Project Overview
I worked on this project during the tenure of my internship, where my mission was to analyze and explain the inner workings of the CNN models, especially how they differentiate features from images of different classes. I performed feature map extraction, then generated a reduced set of Class Activated Mappings from the feature maps and plotted heatmaps, in an effort to understand which portions of images was the CNN observing. Then, I performed feature clustering to understand which features from different images were being grouped together to reach the classifications.
Detailed Description
For maintaining a general approach, I used the VGG16 model and the CIFAR-10 dataset to perform the experiments. Instead of just plugging in CNN models in Computer Vision applications, my intention was to create a structured report as to what influences the CNN's decision-making, which in turn would help us understand how the CNN needs to be tuned for better performance in classification.
