Algorithms mimic biological evolution of the brain to learn better

Researchers at the University of Bern, Switzerland, have developed new algorithms that mimic the process of biological evolution to “learn” to perform tasks more efficiently, reducing the time spent training artificial intelligences.

So-called “evolutionary algorithms” are computer programs that look for solutions to complex problems by imitating the natural human learning system. This approach uses the same principle in which biological fitness directly affects the way an organism adapts to its environment.

“Our brains are incredibly adaptable. Every day, we form new memories, acquire new knowledge or refine existing skills. This contrasts with our current computers, which normally only perform pre-programmed actions”, explains Professor Jakob Jordan, co-author of the study.

Want to catch up on the best tech news of the day? Access and subscribe to our new youtube channel, Canaltech News. Every day a summary of the main news from the tech world for you!

synaptic plasticity

Synapses are the connection points between neurons, which can change as they are used or stimulated. This synaptic plasticity is an important research topic in neuroscience, as it involves all processes of learning and acquired memory.

Scheme and use of evolutionary algorithms (Image: Reproduction/University of Bern)

To develop a machine with this adaptive capability, scientists need models that can operate mechanisms underlying these processes. With these models, it is possible to build artificial intelligence systems based on the processing of biological information, creating machines that learn much faster.

“In all these scenarios, evolutionary algorithms are able to discover synaptic plasticity mechanisms and successfully solve a new task. In doing so, these algorithms have shown incredible creativity and adaptive capability far superior to that found in current AI systems,” adds Jordan.

evolving to learn

The approach developed by researchers is known as “evolving to learn” or “becoming adaptable”. With it, it is possible to confront evolutionary algorithms with three typical learning scenarios. In the first, the computer must detect a repeating pattern without receiving feedback on its performance.

Cartesian genetic programming develops several efficient reward-driven learning rules (Image: Reproduction/University of Bern)

In a second scenario, the system receives virtual rewards for behaving in a desired way. In a third phase, with an oriented learning scheme, the computer is informed about how much its behavior deviated from what was expected.

“We see this system as a promising approach to gaining in-depth knowledge about biological learning principles and accelerating the progress of artificial learning machines, paving the way for the development of intelligent devices capable of better adapting to the needs of their users”, he concludes. Professor Jakob Jordan.

Leave a Comment