On PI-Net deep learning model for classification of images

In this note we discuss the experiment part of the paper “PINet: A Deep Learning Approach to Extract Topological Persistence Images”, where Som et al. trained a base classification model called AlexNet on Cifar10 dataset to get an accuracy of 80%. Then, they concatenated the PIs features with AlexNet base features and trained the model once again to get an accuracy of around 81%. Here we give a slight modification of the PI-Net architecture. Namely, we add two dense layers at the end of the model, the first one has 1024 neurons with ReLu activation and the last one has 10 neurons with Softmax activation, and then we use it as a base classification model on Cifar10 dataset. This enables us to reach an accuracy of 82%.