Neuro-Inspired Plasticity for Biologically Realistic Self-Adaptation of Neural Network Weights

The Prefrontal Cortex is the core of higher level learning and memory. It currently operates much like an AI system, in the sense that its actions are guided via a dopamine based reward function, however there is one critical difference, the PFC has the ability to rewire itself - plasticity. Here, we look to biological studies to find the governing rules of plasticity - competitiveness, memory, and correlation - to create a biologically plausible implementation of plasticity called Hybrid Plasticity. We implement it in continuous time recurrent neural networks (CTRNNS) completing simple working memory tasks. We show that the implementation of plasticity increases the adaptability of the working memory process within networks, while also resulting in a significant decrease in active neurons within the network indicating higher efficiency. We further demonstrate that plasticity results in an increased small world index, indicating high levels of efficiency, parallel processing, and cognitive integration as proposed by Global Workspace Theory. Plasticity CTRNNs are also shown to self organize into brain-like connectivity patterns including inhibitory clusters, excitatory clusters, and inhibitory autapses during their forward pass. Hence, Hybrid Plasticity represents a proof-of-concept solution for bringing forth biologically realistic phenomena within neural networks resulting in increased efficiency and generalizability.