datasets Metric Source Code Year
    UNSUPERVISED MACHINE TRANSLATION USING MONOLINGUAL CORPORA ONLY
    • Multi30k-Task1(en-fr fr-en de-en en-de)
    • BLEU:(32.76 32.07 26.26 22.74)
    		2017
    Unsupervised Neural Machine Translation with Weight Sharing
    • WMT14(en-fr fr-en)
    • WMT16 (de-en en-de)
    • BLEU:(16.97 15.58)
    • BLEU:(14.62 10.86)
    		2018

    Transfer 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

    Reinforcement Learning

    Research Paper Datasets Metric Source Code Year
    Mastering the game of Go without human knowledge the game of Go ElO Rating: 5185 2017

    Email: yxt.stoaml@gmail.com