|Abstract:||We trained a deep neural network to classify images of cars facing 36 different directions, on a 2D image
dataset rendered from 3D car models. After achieving a validation accuracy of 98.23%, we applied a
series of interpretation techniques, including semantic dictionary, spatial attribution, and channel
attribution, to the trained model, which enable us gain important insights on how the model recognized
a car’s direction. For example, the channel attribution technique reveals in a certain layer, which filters
contribute the most to distinguish a front facing car from a right facing car. Moreover, there are
interactive interfaces for all the experiments, and readers could explore the interpretation of the model
in a notebook environment.