Unified model for perception inference
During the course of the CEMNET (A unified framework for perceptual inference in sensory cortices) project, scientists developed a unifying model of perception called Constrained Entropy Maximisation Network (CEMNet). This model overcame issues with inaccuracy that previous neural models had when dealing with naturalistic environments and correlated sensory evidence. The CEMNet model incorporates a theoretical and an experimental component. The computational model represents behaviour in different perceptual environments using a biologically plausible neural architecture. To represent the environment, dedicated units encode presence of specific perceptual features and their cunjunction. The constraint units enabled the detection of the most likely scenario when sensory evidence is presented. Scientists also derived plasticity rules to determine connections between neural units. CEMNET researchers collected experimental data on perceptual integration by subjecting human participants to a visual perception task for which correlations were introduced. Using sophisticated machine learning techniques such as expectation-maximisation, they studied how participants integrated correlated sensory evidence. They found that the experimental evidence is in agreement with the CEMNet model. Simulations have demonstrated the ability of the CEMNet framework to handle complex environments and accurately represent perceptual inference. A key finding, CEMNET researchers revealed the role of experience in building integrated perceptual categories using sensory evidence for decision making. The CEMNET model could be applied to understand perception and perception distortion in healthy and pathological subjects such as those with schizophrenia.
Keywords
Model, perception inference, CEMNET, Constrained Entropy Maximization Network