Neural Correlates of Learning of Complex Movement Recognition

Abstract

The recognition of complex movements is a fundamental function of our everyday life. Results from object recognition, biological motion processing and theoretical modeling indicate that learning could play an important role in shaping the processing of complex movements in the human brain. The aim of this thesis is to systematically investigate visual learning mechanisms, focusing on three main questions: 1. Are humans able to learn to discriminate between very similar complex movement patterns and what are the possible constraints that influence the learning process? 2. What are the neural correlates of the learning process in the different areas of the visual cortex and is there a difference between the learning of human like movements compared to artificial articulated movement patterns at the level of the BOLD response? 3. Is it be possible to implement biologically plausible learning mechanisms into an already existing model for biological movement recognition and can the model be used to simulate the BOLD activity changes obtained from the functional imaging experiments in order to test different hypothesis about the underlying plasticity mechanisms? According to the three main research questions, the thesis is divided into three experimental chapters. Chapter 3 reports a series of psychophysical experiments that investigated whether humans are able to learn to differentiate between complex movement patterns. These patterns belonged to three different groups of movements and were all presented as point-light animations. The first group was composed of natural human movements, the second one consists of movements of artificial skeleton models containing nine segments moving in an articulated fashion and the third one consisted of the same human movements as the first one, but this time the spatial positions of the individual points were scrambled. The main focus of the psychophysical investigations was to identify possible differences in the learning process between the three different groups with respect to the time scale of the learning and the invariance properties of the learned representation. To control for the learning history of the subjects, all stimuli were generated by motion morphing. In this way it was possible to create stimuli that were completely novel to the observer. Additionally, motion morphing by linear combination of prototypical movements allowed to precisely control the spatio-temporal similarity between the individual stimuli. Chapter 4 presents the results of a series of functional imaging experiments that were carried out to identify possible neural correlates of the learning process. Because the stimuli across the conditions that had to be compared were very similar, a special fMRI adaptation paradigm was used to acquire the images. This technique allows to identify possible sub-populations of neurons within the same voxel that contribute to the encoding of different movement stimuli. In addition to the experimental runs, several localizer runs were acquired for every observer to reliably detect the visual areas involved in low-level, mid-level and high-level motion and form processing (namely V1, V2, V3, V3a, VP, V4v, V3b/KO, hMT+/V5, FFA and STSp). By analyzing the fMRI signal separately for each of these areas, it became possible to identify learning processes at all stages of the visual hierarchy. The main focus of the imaging experiments was to pinpoint learning induced neural plasticity mechanisms in the visual cortex and to identify possible differences between the learning of natural human-like movements compared to artificial articulated movement patterns. The final experimental chapter of this thesis deals with the theoretical implementation of the experimental results. The theoretical part was based on an already existing neural model for biological motion recognition, which was extended by the implementation of biologically plausible learning rules. Specific detectors of complex form and optic flow fields were learned automatically along with the temporal order with which these features arise during the movement sequence. Additionally, a neural adaptation mechanism was implemented at the highest level of the model. The goal of the theoretical part was to be able to simulate a whole run of a real fMRI experiment and to determine whether the simulated BOLD responses are in accordance with measured BOLD responses. In the future, this model could then be used to test different hypothesis about how learning shapes the processing of complex movements in the human brain.

Die vorliegende Arbeit beschäftigt sich mit der visuellen Wahrnehmung komplexer biologischer Bewegungen. Vorangegangene Arbeiten legen die Vermutung nahe, dass Lernprozesse bei der Verarbeitung von biologischen Bewegungen von entscheidender Bedeutung sind. Die Dissertation bearbeitet im wesentlichen drei verschiedene Fragestellungen: 1. Können Menschen lernen zwischen sehr ähnlichen komplexen biologischen Bewegungen zu unterscheiden? Gibt es Unterschiede zwischen dem Lernen von menschlichen Bewegungen und artifiziellen Bewegungsmustern? 2. Können wir mit Hilfe von funktioneller Magnetresonanztomographie Veränderungen auf neuronaler Ebene nachweisen, die durch den Lernprozess ausgelöst werden? Sind die gleichen Hirnareale beim Lernen von natuerlichen menschlichen Bewegungen und artifiziellen Bewegungsmustern beteiligt? 3. Können wir ein neuronales Netzwerk entwickeln, mit dem die beobachteten Lernprozesse simuliert werden, um die neuronalen Mechanismen besser zu verstehen? Diese Fragestellungen werden mit Hilfe von psychophysikalischen Messmethoden, funktioneller Bildgebung und theoretischer Modellierung genauer untersucht.