Purchase this article with an account.
BS Tjan, ST L Chung, DM Levi; Limitation of Ideal-Observer Analysis in Understanding Perceptual Learning . Invest. Ophthalmol. Vis. Sci. 2002;43(13):2916.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: Performance for a variety of visual tasks improves with practice. Various attempts have been made to determine the functional nature of perceptual learning. Previously, using an ideal-observer analysis (similar to Gold et al., 1999, Nature), we reported that the improvements in identifying letters in peripheral vision following 6 days of training was due to an increase in sampling efficiency, but not a reduction in intrinsic noise (Chung et al., 2001, pre-OSA). Here we reexamine the functional nature of perceptual learning using the perceptual template model (PTM, Lu & Dosher, 1999, JOSA-A). Methods: We used the data set collected by Chung et al (2001) in which the change in contrast thresholds for identifying single letters embedded in external noise were tracked over a period of 6 days. Six levels of external noise were tested each day. Data were collected using the Method of Constant Stimuli. Thresholds at d'=0.8, 1.7, 2.7 were determined per noise level by fitting a psychometric function to the raw data. This resulted in three threshold-vs-noise-contrast (TvC) functions for each day and each observer. Results: Fitting PTM to the TvC functions revealed that the performance improvements observed in 4 of 5 observers were due to a reduction in internal additive noise. Three of these 4 observers also showed a template retuning. No significant change in the internal multiplicative noise was observed. (The fifth observer did not show any significant change in performance.) Conclusion: Despite individual differences, both internal noise reduction and template retuning appear to be the common mechanisms of perceptual learning in peripheral vision. This interpretation of the results contrasts sharply with the one obtained with an ideal-observer analysis. The discrepancy is due to the omissions of non-linearity and a stimulus-driven stochastic component (multiplicative noise or uncertainty) in the linear ideal-observer model.
This PDF is available to Subscribers Only