A Neural Network Model of Continual Learning with Cognitive Control.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Neuroscience

Background:

• Neural networks face catastrophic forgetting in continual learning, where new data overwrites prior knowledge.
• Humans exhibit effective continual learning and sometimes benefit from blocked trial presentation, unlike standard neural networks.
• Existing research suggests cognitive mechanisms may mitigate forgetting in biological learning systems.

Purpose of the Study:

• To investigate if cognitive control mechanisms can prevent catastrophic forgetting in neural networks.
• To explore the impact of blocked versus interleaved trial presentation on network learning.
• To understand the role of active maintenance and control signal bias in continual learning.

Main Methods:

• Implementing a cognitive control mechanism within artificial neural networks.
• Comparing network performance in blocked versus interleaved learning paradigms.
• Analyzing learned network representations, specifically map-like structures.
• Investigating the influence of control signal bias on learning dynamics.

Main Results:

• Neural networks equipped with cognitive control demonstrated no catastrophic forgetting in blocked learning settings.
• A performance advantage for blocked learning over interleaved learning was observed when active maintenance was biased in the control signal.
• Analysis revealed insights into the network's internal representations and control mechanisms.

Conclusions:

• Cognitive control offers a promising strategy to enhance continual learning in artificial neural networks.
• The findings provide a potential explanation for the observed human advantage in blocked learning tasks.
• A trade-off between memory maintenance and control strength influences learning outcomes.