Adaptive learning with neuromodulation-aware spiking neural networks
Authors: Alejandro Rodriguez-Garcia, Srikanth Ramaswamy
Presenting author: Alejandro Rodriguez-Garcia
Presentation type: Flash talk at SNUFA 2024 online workshop (5-6 Nov 2024)
Abstract
Current implementations of artificial neural networks (ANNs) are limited in their inability to integrate the brain’s functional and morphological diversity, which hinders their capacity for continual and adaptable learning. In contrast, spiking neural networks (SNNs) with heterogeneous neuron populations can generate different modes of neuronal dynamics, making them inherently flexible and allowing them to adapt to task demands. In biological neural networks, this adaptation is partly driven by the release of neuromodulators. Neuromodulators exert influence on neuronal firing patterns and synaptic plasticity, governing network activity to adapt to changing contextual conditions. By fine-tuning neuronal activity and synaptic strength, neuromodulators enable the seamless transition of network activity between functional states, enhancing continual learning. In this work, we introduce a framework for neuromodulation-aware SNNs, where neuron-specific neuromodulation is integrated into heterogeneous SNNs. By leveraging neuromodulators to dynamically shift between different modes of neuronal activity, we demonstrate how these networks become adaptable to context-aware signals. This approach bridges the gap between biological and artificial learning systems, highlighting the potential of neuromodulation to drive more adaptable AI models capable of operating in complex and dynamic environments.