0%

DL with discrete features

Reviews on deep learning w. discrete features

ref:

https://mp.weixin.qq.com/s/HhrnCKUvNHnZDgoGF52Vuw

Properties of Deep learning

  • Super memorizer: easily fit a random labeling
  • super energy sucker
  • overly parameterized ( compression while maintaining similar accuracy is easy )
  • 先略后详(learning process)
    • from easy to difficult
    • from smooth to noisy
  • Robust to massive label noise (compared with MLP, Perceptron)

Learning with more knowledge(which can be discrete features)

Input

  • Multi-modal inputs
  • synthesized inputs
  • others

Output

  • Multi-task learning(more outputs?)
  • more representations
  • more priors