Eintrag in der Universitätsbibliographie der TU Chemnitz
Volltext zugänglich unter
URN: urn:nbn:de:bsz:ch1-qucosa2-969147
Larisch, René
Hamker, Fred H. ; Triesch, Jochen (Gutachter)
Die Auswirkung der Plastizität hemmender Synapses und der entstehenden Netzwerkdynamik auf die Verarbeitung visueller Informationen
The effects of inhibitory plasticity and the emerging network dynamics on processing visual information
Kurzfassung in englisch
To perceive an object, the visual system must process both spatial information about the object (such as shape and texture) and its temporal behavior (such as its moving direction). Simple cells in the primary visual cortex are selective for simple spatial and temporal object properties such as orientation and direction of motion. Their selectivity is thought to be mainly determined by the driving input from cells in the lateral geniculate nucleus, while inhibition of interneurons only sharpens their selectivity. The formation of this selectivity occurs through synaptic plasticity, as has been shown in many experimental and theoretical studies. However, only a few of these studies have investigated whether plasticity at the synapses to and from the inhibitory interneurons is relevant for the formation of selectivity. This leads to the first open question about the relevance of plasticity at the different synapses, especially inhibitory ones. Since the visual system provides information about objects to higher areas along the sensory processing pathway, that object information must be represented within the system. Therefore, in addition to reproducing various experimental phenomena, an artificial visual system should perform well at object identification. This leads to the second open question about the performance of a biologically motivated artificial visual system in object recognition. Furthermore, the selectivity to a stimulus moving direction is strongly connected to the spatiotemporal processing of simple cells. Models that have investigated direction selectivity usually use a purely mathematical description of the neural dynamics without considering how these dynamics can arise within the network dynamics. This leads to the third question of how spatiotemporal processing can only be made possible by network dynamics. Another frequently discussed network mechanism for the origin of direction selectivity is intercortical inhibition. Several publications have discussed whether intercortical inhibition is necessary for the emergence of direction selectivity or whether it only enhances the selectivity caused by different temporal offsets in the thalamocortical connections. This leads to the fourth question about the influence of intercortical inhibition on simple cell direction selectivity. To answer these questions, a network of spiking neurons was used in this work. For the first question, the network included a population of excitatory and inhibitory cells. The plasticity of the synapses followed two spike-timing-dependent plasticity models so that the excitatory cells formed a spatial selectivity similar to that of simple cells. By activating and deactivating plasticity at different synapses, it could be shown that plasticity at synapses related to feedforward and feedback inhibition improves the variety of orientation selectivity of the excitatory cells and thus the representation of the input stimulus. Once an optimal representation had emerged in the network, different object recognition datasets were presented to answer the second question. Using the same receptive fields for all datasets, we evaluated how universal the neuronal representation is and where the limits of the current network approach lie. To answer the third question, the trained network was extended by a layer of retinal ganglion cells whose receptive field consisted of the typical center-surround antagonism. By applying a delay to the transmission of stimulus information in the surrounding field, temporal dynamics could be observed for cells in the excitatory population. To answer the fourth question, the selectivity of different inhibitory circuits and the temporal offsets in the thalamocortical input currents were manipulated. The results show that directional selectivity occurs because the temporal offset in the thalamocortical connections breaks the symmetry between the excitatory and inhibitory currents for opposing directions. This work highlights the functional role of inhibitory plasticity for the representation of spatial information as well as for the emergence of direction selectivity through tuned inhibition.
| Universität: | Technische Universität Chemnitz | |
| Institut: | Professur Künstliche Intelligenz | |
| Fakultät: | Fakultät für Informatik | |
| Dokumentart: | Dissertation | |
| Betreuer: | Hamker, Fred H. | |
| DOI: | doi:10.60687/2025-0077 | |
| SWD-Schlagwörter: | Neuronales Netz , Maschinelles Lernen , Objekterkennung , Sehrinde | |
| Freie Schlagwörter (Englisch): | Spiking neural network , Visual processing , Inhibitory plasticity | |
| DDC-Sachgruppe: | Datenverarbeitung; Informatik | |
| Sprache: | englisch | |
| Tag der mündlichen Prüfung | 13.03.2025 |
