This work developed a model of decision confidence that operates directly on naturalistic, high-dimensional inputs, avoiding the need for simplifying assumptions. A performance-optimized neural network model was trained to make decisions from high-dimensional inputs and to estimate confidence by predicting the probability those decisions will be correct. Unsupervised deep learning methods were used to extract a low-dimensional representation of the model’s training data, which displayed key properties that undermined the presumed optimality of the BE model. The model also accounts for a range of neural dissociations between decisions and confidence, including some features akin to blindsight resulting from lesions to the primary visual cortex.
Related Posts
