Weight-based Neural Network Interpretability using Activation Tuning and Personalized Products

We introduce approaches to simplifying neural networks and enhancing their interpretability using activation-based neuron tuning and personalized weight matrix products. Inspired by the evolutionary principle of the survival of the fittest, we gradually remove neurons with little to no learning efficacy during training and hypothesize that their absence renders opaque models more interpretable. Experimental results pertaining to cancer and diabetes treatment appear to favor our hypothesis and generate more biomedically salient results. Our approaches also allow for interpretations at the sample level, a feature of particular importance in relation to personalized medicine.