Pavlovian Blindsight and Masked Conditioning: A Neural Network Approach

Blindsight, the ability to respond to supraliminal visual stimuli despite extensive primary visual (V1) damage and reports of them as unseen, is central to theories of phenomenal consciousness. However, two related phenomena have not yet been considered in any theory: Pavlovian blindsight and masked conditioning. I propose a unified theoretical account of both by applying an existing neural network model of conditioning to the influential two-stream hypothesis of visual organization. On this hypothesis, V1 projects to the posterior parietal (dorsal) and inferotemporal cortices (ventral). The former also receives projections from the pulvinar, which allows for a V1-independent visual control, characteristic of blindsight. Artificial neural networks were designed after this hypothesis and according to the model. The model was originally proposed as a unified connectionist account of Pavlovian and operant conditioning, but the focus in this paper is on Pavlovian conditioning, where reinforcement (the occurrence of the unconditioned stimulus) is independent of responding. The model’s learning rule combines Hebbian learning with a form of reinforcement learning inspired by the roles of hippocampal and dopaminergic systems in both kinds of conditioning. This signal takes the form of a diffuse temporal difference in the activations of neurocomputational units that simulate hippocampal and dopaminergic systems. A network with pulvinar and without V1 inputs simulated Pavlovian blindsight. A network with both inputs and a larger ventral subnetwork simulated masked conditioning. Implications for the understanding of blindsight are discussed.