Files
Abstract
Deep neural networks are unparalleled tools in Machine Learning that pro-vide near-human performance on image classification tasks [12]. However, their inability to explain their predictions is the primary barrier to adopting them despite their promising accuracy. Many model agnostic interpretation methods [26] have tried to explain the reasoning process behind such "black-box" deep learning models. Prototype-based classification is a classical form of case-based reasoning [19]that claims to be a promising alternative to using black-box image classification models. The current state-of-the-art method among prototype-based classification is the ProtoPNet [7]. However, we’ve identified numerous flaws with this approach that stem from their design choice of using a deep CNN as the base architecture. Essentially, conventional CNNs [17][15] can perform well only with deeper layers. This means that to associate image pixels far away from each other, we are forced to ensure an adequate number of layers (depth) in our model, which adds to more complexity and resource utilization. By switching to a transformer-based architecture, we can eliminate this restriction of "depth" imposed by CNNs.Hence, we propose a transformer-based image classifier architecture that uses the self-attention mechanism to learn what parts of the image are important for the given visual task, which can help encode feature maps into high-level semantic features. By the new trend of using Transformers in computer vision tasks, this work investigates the impact that Transformers can cause in the field of Interpretable Machine Learning with respect to 1. Interpretability and 2.Performance and Resource Utilization. Our results show that while transformers show promising performance in reducing the complexity of the interpretable machine learning models, they hurt their interpretability.