TorchPRISM, a PyTorch software that implements Principal Image Sections Mapping (PRISM) but is easily portable to other deep learning frameworks, is presented in this study. PRISM relies on Principal Component Analysis to visualise the most significant convolutional neural network features. Moreover, PRISM supports Explanation by Example by displaying a comparative set of features between images processed in the same batch. Furthermore, TorchPRISM can generate and provide clustering based on PRISM output that can serve as an indicator for distinguishing ambiguous classes. This study presents experimental results and recommendations for software development.
Explore By Topic