Computational modeling has been a crucial tool in cognitive science to understand human cognitive functions and impairments in neurocognitive disorders. Convolutional Neural Networks (CNNs) exhibit striking similarities to human visual processing systems for object recognition, making them a powerful tool for studying visual processes. In this study, we examined the neurobiological theories, namely, the Excitation/Inhibition (E/I) Imbalance and Internal Noise (IN) in explaining face recognition challenges in autism spectrum disorder (ASD) using CNNs, and revealed that over-excitation and increased noises in the CNNs led to compromised performance on face recognition and atypical patterns of internal representations of face stimuli. This approach enables systematic comparisons between typical and atypical cognition, offering a theory-driven perspective to investigate cognitive challenges and their neurocognitive mechanisms with a computational approach.