"No one is harder on a talented person than the person themselves" - Linda Wilkinson ; "Trust your guts and don't follow the herd" ; "Validate direction not destination" ;

May 14, 2017

Weekend Seminar - Deep learning in production at Facebook

Good Talk - Deep learning in production at Facebook https://lnkd.in/fX7BZif

Notes from Session
Deep Learning Use Cases
  • Event Prediction - Listing top relevant stories for the user, predicting relevance - Approach - Logistics regression + Deep Neural Networks
  • Machine Translation - Automatically machine translated posts generated for users - Approach - Encoder - Decoder Architecture, Using RNN
  • Natural Language Processing - Understand Context of text - Deep Text - Approach - CNN for words + RNN for sequences
  • Computer Vision - Understand pics - Approach - CNN @ massive scale. Understand different aspects of pictures - Classification, Detection, Segmentation 
Scaling the models
  • Computer faster - Tweaks in FFT, TiledFFT, Winograd to reduce convolution computations, NNPack, CuDNN for CPUs
  • Memory Usage - GPU + Activations Memory released and reallocated during different layers of processing in Deep Networks
  • Compress models - Exploit redundancy in model designs, prune them
Good Insights!!!

No comments: