Seeing Convolution Through the Eyes of Finite Transformation Semigroup Theory: An Abstract Algebraic Interpretation of Convolutional Neural Networks Andrew Hryniowski1;2;3, Alexander Wong1;2;3 1 Video and Image Processing Research Group, Systems Design Engineering, University of Waterloo 2 Waterloo Artificial Intelligence Institute, Waterloo, ON 3 DarwinAI Corp., Waterloo, ON fapphryni,
[email protected] Abstract of applications, particularly for prediction using structured data. Despite such successes, a major challenge with lever- Researchers are actively trying to gain better insights aging convolutional neural networks is the sheer number of into the representational properties of convolutional neural learnable parameters within such networks, making under- networks for guiding better network designs and for inter- standing and gaining insights about them a daunting task. As preting a network’s computational nature. Gaining such such, researchers are actively trying to gain better insights insights can be an arduous task due to the number of pa- and understanding into the representational properties of rameters in a network and the complexity of a network’s convolutional neural networks, especially since it can lead architecture. Current approaches of neural network inter- to better design and interpretability of such networks. pretation include Bayesian probabilistic interpretations and One direction that holds a lot of promise in improving information theoretic interpretations. In this study, we take a understanding of convolutional neural networks, but is much different approach to studying convolutional neural networks less explored than other approaches, is the construction of by proposing an abstract algebraic interpretation using finite theoretical models and interpretations of such networks. Cur- transformation semigroup theory.