Purchase this article with an account.
Stefan Ringbauer, Heiko Neumann; Perceptual Learning without Awareness: A Motion Pattern Gated Reinforcement Learner. Journal of Vision 2011;11(11):977. doi: 10.1167/11.11.977.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Problem. Perceptual learning increases the performance of detecting motion directions even if the observer is not aware of the presented motion direction (Watanabe et al., Nature 2001). The results suggest that training with a sub-threshold stimulus affects the neural sensitivity and leads to an increase of decision performance for stimuli above threshold. Which mechanisms of cortical motion processing are involved and how do neural mechanisms of learning account for this achievement?
Method. We propose a neural model of visual motion and motion pattern detection (based on Raudies & Neumann, J. of Physiology Paris 2010). Model V1 detects local motion signals that are integrated in model MT. Spatio-temporal configurations of MT responses are further integrated in model MSTd. Feedback signals from MSTd to MT modulate the activity of MT neurons. MSTd responses are temporally integrated in model LIP which generates a decision. The strength of the connections between MSTd neurons and a decision unit in LIP can be adapted using motion pattern gated reinforcement learning (compare Roelfsema et al., Trends in Cognitive Science 2010). Since sub-threshold stimuli are used for training the maximum input activation in LIP drives this adaptation rather than its output activities. Furthermore, the sum of the connection weights is fixed which decreases the influence of signals induced by task-irrelevant features due to the normalization and also keeps the weights in bounds.
Results. Model simulations quantitatively replicate the findings of Watanabe et al. as there is an improvement of detection performance at an average of 10.8% for stimuli above threshold after the training with sub-threshold stimuli. Due to the normalization of the weights of the connections from MSTd to LIP our model predicts that if two, instead of one, motion directions where trained, the improvement on the decision performance will drop about a third to approximately 7%.
This PDF is available to Subscribers Only