Raphaël Holca-Lamarre, BCCN Berlin / TU Berlin

Learning Representations with Neuromodulators

Biological and artificial neural networks (ANNs) represent input signals as patterns of neural activity. In ANNs, the most widespread method to learn such neural representations is the error-backpropagation algorithm. Although functionally effective, this algorithm bears its limitations and is also unlikely to be implemented in the cortex. In biology, neuromodulators appear to be involved in learning representations. For instance, pairing a stimulus with the release of either acetylcholine (ACh) or dopamine (DA) evokes long lasting increases in the responses of neurons to the paired stimulus. The functional roles of ACh and DA in shaping representations remain largely unknown. In this thesis defence, I present my doctoral work on neuromodulator-based learning in neural networks. The aims of this work are two-fold: first, to gain a functional understanding of ACh and DA transmission in shaping biological representations and, second, to explore neuromodulator-inspired learning rules for ANNs. To address these questions, I model the plastic effects of ACh and DA in a Hebbian-learning network and confirm that stimuli coinciding with greater neuromodulator activation are over-represented in the network. I then simulate the physiological release profiles of ACh and DA. I measure the impact of the neuromodulators on the network's representation and on its performance on a classification task. The results indicate that ACh and DA trigger distinct changes in neural representations that both improve performance. The putative ACh signal redistributes neural preferences so that more neurons encode stimulus classes that are challenging for the network. The putative DA signal adapts synaptic weights so that they better match the classes of the task at hand. The model thus offers a functional explanation for the effects of ACh and DA on cortical representations. Additionally, the learning algorithm yields performances comparable to those of state-of-the-art optimisation methods in multi-layer perceptrons while requiring weaker supervision signals and interacting with synaptically-local weight updates.

Additional Information

PhD defence in the research training group GRK 1589 "Sensory Computation in Neural Systems"

Organized by

Klaus Obermayer / Robert Martin