Using behavioral and neural measures to assess training in scene categorization

• ¹ Wright State University, Department of Psychology, United States

Visual analysis of complex real-world scenes is essential to a variety of professional contexts, ranging from defense intelligence (e.g., overhead imagery analysis and target detection) to urban planning. Expertise in recognizing information-rich yet highly variable scenes is putatively achieved through experience, yet little is currently known about how skills in scene recognition are formed and evolve during learning, and what are the underlying neural mechanisms that support the acquisition and deployment of these skills. The goal of the present study is to quantify the behavioral changes associated with the acquisition of scene expertise, elucidate the neural mechanisms underlying the acquisition of scene expertise, and ultimately, to establish the importance of changes in neural scene representations as novel markers of scene expertise. To track down the development of scene expertise and its underlying neural correlates, we conducted a long-term multi-session behavioral training study in which participants were trained in the categorization of visual scenes, and measured changes in neural responses to these scenes over multiple scanning sessions. We constructed a rich stimulus set consisting of high-resolution color images of scenes varying across five dimensions: Viewpoint (aerial/terrestrial), Naturalness (manmade/natural), and three hierarchical categorization levels: Basic, Subordinate, and Exemplar. For example, the category “airports” contained three airport types (large hub, small hub, military), and each airport type contained ten individual images of specific airports. Critically, each individual scene was presented both in an aerial and terrestrial viewpoint to assess generalization across viewpoints. A group of naïve participants were trained to categorize these scenes for a total of 12 hours across a period of three weeks. Each individual training regimen was comprised of six sessions, interspersed with four functional magnetic resonance imaging (fMRI) scanning sessions. Participants trained on half of the stimuli for five sessions, and in the sixth session they viewed the other half of the scenes. We employed two behavioral metrics to assess scene categorization learning: (1) within-set learning (i.e., learning across the five sessions), and (2) generalization (i.e., transfer of learning) to assess the behavioral effects with the acquisition of scene expertise. Learning occurred within the five sessions (evident in a monotonic decrease in reaction times and increase in accuracy), and notably, we also found transfer of learning, as performance in the sixth session was pronouncedly better than performance in the first four training sessions. Thus, performance substantially improved on all measures, validating our training paradigm. This is a clear demonstration that naïve participants can be trained to develop expertise that goes beyond their initial capabilities in scene categorization. Preliminary analysis of the neuroimaging findings revealed that the effects of training in scene categorization can be measured not only by using behavior, but also by examining how scene representations in scene-selective cortex change as a function of learning. We conducted a region of interest (ROI) analysis, examining the magnitude of neural response in scene-selective regions (i.e., Parahippocampal Place Area (PPA) and the Occipital Place Area (OPA) as well as non-scene-selective visual regions, such as the Fusiform Face Area (FFA) and retinotopic cortex (Early Visual Cortex: EVC). Importantly, all functional ROIs were localized using an independent task. Overall, we observed a robust effect of naturalness, in which manmade scenes evoked a greater response than natural scenes within scene-selective region. Notably, increased levels of training yielded viewpoint effects, evident primarily for the natural scenes. Such an interaction suggests that as participants develop experience with the natural scenes, scene selective regions become more selective for viewpoint, specifically for terrestrial scenes. In comparison, the non-scene-selective visual regions did not demonstrate these effects, suggesting the observed effects are specific and were not due to attentional effects. Together, these results suggest that expertise in scene recognition can be trained in the lab and will form the basis for future studies on the neural substrates of scene expertise. These findings highlight the need for deeper understanding of neural circuits and mechanisms underlying expertise in scene recognition. Overcoming this gap will enable the design of neuroscientific-grounded personalized training regimens.