Fast-BCNN Massive Neuron Skipping in Bayesian.md

Paper title:

Fast-BCNN: Massive Neuron Skipping in Bayesian Convolutional Neural Networks

Publication:

MICRO’20

Problem to solve:

BCNN is a method that places probability distributions over models’ weights in contrast to conventional CNN. It’s implemented by adding a dropout layer after every convolutional layer of a normal CNN. The problems that exist in the method is:

1. The dropout layers have not been studied. They are only added after not that important FC layers.

2. The dropout implementation brought overwhelmed computation loads.

3. BCNN works inefficiently on high-end platforms like GPU.

Major contribution:

1. Fast-BCNN skips the computations for dropped and unaffected neurons. They gives a method that decides if a neuron is considered to be unaffected.

2. They proposed a parallelism scheme that can maximize the contribution of the skipping strategy based on analysis of existing parallel schemes.