Deep Learning Conf 2016

A conference on deep learning.

Debugging DeepNets - practitioners black book

Submitted by Anuj Gupta (@anuj-gupta) (proposing) on May 13, 2016

This is a proposal requesting for someone to speak on this topic. If you’d like to speak, leave a comment.

Section: Full talk Technical level: Advanced Status: Cancelled

Abstract

Secret Sauce to Building production quality Deep Nets

Outline

It will be great if someone can present a talk on the following points

  • Visualize if the weights have saturated
  • Visualize if the gradients have diminished(0) or exploded
  • Strategies to prevent the network from going into state where gradients have diminished(0) or exploded
  • Strategies to recover when gradients have diminished(0) or exploded (gradient clipping)
  • Good toolkits for early diagnosis and Visualization
  • detect bad initialisations
  • Strategies to prevent overfitting
  • badly conditioned activations
  • handle unbalanced datasets
  • Other Do’s and Dont’s and other Dark Arts to build good DeepNets

Speaker bio

Anuj is heading Deep Learning efforts at Freshdesk. you can find more about me on https://in.linkedin.com/in/anuj-gupta-15585792

Comments

Login to leave a comment

You need to be a participant to comment.

{{ formTitle }}
Post a comment...
New comment

{{ errorMsg }}