Design and Architecture of convolutional neural network

Date:

This talk presents an overview of the domain of deep learning.

  • Explored the design principles and difficulties in training large scale deep neural networks such as initialisation strategy, batch normalisation, dropout and residual connection.
  • Implemented activation functions such as sigmoid and ReLU along with hyper-parameter tuning.
  • Evaluated loss functions such as cross entropy, weighted cross entropy, dice, inverse dice and focal loss with corresponding changes in the output.

The resources for the talk are here.