❗️AI models are learning to “reproduce”, without training

Researchers have built a system where neural networks can effectively breed. Instead of using gradient descent to train one model from the top down, the system starts with a population of randomly initialized neural networks and lets evolution do the work.

Models with complementary strengths are paired together. Their parameters are then combined through a crossover process, producing a new “offspring” network that inherits traits from both parents. The offspring compete for survival based on a fitness function. Better-performing models survive and reproduce, while weaker ones are removed.

The system can evolve increasingly capable neural networks across generations without a single step of gradient descent. It’s essentially natural selection applied to AI, random networks, competition, reproduction, inheritance and mutation replacing the traditional training pipeline.

Source.

@aipost 🏴