Learn · Intermediate

Loss of plasticity: why a network that trains forever stops being able to learn

Loss of plasticity is the finding that a neural network trained continuously on a stream of new tasks gradually loses the ability to learn at all - not just the ability to remember. Push it far enough and it performs worse on a brand-new task than a randomly initialized network would, despite all its accumulated experience. It is the single clearest technical obstacle standing between today's models and the idea of an AI that improves from your work every day.

The result that made this undeniable is Loss of plasticity in deep continual learning, published in Nature in 2024 by Shibhansh Dohare, Richard Sutton, and colleagues at the University of Alberta. Their experiments are almost aggressively simple: take standard networks, train them with ordinary gradient descent on a long sequence of tasks, and measure how well they learn each new one. Early tasks go fine. Later ones go worse. Eventually the network - full of learned features, with every parameter free to move - learns new material more slowly than a network that knew nothing.

The first thing to get straight is that this is not catastrophic forgetting, the better-known failure where learning task B erases task A. Forgetting is about the past. Plasticity loss is about the future. A network can remember its old tasks flawlessly and still be unable to take on a new one. Both can happen at once, which is why they get conflated, but the mechanisms and the fixes are different.

The mechanism is best understood through what dies. As training proceeds, an increasing fraction of units become effectively inert. A unit using a common nonlinearity that has drifted into the region where its output is always zero produces a zero gradient forever, so nothing can ever push it back - it is dead weight that still costs memory. Meanwhile the surviving weights grow steadily larger, and large weights make the loss surface stiffer and each update less consequential. And the network's representations lose rank: the many directions the layer could once express collapse toward a few, so however many parameters remain on paper, the number of genuinely independent things the layer can encode shrinks. Understanding plasticity in neural networks traces these symptoms carefully and finds no single cause - it is a syndrome, not a bug.

The analogy that fits is scar tissue. Every injury heals, and the healed tissue holds. But it is less pliable than what it replaced, and after enough of them the joint that has survived everything can no longer bend. Nothing failed; the accumulated adaptations simply used up the range of motion. A long-trained network is stiff in exactly that way - it has committed its capacity to structures that served earlier tasks and has no give left for the next one.

Reinforcement learning ran into this first, because RL is inherently non-stationary: as the policy improves, the distribution of data it sees changes, so the network is always learning a new task in a way supervised training is not. The Primacy Bias in Deep Reinforcement Learning documents the resulting pathology - agents overfit their earliest experiences and then cannot escape those early conclusions, and the fix that works is startlingly crude: periodically throw away and reinitialize part of the network. Losing everything a subnetwork learned is worth it to get its capacity back.

That crudeness generalizes into the best-known remedy. The Nature paper's method, continual backpropagation, keeps standard training but adds a slow trickle of reinitialization: identify the least-used units and reset them to random values, continuously, forever. Fresh randomness keeps arriving, so the pool of adaptable capacity never fully drains. It is the same intuition as regularization - deliberately handicapping the network to keep it healthy - applied to capacity rather than to fit. Weight decay and normalization help too, by attacking the weight-growth and rank-collapse symptoms directly, but none of it restores a fresh network's learning ability.

Why this matters right now: the most-discussed thesis in AI is that the next major advance will be models that keep learning after deployment - improving from the millions of real tasks they perform rather than from a frozen training run. Loss of plasticity is the reason that is a research problem rather than an engineering one. You cannot simply leave the training loop running. A model updated continuously for a year on user sessions would, on the published evidence, become progressively less able to absorb anything new, and would do so quietly - no crash, no error, just a slow decline in how much each day of experience is worth.

It also reframes something about the current paradigm. Every frontier model you use was trained once, at enormous cost, and then frozen - which looks like a cost decision and is partly a technical necessity. Freezing sidesteps the problem entirely. That is why the workarounds people actually ship are external: retrieval to add knowledge without touching weights, fine-tuning with adapters to add a skill in a small isolated module, agent memory to carry context between sessions. All of them keep the base network exactly where it is, because the honest reading of the plasticity literature is that we do not yet know how to let it keep learning without slowly using it up.

Ground Truth. (2026, August 9). Loss of plasticity: why a network that trains forever stops being able to learn. Ground Truth. https://groundtruth.day/learn/loss-of-plasticity.html

@misc{groundtruth:loss-of-plasticity,
  title  = {Loss of plasticity: why a network that trains forever stops being able to learn},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/learn/loss-of-plasticity.html}
}

Topics: fundamentals · continual-learning · training · optimization · reinforcement-learning