Researchers at the University of Bern have demonstrated that evolutionary algorithms can discover entirely new rules for how the brain learns. Their approach, called E2L ("evolve-and-learn"), searches the space of possible synaptic plasticity rules just as natural selection searches the space of possible organisms. In tests across three learning scenarios, the algorithm found rules that solved assigned tasks. In one case, it invented a learning signal that had not been explicitly programmed, and networks using it learned faster than with previously known rules. The work, published in eLife, suggests that evolutionary search could become a standard tool for designing learning algorithms, not just tuning them.
Why it matters
Synaptic plasticity is the mechanism by which neurons strengthen or weaken their connections based on activity. It underlies all learning and memory in biological brains. For AI researchers, plasticity rules are the bridge between biology and machines—a way to make software learn the way brains do.
The catch is that designing good plasticity rules is hard. Researchers start with biological intuition: spike timing matters, neuromodulators like dopamine adjust learning rates, local voltage gates changes. Translating that intuition into equations that actually work requires trial, error, and a fair amount of guesswork.
The Bern team's approach flips this. Instead of hand-designing rules based on biological principles, they let an evolutionary algorithm search the space of possible rules, then test which ones solve a target problem. The algorithm is free to discover solutions that no human would have considered—including, in this case, a novel combination of signals that outperformed existing models.
How E2L works
E2L starts with a population of candidate learning rules, each encoded as a small program. The researchers evaluate each rule on a learning task. Successful rules survive and reproduce, with mutations and operator swaps introducing variation. Over many generations, the population converges on rules that solve the task.
The team tested this approach across three scenarios. In unsupervised learning, the network had to detect a repeating pattern in a continuous stream of input without feedback. In reward-based learning, it received virtual rewards for desired behavior. In guided learning, it received an explicit error signal. In all three cases, evolutionary search produced working rules.
The most striking result came when the algorithm invented a new way of combining signals the researchers had defined into a single, novel signal. Networks using this discovered signal learned faster than networks trained with previously known rules.
Who should care
| Audience | Why it matters | Key trade-off |
|---|---|---|
| AI researchers | Discover rules that hand-design might miss; faster learning without manual tuning | Interpretability suffers; rules may not transfer to other domains |
| Neuroscientists | Generate testable hypotheses about plasticity mechanisms | Discovered rules may not reflect actual brain biology |
| Neuromorphic engineers | Rules optimized for spike-based hardware | Rules must be discretized for implementation |
| Robotics developers | Evolve controllers for specific tasks without expertise in plasticity | Evolved rules are often opaque; hard to debug |
Our analysis
The E2L approach represents a shift in who designs learning algorithms. Hand-designed rules are shaped by the designer's intuitions about what matters. Evolutionary search is shaped by what works. That difference matters when the space of possible solutions is vast and human intuition is unreliable.
The trade-off is interpretability. A rule that wins the evolutionary search may solve the task while remaining difficult to map onto known biological mechanisms. For engineers building adaptive AI, this is acceptable. For neuroscientists hoping to explain how the brain learns, it creates a new challenge: the algorithm has found a solution, but not necessarily one that corresponds to reality.
Two developments to watch. First, whether E2L-discovered rules can be ported efficiently to neuromorphic hardware, where spiking neurons impose discretized, time-constrained computations. Second, whether the approach can find rules for continual learning, where a network must keep adapting without forgetting what it previously learned. Both are open problems, and both matter for real-world applications.
Takeaway
Evolutionary algorithms are now a credible tool for discovering learning rules, not just optimizing network weights. The Bern team's work suggests this approach can find solutions that hand-design misses. The practical limit is that evolved rules often resist easy explanation—a trade-off between raw learning performance and the ability to understand what the system is doing.
Frequently asked questions
What is E2L?
E2L stands for "evolve-and-learn." It is an approach that uses evolutionary algorithms to search for synaptic plasticity rules automatically, rather than hand-designing them.
What did the algorithm discover that was unexpected?
The algorithm combined signals the researchers had defined into a new signal. Networks using this discovered signal learned faster than with previously known plasticity rules.
How is this different from training a neural network?
Training a neural network adjusts its weights under a fixed learning rule. E2L searches for the learning rule itself, evolving populations of candidate rules across multiple generations.
Are the discovered rules biologically plausible?
Some discovered rules are combinations of biologically inspired signals. Whether each one corresponds to actual brain mechanisms is a separate question the researchers did not claim to answer.
Where was this work published?
The work appeared in eLife. The team was based at the University of Bern's Institute of Physiology, with co-authors from RIKEN in Tokyo and the University of Heidelberg. The research was part of the European Human Brain Project.