Overfitting control with Inverse Cross Entropy loss function (ICE)

This is the source code for paper Overfitting control with Inverse Cross Entropy loss function (ICE).

In machine learning research community almost all the time researchers use cross entropy between empirical probability function and target energy function as loss function. In this paper we explore the possibility of doing this in opposite direction in other words taking cross entropy between empirical energy function and the target probability function. We call resulting loss function Inverse Cross Entropy (ICE). We prove theoretically and experimentally that this will help us to have control over overfitting, which is not directly possible with former method.

## Install the dependencies

Install pytorch https://pytorch.org/get-started/locally/

Install websockets

pip install websockets

Run entry point

python main.py

Open http://127.0.0.1:4444/ in browser