Short-term plasticity as cause-effect hypothesis testing in distal reward learning.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Sensory-motor signals often exhibit asynchrony, overlaps, and delays, creating ambiguity in identifying causal relationships between stimuli, actions, and rewards.
• Distinguishing true cause-effect relationships from coincidental occurrences typically requires the repetition of reward episodes.

Purpose of the Study:

• To propose a novel plasticity rule that evaluates hypotheses on cause-effect relationships using both short- and long-term synaptic changes.
• To develop a model that preserves existing network topologies when learning from ambiguous information flows.
• To enhance learning by biasing exploration towards actions that have previously preceded rewards.

Main Methods:

• A new plasticity rule is introduced, utilizing transient weights to represent hypotheses about cause-effect relationships.
• Hypotheses are consolidated into long-term memory only if they consistently predict or cause future rewards.
• Learning is improved by directing the exploration of the stimulus-response space.

Main Results:

• The model demonstrates how beliefs can be consolidated in long-term memory under conditions of ambiguous information.
• It offers a potential solution to the long-standing plasticity-stability dilemma in neural networks.
• The study provides an interpretation for the functional role of short-term plasticity in learning.

Conclusions:

• The proposed model effectively addresses the challenge of learning causal relationships from noisy and ambiguous sensory-motor data.
• It offers a mechanism for stable learning in dynamic environments by selectively consolidating information.
• The findings contribute to understanding how biological neural networks might resolve ambiguity and maintain plasticity.