Evolving neural networks through neuroevolution

Kenneth Stanley offers an overview of the field of neuroevolution, an emerging paradigm for training neural networks through evolutionary principles that has grown up alongside more conventional deep learning, highlighting major algorithms such as NEAT, HyperNEAT, and novelty search, the field's emerging synergies with deep learning, and promising application areas.

Kenneth Stanley offers an overview of the field of neuroevolution, an emerging paradigm for training neural networks through evolutionary principles that has grown up alongside more conventional deep learning. While deep learning focuses on how brain-like structures in computers can learn, neuroevolution addresses how they evolve in the first place, from their architectures to their intrinsic learning dynamics. As in the broader field of deep learning, increases in available computation have led to a renaissance in potential applications of neuroevolution, some of which complement more conventional techniques by offering a path to novel architectures, while others reveal intriguing alternative systems of incentives for learning (even when a gradient is not available). Neuroevolution offers a rich and unique history of exploring creative and divergent algorithms. Kenneth introduces key algorithms, explains their history and motivations, and shares insight into the kinds of applications they enable. Along the way, he touches on available platforms and software packages and potential links to other deep learning frameworks.