A team of computational neuroscientists has unveiled a new artificial neural network model that mirrors how the human brain juggles the competing demands of forming new memories while preserving older ones — a long-standing puzzle in both neuroscience and artificial intelligence. The research, published in late 2024 and gaining renewed attention through follow-up commentary in 2025, suggests that biological neurons employ a far more sophisticated form of “synaptic gating” than previously appreciated, with implications ranging from Alzheimer’s research to next-generation machine learning systems.
The Stability-Plasticity Dilemma
For decades, researchers have grappled with what is known as the “stability-plasticity dilemma.” A brain — or any learning system — must be plastic enough to absorb new information, yet stable enough to retain what it has already learned. Lean too far in either direction and the system either fails to learn or suffers what AI engineers call “catastrophic forgetting,” in which fresh learning overwrites prior knowledge. This trade-off lies at the heart of computational neuroscience, a field that uses mathematical and computer-based models to understand how the nervous system processes information.
The new model, developed by researchers studying cortical microcircuits, proposes that the brain solves this dilemma through a hierarchical system of inhibitory interneurons that selectively gate which synapses can change at any given moment. Rather than allowing all neurons in a region to update simultaneously, only a small, context-specific subset is permitted to undergo plasticity, while the rest remain locked. The work builds on earlier studies from institutions including the Janelia Research Campus, where scientists have been mapping the wiring diagrams of cortical circuits in unprecedented detail.
How the Model Works
At the core of the model is a class of inhibitory neurons known as VIP and SST interneurons. According to the researchers, these cells act like dynamic switches, opening and closing pathways for synaptic change based on incoming sensory context, attention, and behavioral state. When the brain encounters something genuinely novel, the gating circuit relaxes, allowing pyramidal neurons in that region to update their connections. When familiar input arrives, the gates close, protecting existing memory traces from interference.
This biologically plausible mechanism stands in contrast to most current deep-learning systems, which lack any equivalent feature. Modern large language models, for instance, are typically trained once on a vast dataset and then frozen, because continued training tends to corrode their performance on earlier tasks. Engineers at companies including Google DeepMind have been openly searching for biologically inspired solutions to this problem, and the new model offers a candidate framework that could be tested in artificial systems.
Why This Story Matters
The significance of the work extends well beyond theoretical interest. Disorders such as Alzheimer’s disease, schizophrenia, and certain forms of epilepsy involve documented disruptions to inhibitory interneuron function. If the model’s predictions hold, it would suggest that some of the cognitive symptoms in these conditions — particularly difficulties forming new memories without disrupting old ones — could stem from a breakdown in the brain’s synaptic gating system rather than from neuron loss alone. That would open the door to therapies that target interneuron health specifically.
The research also feeds into a broader push within neuroscience to integrate findings from molecular biology, electrophysiology, and behavior into unified computational frameworks. Initiatives such as the NIH BRAIN Initiative have invested heavily in this kind of cross-scale modeling, arguing that only by linking microscopic mechanisms to whole-brain dynamics can researchers hope to crack disorders that have so far resisted treatment.
Expert Reactions
Outside commentators have welcomed the model while urging caution. Independent neuroscientists have noted that the predictions need to be validated in living tissue, particularly through experiments that selectively silence VIP and SST interneurons during learning tasks. Several labs are reportedly already preparing such experiments, using optogenetic tools to manipulate specific neuron populations in mice as they learn new behaviors.
Critics also point out that even if the gating mechanism is real, it is unlikely to be the whole story. Memory consolidation involves the hippocampus, the cortex, and slow processes that play out during sleep — none of which are fully captured by the current model. Still, the framework provides a testable, quantitative starting point, which is something the field has long needed.
What to Watch Next
Over the coming year, expect to see experimental papers attempting to confirm or refute the model’s central predictions, alongside efforts by AI researchers to implement analogous gating mechanisms in artificial networks. If both lines of work succeed, the result could be a rare instance of neuroscience and machine learning advancing in genuine lockstep — each illuminating the other. The deeper question, of course, is whether understanding how the brain balances memory and learning will eventually allow us to repair that balance when it breaks down in disease.
For more stories on the latest in neuroscience, computational biology, and the science of the mind, visit and bookmark science.wide-ranging.com for ongoing coverage and analysis.
