Abstracting Spatial Prototypes through Short-term Suppression of Hebbian Weights in a Continuously Changing Environment

A step towards assumption-free self-organization is proposed. We address the problem of learning a stable representation of the environment from inputs continuously changing at an unpredictable rate. The vulnerability of competitive Hebbian learning to low rate changes is assessed. It is shown that anti-Hebbian suppression of the feed-forward Hebbian weights broadens the range of rates in which learning is possible, and reduces the innuence of the rate on the emerging representation. The resulting robustness during real-time training is demonstrated through simulations, and is compared to an alternative non-synaptic suppression scheme. Some particular passive short-term response properties of high-level visual areas are pointed out as the biological clues for this form of short-term plasticity. The question is raised about a possible stabilizing role of the proposed mechanism in the learning of invariances.