The neuroscientific study of mammalian vision has yielded important achievements in the last decades, but a thorough understanding is still lacking at anatomical and functional levels. From an operational perspective, this understanding would amount to creating an artificial system that reaches the performance and versatility of human vision. A first step to reach this goal requires understanding what neuroscientists call core object recognition, i.e., the rapid and largely feed-forward processing of visual information that mediates the identification and categorization of objects undergoing various identity-preserving transformations (DiCarlo et al., 2012). Electrophysiological experiments on primates have revealed that populations of neurons along the so-called ventral stream – a succession of areas running from the occipital to the temporal cortex – support core recognition, thanks to two key properties: along the pathway, neuronal responses become increasingly more selective to objects identities and increasingly more invariant to their transformations. With similar goals in mind, albeit with an engineering focus, the machine learning field has developed artificial neural networks, with feed-forward multi-layered architectures, that reach human-level performance in various object recognition tasks. Yet these artificial networks are only loosely inspired by the biological ones, they are mainly trained with supervised techniques, and perform on static images, so they fall short at providing a model for the understanding of the mammalian visual system (Kriegeskorte, 2015). Electrophysiological investigations therefore remain an important aspect in vision research. Primates are the closest species to humans, but conducting research with them became more and more difficult (for practical and ethical reasons). Over the past decade, rodents have been used as complementary models to monkeys in the study of visual processing, as they are smaller, reproduce faster, and, importantly, are more suitable for a large batch of experimental techniques. Recent physiological studies in the mouse and the rat have described successions of areas that resemble the visual pathways found in the monkey (Niell, 2011; Vermaercke et al., 2014; Tafazoli et al., 2017), while behavioral studies have shown that the rat visual system is capable of sophisticated object recognition (Zoccolan et al., 2009; Alemi-Neissi et al., 2013). Until recently, most of the vision research has focused on simple, static and parametric artificial stimuli (e.g. bars), leaving the representations of time-varying natural images (i.e. movies) little explored. Natural images are those we see during every-day life: they are characterized by high spatial and temporal correlations, i.e. they contain well-defined structures that have similar color intensities over extended areas, and that remain present in the scene over long intervals (for example, a tree trunk is all brown and doesn’t disappear from moment to moment). Some have argued though that natural images are too complex and still poorly understood to allow well-controlled hypothesis-driven experiments (Rust and Movshon, 2005); others have instead stressed the fact that organisms have evolved within a natural environment and they must have adapted to process natural images in the most efficient way, hence the requirement to use this type of stimuli in vision studies (Barlow, 1961; Felsen and Dan, 2005). The theory that formalizes this hypothesis is named “efficient coding”. The aim of this PhD project is to investigate whether we find evidence in support of the theory in the visual cortex of rats. Specifically, we are addressing two important predictions: 1) that neural responses are increasingly persistent in time, which amounts to measuring if neurons across different areas fluctuate at different timescales in response to the same input (which would be a sign of invariance), and 2) that response distributions successively become sparser (a sign of selectivity). We recorded the neuronal activity in four rat visual areas: in order from the most medial to the most lateral, V1, LM, LI and LL. The results we found are described in two chapters. In the first one, “Representation of natural movies in rat visual cortex”, we observe a tendency towards an increase of slowness estimated with two different measures, and a decrease of sparseness across the four areas. In the second one, “Population decoding”, we are implementing a population decoding technique and show that LL neurons are better than those from other areas at maintaining a self-similar object representation over time. In the last chapter of the thesis we discuss possible implications of our findings.
|Titolo:||Representation of natural movies in rat visual cortex|
|Relatore/i esterni:||Examiner: Balasubramanian, Vijay; Examiner: Wallace, Damian|
|Data di pubblicazione:||22-gen-2018|
|Appare nelle tipologie:||8.1 PhD thesis|