Episodic memories enable powerful algorithms for goal-directed decision making that are context-aware and explainable

Current AI tools are criticized for their lack of context-awareness and explainability. This makes their action selection sometimes undesirable for a given context, or even unsafe. In contrast, our brains can adjust their goal-directed action selection on the fly to the current context, and we can explain our action selection in terms of related experiences. This capability results from a fundamental difference between the way how brains and current AI tools store experiences. Whereas most AI tools transform experiences into parameter values, the brain also stores explicit representations in the form of episodic memories, and uses them for goal-directed action selection. We show that experimental data on neural codes for episodic memories in the brain suggest a method that allows us to port these capabilities into artificial devices. The resulting Episodic Neighbor Algorithm attains high task performance by creating a cognitive map from episodic memories, which provides a clear and explainable sense of direction for goal-directed action selection. Furthermore, this cognitive map can be reconfigured on the fly if the goal or context change, or when new experience is added. The resulting algorithm for goal-directed action selection requires only local rules for synaptic plasticity in shallow neural networks, and is therefore suitable for implementation in energy-efficient edge devices.