Reference : Rhythms in Neuromorphic Reinforcement Learning
Scientific congresses and symposiums : Poster
Engineering, computing & technology : Multidisciplinary, general & others
Rhythms in Neuromorphic Reinforcement Learning
Dethier, Julie mailto [Université de Liège - ULiège > Dép. d'électric., électron. et informat. (Inst.Montefiore) > Systèmes et modélisation >]
Ernst, Damien [ > > ]
Sepulchre, Rodolphe [ > > ]
International Conference on Brain Dynamics and Decision Making
27th of May 2012 - 31st of May 2012
[en] Living organisms are able to successfully perform challenging tasks such as perception, classification, association, and control. In hope for similar successes in artificial systems, neuromorphic engineering uses neurophysiological models of information processing in biological systems to emulate their functions. Our focus lies in the basal ganglia (BG) and specifically on their involvement in action selection and reinforcement learning (RL).
The BG are a group of interconnected subcortical nuclei that participate in cortical-­‐ and sub-­‐cortical loops for limbic, associative, and sensorimotor functions. These loops are topographically organized in relatively discrete channels that loop back, via appropriate thalamic relays, to the same area of cortex they originated from.
The action selection mechanism comes directly from the BG architecture: the parallel channels compete for behavioral resources, conveying phasic excitatory signals–bids for selection–to the input nuclei. Through comparison of input magnitudes, the tonic inhibitory output is withdrawn from selected channels and maintained or increased on non-­‐selected channels, releasing or preventing action, respectively. This action selection model can be exploited in Cognitive Pattern Generators, analogue to the motor system's Central Pattern Generators, rhythm generators that operate to organize cognition.
The BG play also a critical role in reward and RL circuits. Phasic firing in midbrain dopaminergic neurons complies with the reward prediction error signal of contemporary learning theories. This mechanism could explain cognitive functions, e.g. conditioning memory, and dysfunctions, e.g. Parkinson’s and schizophrenia.
Modeling rhythms at the neurocellular level could introduce the rhythmic component required at the network level for both action selection and RL. The first step in this project is the modeling of the BG and their parallel processing loops with this rhythmic component, a subject of ongoing research.

There is no file associated with this reference.

Bookmark and Share SFX Query

All documents in ORBi are protected by a user license.