The difference in the perception of the Necker cube by autistic and neurotypical observers with lower reversal rates and reduced from-above bias in autistic individuals has been reported. Prior literature used a multi-parameter connectionist model for the Necker cube perception in autistic observers. We take a more delimited approach by examining whether this behavioural pattern of reversal rates, dwell-time bias, and undefined percepts can be reproduced through variation of neural noise alone in a simple Hopfield attractor network. When this Hopfield network is trained on the patterns that are a vectorization of the Necker cube image, it is observed that varying only the temperature (noise parameter) is insufficient to reproduce these results since with an increase in temperature, the dwell time bias reduces while reversal rates increase and vice versa for lower temperature. Hence the attractor asymmetry (parametrised by gamma) and temperature need to be simultaneously varied to reproduce that pattern. The undefined reports experimentally found are treated as states closer to the ambiguous input than to either dominant percept in this Hopfield model. The input vector representing the Necker Cube was manually verified. To help in the coding for the modified Hopfield Network, Gemini was used in the Colab notebook environment.
Panda et al. (Tue,) studied this question.