Final Report Summary - LEARN2SEE (Invariant visual object representations in the early postnatal and adult cortex: bridging theory, model and neurobiology)
The goal of the LERAN2SEE project was to understand what neuronal processes underlie visual object recognition in rats, and what learning principles drive the development of such processes during early postnatal development. Our research was based on a variety of experimental and theoretical/computational approaches, including high-throughput behavioral testing (i.e. psychophysics), neurophysiology, histology, rearing of newborn rats in visually-controlled environments, and machine learning tools (such as information theory, pattern classification, convolutional neuronal networks, and generalized linear models). At the behavioral level, we have performed various studies to investigate high-level processing of both shape and motion information in rats engaged in visual discrimination tasks. One study revealed that rats’ proficiency in visual object recognition is accounted for by the complexity of their perceptual strategy. In another study, we have shown that rats are able to correctly discriminate the direction of drifting plaids (i.e. complex patterns, typically used to study high-level motion processing in primates), thus providing the first behavioral evidence, to our knowledge, of the processing of global motion in rodents. In another study, we have compared the perceptual strategies of rats engaged in the recognition of solid gratings using the tactile, the visual or the visuo-tactile modality. At the neurophysiology level, we have performed a number of studies aimed at understanding how shape and motion information is represented in rat primary visual cortex (V1), as well as the progression of higher-order visual cortical areas that run laterally to V1: areas LM, LI and LL. In a first study, we have found evidence that the ability of neuronal representations to successfully support object recognition increases along the areas progression, being maximal in LL, thus providing very compelling evidence about the existence of a rodent object-processing pathway. This conclusion was confirmed by a second study, in which we probed V1 and LL with simpler visual stimuli (e.g. drifting gratings) and found critical differences in the spatiotemporal tuning properties of neurons in these two areas. In yet another study, we have tested how visual neurons recorded from V1 trough LL encode natural movies. In two other studies we have investigated how V1 and LM neurons encode global motion information, finding neurons that respond in a consistent way to the direction plaid and grating stimuli, thus behaving as the “pattern cells” found in primate visual cortex. As mentioned above, all this experimental work was supported by computational approaches for modeling and interpretation of the collected data. In addition, we also performed a purely computational study, where we have shown that the intrinsic dimension of object representations across the layers of deep convolutional networks evolves according to a typical hunchback trend and reaches, in the last hidden layers, very low values that predict very well the classification performance of the network. Finally, we have designed and built a system that allows rearing newborn rats in visually controlled environments, where the statistics of the visual input is under the control of the experimenter. This apparatus allowed testing computational hypotheses about the impact of the environment on the development of visual cortex with unprecedented accuracy and flexibility. Specifically, we performed two neurophysiology studies employing rats reared in such visually controlled environments, where the temporal continuity of their visual experience was destroyed. This revealed that experiencing a temporal continuous visual inputs during postnatal development is essential for the normal development of V1 and LL, which otherwise do not develop some key tuning properties (such as the ability to encode stimulus orientation in a translation-invariant manner) that are crucial to support visual object recognition. Overall, we believe that the array of scientific questions addressed by the LEARN2SEE project have substantially improved our understanding of visual processing in the rat brain and, more in general, have revealed how the development of some key properties of visual object representations critically depend on adaptation to the visual environment during early postnatal life.