Semantic Anchors Facilitate Task Encoding in Continual Learning

Area of Science:

• Cognitive Science
• Neuroscience
• Artificial Intelligence

Background:

• Human learning efficiently integrates prior knowledge.
• Traditional task learning research often isolates abstract rules, ignoring semantic context.
• Semantic knowledge and labels may facilitate new task acquisition.

Purpose of the Study:

• To investigate if semantically rich task embeddings and labels improve novel task learning.
• To assess the impact on rule encoding, forgetting, and interference.
• To explore the underlying mechanisms using decision-making tasks and computational modeling.

Main Methods:

• Experiments involving novel task learning with varying semantic richness of stimuli and labels.
• Value-based decision-making tasks and reinforcement learning modeling.
• Artificial recurrent neural networks fitted to human performance data.

Main Results:

• Semantically rich settings and labels reduced task forgetting.
• This benefit persisted across different label types (pictorial, words) and compared to meaningless labels.
• Semantic embeddings facilitated efficient, feature-specific processing and improved task separation in models.

Conclusions:

• Semantically rich task rules and labels enhance novel task learning and robustness.
• This approach offers insights into human continual learning advantages over artificial agents.
• Utilizing semantic context is crucial for efficient and stable learning systems.