Modelling continual learning in humans with Hebbian context gating and exponentially decaying task signals

Area of Science:

• Computational neuroscience
• Artificial intelligence

Background:

• Humans exhibit task-specific interference, performing worse on multiple tasks simultaneously compared to sequential learning.
• Standard deep neural networks show the opposite pattern, often benefiting from simultaneous training.

Purpose of the Study:

• To develop novel computational constraints for artificial neural networks that mimic human sequential learning costs.
• To enable artificial neural networks to learn tasks in sequence without forgetting, addressing the limitations of standard models.

Main Methods:

• Augmented standard stochastic gradient descent with "sluggish" task units and a Hebbian training step.
• Introduced computational constraints inspired by primate prefrontal cortex gating mechanisms.

Main Results:

• "Sluggish" units created a training switch-cost, biasing representations towards joint, context-ignoring ones during interleaved training.
• The Hebbian step fostered orthogonal representations, effectively preventing interference by creating a task-unit gating scheme.

Conclusions:

• The proposed model successfully replicates human performance differences between blocked and interleaved training curricula.
• The model's findings suggest that misestimation of category boundaries contributes to performance variations in human learning.