BIOPOL: Biologically Derived Principles of Online Learning

The BIOPOL project explores how AI systems can learn continuously online, like the human brain, rather than relying on static training datasets. Based on accurate models of biological neurons, this approach enables adaptive, self-updating systems that improve through real-world experience.

The Model

The team’s model incorporates inhibitory-excitatory balance dynamics, local learning rules, and adaptive filters in a successive approximation scheme that adjusts to new inputs in real time. This allows the system to avoid backpropagation and evolve without an offline training phase.

Early Results

Early results using the MNIST benchmark show promising adaptability.

Contribution to Neuromorphic Computing

This line of research contributes to the emerging field of neuromorphic computing: AI that is both efficient and alive to its environment, capable of ongoing learning and adaptation much like biological organisms.