Visual Learning for a Mid Level Pattern Discrimination Task - eScholarship

CogSci2000 Visual Learning for a Mid Level Pattern Discrimination Task I. Fine ( fine@salk.edu ) Department of Psychology; University of California, San Diego, 9500 Gilman Drive La Jolla, CA 92093-0109, USA Robert A. Jacobs ( robbie@bcs.rochester.edu ) Department of Brain and Cognitive Sciences, University of Rochester Rochester, NY 14627, USA Abstract Our goal was to examine the plasticity of the human visual system at mid to high levels of visual processing. It is well understood that early stages of visual processing contain cells tuned for spatial frequency and orientation. However images of real-world objects contain a wide range of spa- tial frequencies and orientations. We were interested in how different spatial frequencies and orientations are com- bined. We used a pattern discrimination task - observers were asked to discriminate small changes in a “wicker- like” stimulus consisting of six superimposed sinusoidal gratings. Observers were asked to discriminate a 15% spa- tial frequency shift in two of these sinusoidal components, which were masked by four noise components. We found large amounts of perceptual learning for this task – over eight sessions of training observers’ average percent cor- rect increased by 31%, corresponding to their thresholds dropping to a third of their initial values. Further experi- ments suggest that learning was based on changes within a mid level stage of processing intermediate between low- level analyzers tuned for orientation and spatial frequency and high-level pattern matching or object tuned cells. This mid level stage seems to be “very roughly Fourier” and combines information from individual gratings using prob- ability summation. This stage of processing is also re- markably plastic compared to earlier stages of processing. Introduction A great deal is known about low level visual pattern analyzers and their role in visual perception. At early stages of processing retinal input is represented by low level analyzers tuned for spatial frequency and orientation with receptive fields of limited spatial extent - properties very similar to simple cells in V1 (see Graham, 1989 for a review). However images of real-world objects contain a wide range of Fourier components, and therefore the combination of information across these low level ana- lyzers is necessary to reliably recognize objects. Evidence suggests that there may be mid level mechanisms selec- tively pooling information across low level analyzers tuned for a wide range of spatial frequencies or orienta- tions (e.g. Georgeson, 1992; Derrington & Henning, 1989; Burr & Morrone, 1994; Graham & Sutter, 1998; Olzak & Thomas, 1999). It has been argued that relatively early stages of the visual system (V1) change with training (e.g. Ball & Sekuler, 1987; Fahle & Edelman, 1992; Sagi & Tanne, 1994; Ahissar & Hochstein, 1995,1996; Saarinen & Levi, 1995; Fahle & Morgan, 1996; Schoups & Orban, 1995). In addition, some learning effects have been noted (Ol- zak, personal communication, 1995; Fiorentini & Berardi, 1981) for tasks involving compound grating discrimina- tions thought to involve mid level mechanisms. The following experiments provide support for the ex- istence of mid-level mechanisms pooling over analyzers tuned for spatial frequency and orientation. These mid level mechanisms are shown to be far more adaptable as a function of experience than low level analyzers. Experiment 1 The purpose of Experiment 1 was to measure learning for a complex “wicker” stimulus that required observers to combine information over a wide range of spatial fre- quencies and orientations. Methods stimulus mask stimulus mask time stimulus mask stimulus mask Figure 1: Diagram of the task used in the experiment. Five observers were asked to perform a four alternative forced choice discrimination task (Figure 1). Four stimuli were presented sequentially in time. A two-dimensional white noise pattern was presented after each stimulus to reduce afterimage interference. Observers were asked to