You’ve Got the Wrong Outfit: Evaluating Deep Learning Paradigms on Digit and Fashion Recognition

To build more accurate and trustworthy artificial intelligence algorithms in deep learning, it is essential to understand the mechanisms driving classification systems to identify their targets. Typically, post hoc methods are used to provide insights into this process. By contrast, in this work, we investigate the possibility of using class activation maps in combination with contrastive loss to enhance the reliability of training of a deep learning model. MNIST and Fashion MNIST datasets are considered in our investigation since they have already proven a practical starting point for assessing an almost tautologic training strategy for deep learning algorithms, given that classification targets are the primary significant content of the images in these datasets. Starting from the raw comparison of accuracy and system complexity of the proposed approach, a further investigation of the technique’s feasibility in a deep learning study is conducted over six random seed splits of the training data and model performance. A modern deep learning network, such as ConvNeXT, determines whether a more robust architecture trained with the proposed mechanics provides better insights than a simple convolutional neural network. This investigation also addresses the importance of skip connections, structured learning layers, and feature map dimensions in the learning process.