Feature Sleuth: Understanding and expediting learning in Convolution Neural Networks using feature maps
By Anand Mayank, Bhupathiraju Akhilesh Varma, Chandrala Rohini
 |
|---|
Abstract
Neural networks are heavily used in image processing and classification. However, understanding the internal workings of neural networks has always been a tricky task. One way to look at how a convolutional neural network (CNN) works is by visualizing the feature maps which are obtained after passing a filter through an image. To better understand these feature maps and their purpose in CNN, we experimented with multiple ways of utilizing these in accelerating and enhancing the training of CNNs. Firstly, we proposed a similarity-based classification in which, while training the network, the feature vector is computed using the feature map of the input image data and this is compared with the previously cached feature vector for images trained so far for each label after every batch of training. If a valid match, on comparison with the true label is found, the image will be assigned the default label, thus skipping the further training process of subsequent convolution layers, which boosts and increases the efficiency of the neural network. Secondly, to achieve improved classification, all the feature maps obtained from a fully trained CNN model were once again passed for training to the CNN with the same model architecture. This can add regularization effect to the model and helps when a model overfits the training data.