Pruning recurrent neural networks replicates adolescent changes in working memory and reinforcement learning.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Developmental Neuroscience

Background:

• Adolescence features cognitive improvements alongside decreased skill acquisition speed.
• Significant synaptic pruning occurs in the cortex, particularly the prefrontal cortex, during this developmental stage.

Purpose of the Study:

• To model adolescent brain development using recurrent neural networks (RNNs).
• To investigate the computational benefits of synaptic overproduction and pruning on cognitive tasks.

Main Methods:

• Trained RNNs on working memory and reinforcement learning (RL) tasks.
• Applied incremental synaptic pruning with retraining to the trained networks.

Main Results:

• Pruned networks showed increased resistance to distraction in working memory tasks.
• Pruned RL networks demonstrated more accurate value estimates and consistent optimal choices.
• Pruned networks exhibited slower learning for some new problems, indicating a trade-off with flexibility.

Conclusions:

• Synaptic pruning enhances cognitive performance, mirroring adolescent development.
• Overproduction and subsequent pruning of synapses offer a computationally advantageous strategy for brain development.
• Cognitive competence gained through pruning may come at the expense of learning flexibility.