Neural learning rules for generating flexible predictions and computing the successor representation.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• The hippocampus plays a role in memory-guided behaviors.
• The successor representation (SR) is a predictive map inspired by reinforcement learning.
• Current models lack a neural mechanism for SR generation.

Purpose of the Study:

• To propose a neural mechanism for the successor representation (SR).
• To demonstrate how recurrent neural networks can compute the SR.
• To investigate the biological plausibility and flexibility of SR computation in neural circuits.

Main Methods:

• Simulated a recurrent neural network (RNN) with synaptic weights matching the transition probability matrix.
• Analyzed RNN dynamics to show natural SR calculation.
• Investigated the effect of network gain on the predictive horizon.
• Derived biologically plausible learning rules for SR acquisition in RNNs.
• Tested the model using realistic inputs and compared it to hippocampal data.

Main Results:

• RNN dynamics naturally compute the SR when synaptic weights align with transition probabilities.
• Network gain provides a flexible mechanism to modulate the predictive horizon.
• Derived learning rules enable RNNs to learn the SR in a biologically plausible manner.
• The model successfully replicated hippocampal data from random foraging tasks.

Conclusions:

• Recurrent neural networks offer a viable neural substrate for generating successor representations.
• The SR is more neurally accessible than previously assumed.
• This framework supports the SR's role in diverse cognitive functions beyond memory, such as planning and decision-making.