Feature specific analysis of a deep convolutional neural network for ageing classification
Shaun Barney, Scott Stainton, Michael Catt, Satnam Singh Dlay · 2018
Deep learning is a sub-field of machine learning with their models being vaguely inspired by the communication patterns and information processing of a nervous system. In a deep learning model each layer of the network, which is build up of multiple neurons, transforms its input data into slightly more abstract and composite abstractions. Within a convolutional neural network (CNN), the first layer of neurons may be transform its raw input pixels to encoded edges, with the next layer encoding arrangements of edges, with the next layer more abstract features such as eyes and noses. However, locating which exact neurons in each layer are looking for each abstract feature is still a hot area for research. Here we show a novel way to pinpoint each neuron in the hidden layers that is searching for a specific feature that is believed to play a pivotal role in the classification task. We found that by removing the feature of interest from the original image and running this through the CNN, each neurons activation map can be compared to the activation generated by the original image using t-SNE. This allowed us to successfully located the individual neurons in the network that are significantly effected by the removal of this feature and hence changing the classification. This method allows us to get a better understanding of what the network has learnt and how important this learnt information is when coming to the final classification.