Correspondence between neuroevolution and gradient descent.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Neural network training commonly employs gradient descent.
• Neuroevolution, utilizing stochastic mutation, is an alternative training paradigm.
• These methods are often perceived as fundamentally distinct approaches.

Purpose of the Study:

• To establish an analytical connection between neuroevolution and gradient descent.
• To demonstrate the equivalence under specific conditions and for various network architectures.
• To bridge the understanding between two major families of neural network training algorithms.

Main Methods:

• Analytical derivation of the equivalence in the limit of small mutations.
• Numerical simulations to validate the correspondence for finite mutations.
• Testing on both shallow and deep neural network models.

Main Results:

• Training neural networks via conditioned stochastic mutation (neuroevolution) is analytically equivalent to gradient descent with Gaussian white noise for small mutations.
• This equivalence is observable in numerical simulations for finite mutations.
• The correspondence applies to both shallow and deep neural networks.

Conclusions:

• Neuroevolution and gradient descent are fundamentally linked training methods.
• This finding unifies perspectives on neural network optimization.
• The study provides a theoretical bridge between evolutionary and gradient-based learning.