Every day, the countless tasks we carry out during our life require us to answer a key question: "what is it I'm looking at?" Robustly recognize objects in spite of drastic changes in their appearance is a basic requirement of our survival, and a hallmark of the computational capacity of primate neocortex; our visual system is in fact able to solve this problem without any apparent effort and with surprising and unparalleled speed. In order to make this possible the visual cortex must be able to build, at a certain point, a representation of a visual stimulus that is invariant to the so-called identity preserving transformations (i.e. rotations, translations, scaling, background and/or lighting variations). In primates it was shown that this task is accomplished through processing visual information along a dedicated cortical pathway called the ventral stream, originating in V1 and reaching its highest level in the inferotemporal cortex (DiCarlo, Zoccolan, et al., 2012; Rolls, 2012). This stream is made up, at least for what concerns core object recognition, by a series of feedforward modules organized in a hierarchical fashion. These areas sequentially build more and more complex representations, as their neuronal tuning properties change from a preference for very simple stimuli (e.g. oriented edges) in the low-level areas, to a selectivity for complex patterns in the inferotemporal cortex (Desimone et al., 1984; Kobatake and Tanaka, 1994). We are, however, still far from guring out how the ventral stream solves the invariance problem, as the monkey visual system is a dramatically complex machine, composed by more than 35 areas reciprocally connected by hundreds of pathways (Felleman and Van Essen, 1991). For this reason, and over growing concerns about the use of primates in research, rats have been considered with ever increasing interest as models of the mammalian visual system. Although far simpler and with a visual acuity ten times lower than primates, rats allow access to a wider array of experimental techniques (e.g. two photon imaging, calcium imaging, optogenetics), and recent studies have shown that they are able to invariantly recognize visual objects undergoing identity-preserving transformations (for a review, see Zoccolan, 2015). Anatomical (Coogan and Burkhalter, 1993; Sereno and Allman, 1991) and electrophysiological (Tafazoli, Safaai, et al., 2017) studies indicate that along the progression of rat extrastriate areas running lateral to primary visual cortex (V1) low-level information is gradually pruned in favor of representing more complex features (Tafazoli, Safaai, et al., 2017). In this study we sought to further examine the validity of the rat as a model of the mammalian visual system by investigating in deeper detail the representation of visual objects along the rat putative ventral stream, and how it varies across its different areas. To this extent we recorded in-vivo extracellular activity in anesthetized rats passively exposed to a rich stimulus set, designed to explore a wide range of visual features and transformations. We analyzed the response to the stimuli both at the single-unit level, using information theory to characterize the amount of information conveyed by neuron of different areas about features of different complexity, and at the population level, checking whether and how the complexity of the population code changes across areas and how the different features are encoded by the extrastriate areas. Our analysis shows that the information about low-level properties of the stimuli is progressively reduced both at the single neuron and at the population level. Single neuron analyses show that the mutual information between neuronal response and low level stimulus properties (e.g. total luminosity or position) decreases along the putative ventral stream, while the relative weight of the information about mid- and high-level stimulus properties increased, all in accordance with the information conservation principle. At the population level we used two different clustering algorithms (complete linkage hierarchical clustering and K-means) to group the stimuli according to their representation and then compare the result with the properties of the stimulus set at different levels. The results of the complete linkage hierarchical clustering show that in the early stages of visual processing stimuli sharing similar low-level properties get clustered together in the rst nodes (the ones closer to the root) of the hierarchical tree, while the same does not happen in later stages. Then we divided the stimulus set in categories based on properties of different complexity and we compared these categories to the results of the K-means clustering in the different areas using Adjusted Mutual Information, a well-established clustering comparison metric. The comparison yields similar results to the single neuron analyses, with low level areas' populations carrying more information about low level stimulus properties and with information about high level properties, after ltering out the effect of the low-level ones, staying constant across the hierarchy and thus increasing in relative importance. These results provide further evidence that the functional organization of rat lateral extrastriate areas resembles the one of the primate ventral stream, further promoting the viability of rodents as models of high-level visual processing in mammals.
|Titolo:||Functional evidence of hierarchical object processing in rat lateral extrastriate cortex|
|Data di pubblicazione:||31-gen-2018|
|Appare nelle tipologie:||8.1 PhD thesis|