Mimicking evolution helps artificial intelligence learn much faster

Summary

When humans or animals learn a new skill, they do not start from a blank slate. Biological brains come equipped with built-in structures shaped by millions of years of evolution. This innate wiring allows nature to perform "few-shot learning", meaning organisms can pick up new tasks after just a few attempts. In contrast, standard artificial neural networks (computer systems designed to process information like a brain) usually start completely random and blank. Because they lack this evolutionary head start, these artificial networks require massive amounts of training data to learn comparable skills.

To see if computers could benefit from nature's approach, researchers developed a novel algorithm that combines the rules of natural selection with standard machine learning techniques. By simulating evolutionary conditioning, they hoped to build artificial networks that learn more like living brains. Interestingly, the researchers found that this evolutionary simulation by itself did not instantly create a highly capable system. On its own, the conditionally evolved network performed about the same as a basic, untrained computer model.

However, a major difference emerged once the researchers actually started teaching the networks. The artificial models shaped by simulated evolution showed unique, hidden learning dynamics that allowed them to rapidly fine-tune their skills. Instead of blindly guessing through trial and error, the artificially evolved networks used their built-in foundational traits to accelerate the learning process and quickly reach optimal performance.

The story in numbers.

Machine learning contexts evaluated (reinforcement and supervised)

2

In vivo biological experiments (purely computational study)

0