Exploring Neuromodulation for Dynamic Learning

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Continual learning systems need to adapt to new tasks using limited data.
• Neuromodulation in the central nervous system facilitates dynamic learning across species.

Purpose of the Study:

• To embed neuromodulatory plasticity into dynamic learning architectures for efficient few-shot learning.
• To develop power and area efficient few-shot learning systems inspired by biological neuromodulation.

Main Methods:

• Introduced ModNet, an architecture with a modulatory layer in a random projection framework, enhanced with attention and compartmentalized plasticity.
• Developed a modulatory trace learning rule for deeper networks, utilizing time-dependent traces updated via simple plasticity rules.
• Integrated an inbuilt modulatory unit to regulate learning based on context and internal state, enabling self-modification of weights.

Main Results:

• ModNet architectures demonstrated rapid learning, completing benchmark image classification tasks in just 2 epochs.
• The modulatory trace learning rule achieved 98.8% accuracy on the Omniglot dataset for few-shot image classification.
• Achieved state-of-the-art few-shot learning performance with significantly reduced computational resources (20x fewer trainable parameters).

Conclusions:

• Neuromodulatory plasticity is a viable mechanism for creating efficient and adaptable few-shot learning systems.
• The proposed ModNet and learning rules offer a computationally efficient approach to few-shot image classification.
• This biologically inspired approach advances the development of AI systems capable of rapid, low-data learning.