Abstract
Perceptual learning is a long-lasting perceptual sensitization after extensive training. Because the learning is specific for a trained eye, feature or position, greater plastic changes may occur in lower visual cortex. However, how an external performance feedback during the training is utilized for this sensory learning remains unclear. If it works as a supervised signal, its information quality is most important. By contrast, if it is used for a plasticity control, a feedback that boots perceptual learning can be made even when it does not reflect subjects' actual performance. Here, using the feedback manipulation procedure, we show that subjects implicitly evaluate the statistical characteristics of the performance feedback and utilize it for the plasticity control. Subjects performed a grating discrimination task 40 times within each block. After each block, the feedback was presented to subjects. Subjects showed a significant sensitivity increase during 30 blocks when the block feedback reflected actual subjects' accuracy (control condition). This learning tendency was well-explained by a linear regression (basic learning tendency). In the feedback manipulation condition, the block feedback followed the basic learning tendency with Gaussian noise and did not reflect subjects' actual performance. When we made the gradient of the feedback larger than that of the basic learning tendency, subjects' sensitivity increase was significantly greater compare to the control condition. Surprisingly, when we made the variance of the Gaussian noise of the feedback smaller, the learning was boosted more. We confirmed that subjects were not aware of the feedback manipulation. We further showed that these behavioral results can be interpreted by the optimal statistical framework using Kalman filter estimation. We argue that, if subjects' learning dynamics is identified, we can make the feedback that maximizes the perceptual learning even when it does not reflect subjects' actual performance.
We thank to Yuka Furukawa for her support with data collecting. This research was supported in part by the National Institute of Information and Communication Technology (NICT-KARC) and Human Frontier Science Program (HFSP) to M.K.