The idea is to use Genetic algorithms to automatically learn Neural Network topologies as opposed to learning using Grid Search. We start with the NEAT algorithm which is depicted.
We extend the NEAT algorithm to be able to learn arbitrary Deep Neural Topologies using Genetic Algorithm which we call DeepNEAT. The following modifications are needed for this to work.
A relatively simple easier task of character recognition is selected for performance evaluation. We chose Devanagari Script Character Recognition whose SOTA is shown.
https://ieeexplore.ieee.org/document/7400041
The results we got were very promising. We got a network topology which was less complex than the SOTA mentioned in the paper but performed better on the splits.
