Date Published: February 17, 2004
Publisher: Public Library of Science
Author(s): Gregor Rainer, Han Lee, Nikos K Logothetis
Abstract: One of the most remarkable capabilities of the adult brain is its ability to learn and continuously adapt to an ever-changing environment. While many studies have documented how learning improves the perception and identification of visual stimuli, relatively little is known about how it modifies the underlying neural mechanisms. We trained monkeys to identify natural images that were degraded by interpolation with visual noise. We found that learning led to an improvement in monkeys’ ability to identify these indeterminate visual stimuli. We link this behavioral improvement to a learning-dependent increase in the amount of information communicated by V4 neurons. This increase was mediated by a specific enhancement in neural activity. Our results reveal a mechanism by which learning increases the amount of information that V4 neurons are able to extract from the visual environment. This suggests that V4 plays a key role in resolving indeterminate visual inputs by coordinated interaction between bottom-up and top-down processing streams.
Partial Text: It is well established that learning can have a strong impact on neural responses to visual stimuli in high-level association cortices such as inferior temporal (IT) or prefrontal (PF) cortex, where the activity of single neurons reflects learning in pair association, object identification, or categorization tasks (Sakai and Miyashita 1991; Logothetis et al. 1995; Booth and Rolls 1998; Kobatake et al. 1998; Erickson and Desimone 1999; Rainer and Miller 2000; Freedman et al. 2002; Sigala and Logothetis 2002). In these studies, learning is thought to modify neural activity to represent task-relevant attributes, such as trained views of three dimensional objects (Logothetis et al. 1995) or associations between paired visual stimuli (Sakai and Miyashita 1991; Erickson and Desimone 1999). The learned representations often exhibit invariance for stimulus features such as size (Logothetis et al. 1995), rotation (Booth and Rolls 1998), or stimulus degradation (Rainer and Miller 2000). Similar neural activity to within-category stimuli during categorization (Freedman et al. 2002) can also be thought of as a learning-dependent form of invariance. Several lines of evidence suggest that these learning effects involve synaptic plasticity and thus represent long-lasting modifications to visual association cortices.
We found that learning resulted in significant and robust improvements in monkeys’ ability to identify degraded stimuli. Behavioral performance varied systematically with coherence (Figure 2A). Monkeys performed at chance level (50% correct) when stimuli were presented at 0% coherence and thus contained no task-relevant information. For degraded stimuli (35%–65% correct), monkeys performed significantly better with familiar than with novel stimuli (t-test, p < 0.01). For undegraded stimuli at 100% coherence, the monkeys' performance was near ceiling for both novel and familiar stimuli (92% and 95% respectively; t-test, p = 0.12). Learning-dependent performance improvements for degraded stimuli were highly consistent across stimuli and monkeys. There were in fact no significant differences in the monkeys' performance to each of the familiar stimuli across sessions at all coherence levels (one-way ANOVAs, p > 0.1), and this was also true for novel stimuli. In addition, performance for novel and familiar stimuli did not differ significantly between the two monkeys at any coherence level (t-tests, p > 0.1). Note that the monkeys’ excellent perfor-mance with undegraded novel objects reflects the fact that they have acquired the rule of the DMS task and are thus able to perform it near ceiling with novel stimuli. The timecourse of this learning-dependent difference in performance is shown in Figure 2B. Session 1 represents a session in which a set of four initially novel stimuli is arbitrarily chosen and kept constant in subsequent sessions, thus becoming more and more familiar. Comparing performance for these stimuli with performance of novel stimuli that are randomly chosen in each session reveals that it takes several sessions for the learning effect to appear. Performance averaged across the first five session was similar for novel and familiar stimuli (t-test, p = 0.43). Furthermore, the learning-dependent difference in performance appeared to asymptote after around ten sessions. In summary, learning led to robust improvements in the monkeys’ ability to identify degraded natural images while the monkeys performed near ceiling for novel and familiar undegraded images.
V4 neurons are generally conceptualized as detectors of visual features of intermediate complexity, such as non-Cartesian gratings (Gallant et al. 1996) or contour features (Pasupathy and Connor 1999). We have found that learning does not affect how V4 neurons respond to undegraded natural images, both in terms of mean firing rate and information communicated about these stimuli. This absence of learning-dependent differences suggests that this V4 selectivity for features of intermediate complexity is not modified by learning, at least during the several weeks of training in the adult monkey during our task. Basic response properties of V4 neurons thus appear not be altered by learning, similar to findings in V1 that have found that parameters such as receptive field size or orientation tuning width remain unchanged even after extensive training (Crist et al. 2001).