2018 DS/ML digest 30

2018 DS/ML digest 30

Posted by snakers41 on November 28, 2018

Buy me a coffeeBuy me a coffee

Become a PatronBecome a Patron


Posts / articles

ML industrialization insanity / huge datasets

  • Tencent Ml images - 17M image dataset - marginal improvement on Imagenet with ResNet-101 (80.7%+);
  • Huge Networks from Google - 10x - 30x size, 84%+ in Imagenet;
  • Training Imagenet in 224s on a huge cluster - torus all-reduce;

NLP

  • Multi-Task Learning
    • Normally I would say that this is insanity, but these SOTA approaches so far to seem to work only on paperж
    • Demo - type in something real;
    • Blog;
    • Code;
    • The idea of multi-taks learning is sane - because model capacity is too high for such tasks;

TDS picks:

Picked Papers