/BSnet

Boolean Structured Deep Learning Network (aka BullShit net)

Primary LanguagePython

BSnet

  • Boolean Structured Deep Learning Network
  • The network is designed to mimic monotone circuit in Boolean algebra theory
  • Tell the LeNet bully to get lost
  • No data augmentations, no regularization such as weight decay and dropout

Main Takeaways

  • My model has 30000 parameters compared to LeNet 60000 parameters
  • Model based on the theory of monotone circuit of Boolean algebra
  • Under certain conditions, the training optimization function is convex
  • Use fully connected layers without overfitting
  • Able to be trained on a laptop without GPU
  • Able to achieve 80% classification accuracy after masking 70% of the input image pixels to value 1. A normal deep learning model can only achieve 60%

How to Run

The commands are designed to run on Windows OS. If you are using Linux, adapt the commands accordingly.

Run the command to train a BSnet

python keras_first_network_bsnet.py >> bsnet.txt

Run the command to train a Normal Relu Network

python keras_first_network_normal.py >> normal.txt

Run the command to plot out the accuracies curves

python plot_acc.py

Model

Network design

Experiment Results

Experiment results

Links

BSnet paper link

BSautonet paper link

BSautonet GitHub

Discrete Markov Random Field Relaxation

Slideshare

That's it. Have a Nice Day!!!