Hybrid Plasticity: A Biologically Realistic Meta-Learning Algorithm to Increase the Adaptability and Efficiency of Artificial Intelligence Systems

Overview

Current Artificial Intelligence (AI) systems can learn arbitrarily complex patterns but struggle to generalize and adapt to new situations, making them unsuitable for real-world deployment. Therefore, I propose Hybrid Learning (HL), a meta-learning rule that allows AI to learn and remember new concepts. HL is a mathematical rule I derived to mimic prefrontal plasticity, imitating the brain’s information coding and retention process. During training, synapse-specific HL coefficients are learned that allows the network weights to use environmental input to guide convergence to a set of optimal weights as governed by an attractor in the weight-phase-space. I test HL on artificial neural networks (ANN), completing OpenAI reinforcement learning (RL) tasks, and recurrent neural networks (RNN) completing complex memory tasks. HL significantly outperforms many standard RL approaches, including Proximal Policy Optimization and Deep Q-Learning, while simultaneously increasing the adaptability of ANNs significantly more than other meta-learning approaches, including Hebbian Learning. Furthermore, in RNNs, HL significantly increases adaptability and performance while reproducing neurobiological phenomena within network connectivity. HL networks exhibit close mimicry of GABAergic circuits, as they start unorganized but self-organize into inhibitory and excitatory clusters, develop strong inhibitory autapses, and demonstrate high levels of efficiency, parallel processing, and cognitive integration. These findings provide evidence for two neuroscientific theories, the Synaptic Theory of Working Memory and the Global Workspace Theory of Consciousness. Experimentation indicates that HL is a state-of- the-art meta-learning algorithm that increases the performance and generalizability of AI systems while also acting as a tool to study the brain using artificial networks.