Emergence of associative learning in a neuromorphic inference network

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Artificial Intelligence

Background:

• Predictive coding and active inference theories propose the brain uses generative models to predict sensory input and minimize prediction errors.
• Previous applications of these theories modeled neural networks at a mesoscopic scale.
• The validity of modeling individual neurons as inferring agents within a biologically plausible architecture remained unexplored.

Purpose of the Study:

• To explore the validity of modeling neurons as inferring agents in a biologically plausible architecture using predictive coding and active inference.
• To simulate associative learning in a simplified cerebellar circuit.
• To implement the model in low-power hardware for neuromorphic applications.

Main Methods:

• A simplified cerebellar circuit was modeled with individual neurons acting as Bayesian agents.
• The model simulated the classical delayed eyeblink conditioning protocol.
• Neurons and synapses minimized prediction error, serving as the network's cost function, and the network was implemented on a low-power microcontroller.

Main Results:

• Persistent changes in synaptic strength, mirroring neurophysiological observations, emerged through local prediction error minimization.
• The model successfully demonstrated associative learning.
• The hardware implementation showed remarkably efficient performance in unsupervised learning tasks compared to conventional neuromorphic architectures.

Conclusions:

• An ensemble of neurons minimizing free energy within a biologically plausible architecture can replicate natural self-organization, such as associative plasticity.
• Neuromorphic networks of inference units can learn unsupervised tasks without pre-programmed learning rules.
• This approach offers a potential pathway toward novel brain-inspired artificial intelligence.