| dc.description.abstracteng | Many neurons in sensory pathways respond selectively to a narrow class of
stimuli such as faces or specific communication calls. At the same time
neural processing is robust to a large degree of natural variablity within
such complex stimulus classes. For instance, face detection must be robust
with respect to a particular hair or eye color and speech processing must
tolerate large differences between female and male vocalizations. It is
hypothesized that such difficult perceptual invariances might be subserved by
populations of neurons that specialize on different substructures within a
sensory object category. However, it is unclear how specialization emerges within a population of neurons during learning. Coordination of individual learning processes between neurons, that would enable a division of the task, has remained biologically challenging. State-of-the-art machine learning approaches that focus mainly on generating diversity to improve classification performance, assume full transparency of all learning processes and therefore provide only limited insights into the mechanisms that lead to specialization in networks of neurons. By contrast, biologically plausible models that are based on majority voting do not result in specialized neural ensembles. The goal of this thesis is to bridge the gap between biological plausibility and high classification performance by means of specialization.
We show that specialization within populations of
sensory neurons can emerge by a weakly selective training algorithm that only relies on a global supervisory feedback signal, the Tagging algorithm. In response to this global feedback, neurons decide individually whether to engage in a learning step based on the confidence of their individual decision. We show that the Tagging algorithm induces specialization within neuronal populations not only for specifically tailored classification problems, but also for a real-world spoken digit detection task. Additionally, we demonstrate that weakly selective learning can be applied to a broad range of neuron models, ranging from perceptrons to spike escape models, and can be combined with other learning concepts such as reinforcement learning. | de |