Artificial intelligence systems — and human brains — often stumble at a surprising moment: right when they start learning something new, their performance on old tasks briefly collapses. Now, a team of researchers has borrowed a brain mechanism to soften that dip.
The research
Alejandro Rodriguez-Garcia, Anindya Ghosh, and Srikanth Ramaswamy studied a problem in continual learning called the stability gap: a temporary performance drop on previously learned tasks when new ones are introduced. Their work appears on arXiv (submitted July 18, 2025, revised August 10, 2026; arXiv:2507.14056v3), spanning machine learning, artificial intelligence, and neurons and cognition.
Standard optimizers like momentum-SGD and Adam already provide some balance between adapting to new data and keeping old knowledge. But the researchers found these optimizers still show clear stability gaps — even under ideal "joint training," where all tasks are trained together at once. That makes joint training a clean setting to isolate the gap's causes.
Inspired by how noradrenergic bursts in the brain suddenly increase neuronal gain under uncertainty, the team added a dynamic gain scaling mechanism. Think of it as a two-timescale optimizer: it briefly boosts the effective update magnitude at task transitions while dynamically reparameterizing the weights that govern the forward pass. The goal: reduce the curvature amplification that happens when the data distribution shifts abruptly.
They tested dynamic gain scaling on domain- and class-incremental versions of MNIST, CIFAR, and mini-ImageNet benchmarks under task-agnostic joint training. The method attenuated the stability gaps while maintaining competitive accuracy. In other words, the system didn't just trade old skills for new ones — it kept both steadier.
Why it matters
Your brain faces the same stability–plasticity dilemma every day. When you learn a new software tool, a language, or a driving route, your older habits can wobble. Noradrenaline — a neuromodulator linked to alertness and uncertainty — helps your neurons temporarily turn up their "gain," making them more responsive without wiping out existing connections. This study shows that same principle can be translated into a mathematical optimization rule.
For anyone interested in cognition, the takeaway is that sharp learning transitions aren't a personal failing; they're a fundamental trade-off. The research also hints that brief, strategic boosts in attention or arousal — the kind that noradrenaline provides — may help you cross those boundaries more smoothly.
What you can do
- When switching between demanding tasks, give yourself a short pause or novelty boost — a walk, a change of scene — to mimic a controlled gain increase.
- Avoid stacking multiple brand-new skills in one sitting; let each transition settle before adding the next.
- If you feel slower on an old task after starting a new one, recognize it as a temporary stability gap, not lost ability.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.