Evolving Decomposed Plasticity Rules for Information-Bottlenecked Meta-Learning

Artificial neural networks (ANNs) are typically confined to accomplishing pre-defined tasks by learning a set of static parameters. In contrast, biological neural networks (BNNs) can adapt to various new tasks by continually updating their connection weights based on their observations, which is aligned with the paradigm of learning effective learning rules in addition to static parameters, e.g., meta-learning. Among broad classes of biologically inspired learning rules, Hebbian plasticity updates the neural network weights using local signals without the guide of an explicit target function, closely simulating the learning of BNNs. However, typical plastic ANNs using large-scale meta-parameters violate the nature of the genomics bottleneck and deteriorate the generalization capacity. This work proposes a new learning paradigm decomposing those connection-dependent plasticity rules into neuron-dependent rules thus accommodating $O(n^2)$ learnable parameters with only $O(n)$ meta-parameters. The decomposed plasticity, along with different types of neural modulation, are applied to a recursive neural network starting from scratch to adapt to different tasks. Our algorithms are tested in challenging random 2D maze environments, where the agents have to use their past experiences to improve their performance without any explicit objective function and human intervention, namely learning by interacting. The results show that rules satisfying the genomics bottleneck adapt to out-of-distribution tasks better than previous model-based and plasticity-based meta-learning with verbose meta-parameters