Poster F29, Tuesday, March 27, 8:00-10:00 am, Exhibit Hall C
Fast synchronization and slow synaptic learning as a solution to the stability-plasticity dilemma
Pieter Verbeke1, Tom Verguts1; 1Ghent University
The human ability to adapt to a constantly changing environment is remarkable. This relies on the ability to learn quickly about associations between perceptual, motor, and goal representations. Nevertheless, fast learning in neural networks typically leads to forgetting of older information; this is unfortunate because one would like to retain environmental regularities without disruption from novel information. Thus, there exists a tradeoff between being sufficiently adaptive to novel information (plasticity) while retaining valuable earlier regularities (stability). We propose that the brain deals with this issue by relying on two separate, yet interacting learning mechanisms in the same neural structures. The first, fast learning mechanism implements binding-by-synchronization (Fries, 2015) (sync learning). Here, perceptual, motor, and goal representations are bound together by synchronization of neural firing. The second, slow learning, mechanism corresponds to classical synaptic learning by (reward-modulated) Hebbian learning. To implement this hypothesis, we adapted the Verguts (2017) model and tested it on a reversal learning task. Simulations demonstrated that a model using only synaptic learning could not deal with sudden changes in task rules. A model using only sync learning could flexibly deal with task rule changes, but overwrote earlier learned rules. Combining sync learning with synaptic learning however, allowed the model to deal with task rule changes without overwriting earlier information. Thus, the resulting model combined (fast) plasticity using sync learning with (slow) stability using synaptic learning to address the stability-plasticity dilemma. In addition to solving this computational problem, we compare the model to neurophysiological and –anatomical data.
Topic Area: EXECUTIVE PROCESSES: Goal maintenance & switching