Adult neurogenesis acts as a neural regularizer.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Adult neurogenesis continuously generates new neurons in the hippocampus.
• These new neurons integrate into existing circuits, potentially introducing "wiring" noise.
• Noise injection is a known regularization technique in machine learning to improve model generalization.

Purpose of the Study:

• To investigate if a neurogenesis-like process can act as a regularizer in computational models.
• To determine if this process enhances generalization in a category learning task.

Main Methods:

• A convolutional neural network (CNN) was trained on the CIFAR-10 object recognition dataset.
• Neurogenesis was modeled as a weight reinitialization process for a subset of hidden layer neurons.
• Performance was compared against networks with no neurogenesis and conventional noise injection methods (dropout, weight decay).

Main Results:

• Neurogenesis significantly enhanced generalization performance on unseen test data.
• Neurogenic networks performed comparably to or better than networks using conventional noise injection techniques.
• The findings support the role of neurogenesis as a noise-injection-based regularizer.

Conclusions:

• Neurogenesis may facilitate generalization in hippocampal learning by acting as a natural regularization mechanism.
• This computational model expands our understanding of the cognitive roles of adult neurogenesis.
• The study highlights potential parallels between biological learning and machine learning regularization.