New approach for DNN architecture design

Major Activities

Training deep networks suffers from the common problems such as: i) Need for a gigantic training set to confront over-fitting issue; ii) architecture adaptability problems, e.g., any amendments to a pretrained DNN requires retraining from the scratch. Therefore, there is a need to develop a computationally efficient procedure for training a DNN without making any compromise on the accuracy achieved by traditional back propagation algorithm.

In this project, we propose a two stage Algorithm for progressively growing neural networks.

Significant Results

The proposed framework has been demonstrated on regression problems such as the Boston housing price prediction problem where the method outperforms a baseline network.

In particular we also demonstrate that the procedure can be used for progressively learning the solutions of PDE. Figure below shows the adaptive approach where one progressively learn the solution of Poisson’s equation.