2018 DS/ML digest 21

2018 DS/ML digest 21

Posted by snakers41 on August 21, 2018

Buy me a coffeeBuy me a coffee

Become a PatronBecome a Patron

2018 DS/ML digest 21

News / market / aticles:

  • New Nvidia GPUs:
    • TLDR - +20-30$ performance on the Ti model for CNNs;
    • Only 12GB of memory on the Ti model;
    • Founders’ edition costs US$1500 if ordered from the CIS;
    • Used 1080Ti can be bought 3x cheaper now;
  • A small post on finding optimal LR
  • How Google’s Smart Linkify works
  • How to publish medical CV results properly
  • ICML papers by Topic
  • Apple has its own “easy toolkit” for CNNs. Have not heard about people using it in practice;
  • Keras vis applied to emotion recognition. TLDR - despite Google’s attempts on distill / colab etc - there are no REAL ACCESSIBLE options for CNN LAYER visualization;
  • Data Engineering on Upwork https://habr.com/post/417745/


  • Face recognition is limited by noise the datasets =)

Implementations / code / cool libraries

  • Searching similar vectors on billion scale / building KNN graph on GPUs:
    • Facebooks FAISS - it has KNN graph running on several GPUs
    • UMAP works well up to 1m-3m, then you may face this