Interpreting Neural Networks through Mahalanobis Distance

Neural networks have revolutionized machine learning, achieving remarkable success across diverse applications. Central to their efficacy is the use of activation functions, which introduce non-linearity and enable the modeling of complex relationships within data. While Rectified Linear Units (ReLU) have gained prominence due to their simplicity and effectiveness [Nair and Hinton, 2010], the exploration of alternative activation functions remains an open and valuable area of research [Ramachandran et al., 2018]. Neural network units are often viewed as linear separators that define decision boundaries between classes [Minsky and Papert, 1969] with larger activation values suggesting stronger contributions of features to those decisions. Our work challenges this perspective, exploring how individual neurons can be understood through the lens of statistical distance measures.