A neurocomputational basis of face recognition changes in ASD: E/I balance, internal noise, and weak neural representations

Individuals with Autism Spectrum Disorder (ASD) are known for their socio-communicative challenges, including face recognition. Despite mounting evidence in behavioral studies, the neurocomputational basis of these challenges remains unclear. Meanwhile, neurobiological theories propose that ASD may arise from an imbalance of excitatory and inhibitory signals (E/I imbalance) or excessive internal noise (IN). However, studies with humans can hardly provide causal evidence. Therefore, this study employed Conventional Neural Network (CNN) models to simulate face recognition in typical populations and ASD based on the claims of I/E imbalance and IN theories. By varying the positive slope in the ReLU activation function (simulating E/I imbalance) and random noises added to the weights (simulating internal noise), we showed that CNN models with non-optimal ReLU slope or noised weights led to poorer performance in face recognition and atypical neural representations of faces. Overall, simulations based on the E/I imbalance theory seem to encompass a broader range of behavioral profiles in ASD. Our approach to using CNN models to test neurobiological theories is highly theory-driven, and our results provided causal evidence to how neurobiological factors could influence face recognition in ASD. This framework could be easily adapted to test in other neurobiological disorders, providing a plausible bridge between neurobiological theories and behavioral and neuroimaging research on humans.