sooryakiran / harts

Hardware Architecture Search (Harts), finds an optimal of logic gates to perform a prediction.

HARTS

Hardware Architecture Search (Harts), finds an optimal architecture of logic gates to perform a prediction. The search space is discrete, but we make it continious by taking choices based on softmax over a trainable weights. The weights can be trained using back propagation. A possible hardware architecture / arrangement of logic gates can be obtained by going greedy on the optimised weights.

DISCLAIMER : This is not a Binary Neural Network. This can be considered rather as an 'architecture search'.

Classification of MNIST using logic gates

The above code can be used to generate optimal hardware architectures for prediction tasks. Here we present an example on the MNIST dataset. We were able to achieve a testing accuracy of 73% using just 5 layers of logic gates. The fact that such a search converges makes me happy :)

Training loss

Training accuracy

TODO

1. HDL generation from the trained networks