Researchers at the Technion – Israel Institute of Technology have developed a mathematical description of learning that can expose hidden differences between networks that behave identically, according to a study published by the university.

Learning in the brain involves changes in connections between neurons, but the results are visible in the behavior of the entire system. The new research, led by Prof. Omri Barak and graduate student Yoav Gur from the Ruth and Bruce Rappaport Faculty of Medicine, examined this relationship in mathematical models of recurrent neural networks — artificial models used to study computations in the brain.

The researchers created a mathematical framework that calculates how a network changes during learning using a small number of variables. This analysis led to a distinction between two types of features: those that determine the network's actual behavior, and those that do not affect current performance but influence how the network learns in the future. These "hidden" features are not apparent when examining the system's behavior alone, but may be revealed when the system is required to learn a new task.

In one example from the study, two networks with different internal structures responded identically to the same inputs. However, when trained under the same conditions, they developed different learning trajectories: one showed a period of significant slowdown in improvement, while the other improved more continuously. This demonstrates that learning itself can serve as a research tool, revealing differences not visible when examining only the system's current function.

The research also showed that under certain conditions, these hidden features can retain traces of learning history. Networks initially trained on different tasks and then on a shared task achieved similar performance, but their hidden structural features still allowed researchers to identify which training history each network had undergone.

"The question of how changes in the structure of a complex system affect behavior recurs in many fields of science," Prof. Barak said, according to the Technion. "In the networks we studied, we found that understanding what the system does today is not enough to describe how it will learn in the future. The way it changes as a result of learning can reveal structural differences that remain hidden in its current function."

The results were obtained in mathematical models and computer simulations. They suggest a direction for experiments in biological systems: to examine how systems with similar behavior change following the same learning process, and to check what internal differences can be revealed this way.

The research was supported by the Israel Science Foundation (ISF) and the Human Frontier Science Program. The paper was selected for an oral presentation at NeurIPS 2026, one of the world's leading conferences in artificial intelligence and computational neuroscience. Out of 30,709 papers submitted to the conference, only 112 were accepted for this track (approximately 0.36%).

For university students studying neuroscience, artificial intelligence, or related fields, this work offers a new way to think about learning systems — whether biological or artificial. It suggests that observing how a system learns a new task can reveal structural properties that a snapshot of its current performance cannot. This could have implications for understanding brain plasticity, designing more adaptable AI systems, and even diagnosing network-based differences in biological or engineered systems.