Integrating unsupervised and reinforcement learning in human categorical perception: A computational model

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Categorical perception involves perceptual system tuning during categorization tasks.
• Existing models consider task-independent (unsupervised learning, UL) and task-dependent (reinforcement learning, RL) effects separately.
• The interaction between UL and RL in categorical perception remains understudied.

Purpose of the Study:

• To investigate the interaction between unsupervised learning and reinforcement learning in the emergence of categorical perception.
• To propose and test a neuro-inspired computational architecture integrating UL and RL processes.

Main Methods:

• Developed a system-level computational model with a perceptual component combining UL and RL.
• Tested the model using a categorization task to evaluate performance under different UL/RL balances.

Main Results:

• A balanced integration of UL and RL resulted in optimal categorical perception and task performance.
• Excessive UL led to failure in identifying task-relevant features.
• Excessive RL caused slow initial learning and sub-optimal performance.

Conclusions:

• The interplay between unsupervised and reinforcement learning is crucial for effective categorical perception.
• Model findings align with experimental evidence on extrastriate cortex activation.
• Extreme model cases may explain sensory alterations observed in autistic individuals.