Multi-Timescale Memory Dynamics Extend Task Repertoire in a Reinforcement Learning Network With Attention-Gated...

Area of Science:

• Cognitive Science
• Neuroscience
• Artificial Intelligence

Background:

• Reinforcement learning and memory are crucial for neural network models.
• The Attention-Gated MEmory Tagging (AuGMEnT) model excels at animal learning but struggles with hierarchical tasks.
• Hierarchical tasks require managing information across different timescales.

Purpose of the Study:

• To address the limitations of the AuGMEnT model in hierarchical tasks.
• To develop a novel neural network architecture capable of handling multi-timescale memory demands.
• To improve reinforcement learning models for complex cognitive tasks.

Main Methods:

• Introduction of a hybrid AuGMEnT network incorporating both leaky (short-term) and non-leaky (long-term) memory units.
• Designing memory units that facilitate low-level information exchange while preserving high-level information.
• Evaluating the hybrid AuGMEnT on cognitive reference tasks: sequence prediction and 12AX.

Main Results:

• The hybrid AuGMEnT network demonstrates improved performance on hierarchical tasks compared to the original AuGMEnT.
• The model successfully distinguishes and manages information requiring different memory retention periods.
• Effective handling of both short-term and long-term memory components within a single architecture.

Conclusions:

• The hybrid AuGMEnT architecture offers a viable solution for reinforcement learning models facing multi-timescale memory challenges.
• This approach enhances the capability of artificial intelligence in mimicking complex cognitive functions.
• Future research can explore this hybrid memory system in more complex cognitive architectures and tasks.