Metaplasticity as adaptive gradient preconditioning for incremental learning

The human brain possesses a remarkable ability to learn continuously. A child can learn to recognize a dog, then learn to ride a bicycle, and then learn a new language, all without forgetting how to identify a dog. This seamless flow of new knowledge does not overwrite old memories. In contrast, artificial intelligence systems, which are built on mathematical models inspired by the brain, struggle profoundly with this same task. When these computer programs learn a new task, they often suffer from "catastrophic forgetting," a phenomenon where the new learning violently erases the old. The system becomes unstable, unable to hold onto its past while trying to grasp its present. Scientists have long sought to solve this "stability-plasticity dilemma," trying to find a way for machines to be flexible enough to learn new things but stable enough to keep what they already know.

For decades, researchers have tried to mimic the brain's solution by adding extra rules or saving old data to review later. However, these approaches often require the computer to know exactly when one task ends and another begins, or they demand massive amounts of memory to store past examples. These requirements feel artificial and do not reflect how biological intelligence actually works. A new study from researchers at the University of Sydney proposes a different path. Instead of forcing the computer to remember old data or stop to calculate importance after a task is finished, they suggest the system should adjust its own learning speed in real time, moment by moment, just as a living synapse does. They call this approach SynGAP, a method that treats the learning process not as a series of distinct steps, but as a continuous, flowing river where the current naturally shapes the banks.

The core idea draws from a biological concept known as metaplasticity. In the brain, a synapse is not just a simple switch that turns on or off; it has a history. The state of a synapse changes based on how much it has been used in the past, making it either more or less likely to change again. This is the "plasticity of plasticity." If a connection has been strengthened over a long time, the brain makes it harder to alter, protecting that memory. If a connection is new or unused, it remains flexible. The researchers realized that artificial neural networks could benefit from this same principle. Instead of waiting until a task is complete to decide which parts of the network are important, their method allows the network to constantly monitor its own history. It keeps a running average of how much each part of the network has contributed to past learning, effectively creating a memory of importance that updates with every new piece of data.

To put this into practice, the team developed a system that acts like a dynamic valve for the learning process. When the computer tries to learn something new, it calculates the standard adjustments needed to improve its performance. However, before these adjustments are applied, they pass through a filter. This filter checks the history of each connection. If a connection has been critical for past tasks, the filter gently slows down the change, preventing the new learning from destroying the old. If the connection is not critical, the filter allows the change to happen freely. Crucially, this happens without the computer needing to know what task it is currently doing, and without needing to store a large library of old examples to look back at. The system simply flows forward, adjusting its own sensitivity based on what it has already learned.

The researchers tested this approach on two challenging benchmarks where a computer must learn to recognize different categories of images one after another. In one test, the system had to learn 100 different classes of images, split into ten groups. Standard learning methods, which simply try to minimize errors as they go, forgot the first groups almost entirely, retaining less than 10% accuracy on the old tasks. Methods that tried to save old data or add extra penalties to the learning process performed better but still struggled, often forgetting more than half of what they had learned. The new SynGAP method, however, achieved a significantly higher level of success. It retained an average accuracy of nearly 27% on the old tasks, which is more than four times better than the standard regularization method and nearly 10% better than the method that relies on saving old data. It also reduced the measure of forgetting by over 10% compared to those established techniques.

In a second test involving a continuous stream of object recognition, the results were similarly strong. The new method reached an accuracy of about 68%, outperforming the best standard learning tools by roughly 10 percentage points. Perhaps most importantly, the system did not become rigid. Some methods that protect old memories too well make it impossible for the computer to learn anything new. The researchers found that their approach struck a delicate balance: it protected the past without freezing the future. The system remained flexible enough to adapt to new information while keeping the core structure of its old knowledge intact.

The success of this approach lies in how it changes the nature of the learning update. Traditional methods often try to fix the problem by adding a penalty to the final goal, essentially telling the computer, "Don't move too far from where you were." This can distort the learning path and create conflicts as the number of tasks grows. The new method, by contrast, modifies the learning step itself. It acts as a geometric guide, steering the computer away from directions that would destroy old memories while allowing it to move freely in directions that do not interfere. This happens continuously, driven by the flow of data itself, rather than by external signals or stored memories. The researchers demonstrated that by mathematically formalizing this biological process, they could create a system that is both robust and efficient, requiring very little extra memory to operate.

While the system is highly effective, the researchers note that it is not a perfect solution for every possible scenario. It currently relies on a simplified view of the network's internal geometry, which works well for many cases but might need refinement for extremely complex problems. It also still requires a small amount of stored data to function, though far less than previous methods. Despite these limitations, the work offers a significant step forward. It shows that by looking closely at how biological systems manage the tension between stability and change, and by translating those principles into a continuous, real-time process, we can build artificial intelligence that learns more like a living mind. The study suggests that the key to preventing catastrophic forgetting is not to stop the learning or to hoard the past, but to let the history of learning shape the very way the future is learned.