Designing Interpretable Image Recognition Models Based on Neuroscience

Recent advancements in AI have enabled models to reach human-level skills in tasks such as image recognition and text generation through mathematical improvements and architectural recombination. However, a major challenge for AI models is the “black-box” problem, meaning it is difficult to understand how the model makes decisions. Therefore, we are motivated to build explainable and interpretable AI from a neuroscience perspective, potentially solving the black-box problem. In the neuroscience domain, Gabor filters can express the orientation selectivity of neurons in the primary visual cortex, thus some studies have explored incorporating Gabor filters into Convolutional Neural Networks (CNNs) to imitate the neural structure of the brain. In this study, we focused on the neural structure of the visual cortex, particularly how neurons in the primary visual cortex construct the orientation selectivity, and aimed to determine if Gabor filters could play a pivotal role in building AI with explainable and interpretable architectures.