An explainable artificial intelligence approach to spatial navigation based on hippocampal circuitry.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Mammalian spatial navigation is highly efficient, learning complex environments like mazes in few trials.
• Current deep learning algorithms struggle with learning long object sequences for trajectory planning, requiring extensive training.
• This highlights a gap between biological cognitive function and artificial intelligence capabilities.

Purpose of the Study:

• To extend a previously proposed single learning trial (SLT) model for object sequence learning.
• To introduce the capability for artificial agents to navigate and learn optimal paths in a classic four-arms maze in a single trial.
• To demonstrate robust and efficient implementation of spatial navigation as a fundamental cognitive function using an extended SLT (e-SLT) network.

Main Methods:

• Developed an extended single learning trial (e-SLT) computational model.
• Incorporated neural components such as place cells, head-direction cells, and object-coding cells.
• Simulated the e-SLT network within a classic four-arms maze environment.

Main Results:

• The e-SLT model successfully learned to navigate the four-arms maze, identifying the correct path to the exit in a single trial.
• The network demonstrated the ability to ignore dead ends and efficiently learn the trajectory.
• Conditions for robust and efficient cognitive function implementation within the e-SLT network were identified.

Conclusions:

• The e-SLT model provides a framework for understanding hippocampal circuitry and operation in spatial navigation.
• This research offers insights into how biological brains achieve rapid learning of complex environments.
• The findings may serve as a foundation for developing next-generation artificial intelligence algorithms for spatial navigation.