High-Level Vision in Rodents

The human visual system is able to recognize objects despite tremendous variation in their appearance due to changes in size, position, viewpoint, illumination, etc. This ability, referred to as invariant object recognition, requires the construction of representations that are tolerant to these identity-preserving transformations. This forms the basis of higher visual processing, but its neural underpinnings are poorly understood. Rodents have been the animal model of choice for many research topics, e.g. learning, memory, spatial navigation, drug development, and all have experienced great advances thanks to methodological innovations like two-photon calcium imaging, optogenetics and genetic manipulation most often applied in rodents. This dissertation explores the possibility to extend the use of rodents to studies of high-level vision, where traditionally nonhuman primates (e.g. macaques) are the model of choice. Rodents have been largely overlooked because their brains are often assumed to lack advanced visual processing machinery. However, recent studies have shown a surprising degree of functional specialization in at least nine visual areas that surround rodent primary visual cortex (V1). At the same time carefully designed behavioral studies have indicated that they are capable to solve complex visual tasks. These arguments are further discussed in Chapter 1, together with an overview of the main concepts and state-of-the-art in the field of high-level vision. In Chapter 2, we investigated how simple, parametric stimuli are represented by the rat visual system. After being trained on a simple discrimination between two stimuli ( prototypes ), they effortlessly generalized to new stimuli that varied quite a lot in the orthogonal direction in stimulus space. It did not matter to the rats whether the category boundary between classes was aligned with the dimensions of the stimulus space (orientation and spatial frequency). A markedly different result is obtained when humans perform a similar task. We conclude that rats are simply computing the similarity between each of the newly presented stimuli and the prototypes they had learned in a previous phase, a strategy that is sometimes more optimal than the rule-based strategy used by human participants. Chapter 3 is dedicated to the shape discrimination experiment that was published in Current Biology (Vermaercke & Op de Beeck, 2012). One of the main achievements of this study is that we successfully applied a multivariate paradigm called Bubbles to behavioral research in rats. The Bubbles paradigm has mostly been used in humans and monkeys. It involves the partial occlusion of stimuli used in different trials. By applying reverse correlation to correct and incorrect trials, this technique allows the visualization of the crucial information that drives task performance. We used this paradigm in a discrimination task that required rats to tolerate changes in stimulus position and were able to pinpoint the underlying strategy used by rats. We performed a similar experiment with human subjects, which allowed us to directly compare both species. We concluded that rats were using a less complicated, but still a quite flexible strategy. In Chapter 4, we recorded single unit responses for a stimulus set of six shapes. Using the responses of populations of neurons in five areas along the cortical what pathway, we could show how the representation of these shapes changed gradually. By comparing responses to shapes at two positions within each cell s receptive field, we noticed that populations in higher areas are more tolerant to these position changes. This could be considered a neural substrate for solving the task of Chapter 3. We also compared the neuronal similarities between shapes (related to the representational space containing these shapes) in different areas to behavioral discrimination performance and found good correlations in higher areas, whereas V1 responses were only correlated to physical differences between the shapes. Chapter 5 summarizes the main results of the studies presented in this dissertation and finally we conclude that rodents are a promising model of high-level vision, if their limitations are taken into account. This conclusion is also supported by recent studies that have tried to answer the same question, which clearly indicates the great importance of this work. Indeed, being able to investigate (rudimentary) forms of high-level vision in an animal model that allows the use of new techniques to study neural processes at the system level, will be a huge step towards understanding the computational underpinnings of complex visual processing in the human brain.nrpages: 153status: publishe