mahak13 / interpretability-of-ml-models

Exploring Interpretability of Machine Learning models for object detection. Used GoogleNet and VGG-19

Abstract

Neural Networks can undoubtedly achieve great performance on a majority of tasks nowadays. But along with that, there is a growing need for them to be able to explain their decisions. They are famously called 'black boxes', which means it is extremely difficult to understand their outputs and how they reached there. This paper aims to extend the works in [1] and explore different interpretability techniques and understand their effectiveness, for the task of object detection. This paper has used convolutional neural networks but these techniques can be tested for different tasks and networks as well. This paper explored visualizing channel attributions, features, and activations of neuron groups to understand more about what the network is learning. It was observed that these techniques in combination with interactivity can make neural networks more explainable in particular cases, not all. For a better understanding of this particular task, with Inception V1, visualizing neurons with the highest activations and also neuron groups gives more information in understanding what the neural network understands at each layer. For VGG-19, visualizing neuron groups was the most effective. [1] C. Olah, The building blocks of interpretability, Distill 3.3