Abstract:
The pursuit of a thorough understanding of human visual perception and sensation has occupied many great minds for centuries. In recent years, however, Vision Neuroscience has seen a surge in advancements due to the mutually beneficial symbiosis with the field of Machine Learning. Especially the adaptation of Artificial Neural Networks (ANNs) as system identification models has proven to be of great value for modeling the complex computations that are performed by cortical neurons for visual processing. Such system identification models using ANNs are now the state-of-the-art for predicting neural activity as a response to natural stimuli. Additionally, these models have proven exceptionally useful for traditional biological Neuroscience as well: Experiments whose search space would be unfeasible to investigate directly in an animal (in-vivo) can be first run on a system identification model instead (in-silico). The models can then help to shrink the search space in-silico to a size which is more manageable for the actual in-vivo experiment. This thesis builds on the latest advancements of ANN system identification models of mouse visual cortex for neural prediction and in-silico experiments. It contributes to the knowledge of the field in three major directions: It first introduces a novel ANN architecture which is exceptionally data-efficient and which can cleverly make use of retinotopy by incorporating anatomical information additionally to the functional data. Combined with a novel method of matching and tracking neurons over separate experimental sessions, the resulting model sets a new state-of-the-art for neural response prediction and, more importantly, yields features that generalize between animals. Since such models, however, only capture a part of the full neural response distribution while many normative theories require knowledge of the full distribution, the second contribution of this thesis sets the stage for full-likelihood models: By mathematically deriving the full posterior predictive distribution of the most popular distributions in Neuroscience, the upper bounds of a likelihood-based performance metric are obtained. This metric makes it possible to accurately evaluate full-likelihood models and, more precisely, get insights into exactly which parts of the model need the most attention for improvement. In the final contribution of the thesis the framework for full-likelihood models and the generalizing state-of-the-art features are then used to obtain neuroscientific insights: The analysis of the variability of neural firing. To this end, novel types of stimuli that aim at driving or suppressing the variance conditioned on a stimulus are generated. The results indicate that the variance of the firing rate might be affected by the size of the stimulus, much like known center-surround effects for the mean firing rate.
