Recycling the Stability–Plasticity Conflict in Continual Learning
Research question. Can a single recycling budget govern both unit reinitialisation and subspace eviction, so that a network retains old tasks without losing the plasticity that lets it keep learning new ones?
- Dates
- October 2026–present
- Research areas
- Continual Learning, Robot Learning, Representation Learning
The conflict
Two lines of work each solve half of continual learning and break the other half.
Continual Backprop keeps a network learning indefinitely by continually reinitialising low-utility units. It never saturates, it never goes rigid, and it stores no data. It also forgets: nothing protects what earlier tasks built, so a reset is free to destroy it.
Subspace methods — GPM, and TRGP for transfer — do the opposite. They record the directions earlier tasks relied on and project new gradients away from them. Old tasks survive. But the protected subspace only ever grows, the space of permitted updates only ever shrinks, and the network eventually cannot learn at all.
Grafting one onto the other naively fails in a specific way, and the failure is the interesting part: reinitialisation and projection are fighting over the same resource. The reset mechanism does not know which units carry protected structure, so it destroys exactly what the projection is trying to preserve.
The idea being tested
One budget, two worlds. Units are recycled at a constant rate; directions are evicted when the subspace budget is full; and the same budget governs both. Three couplings are what make that more than a bolt-on:
- Unit → direction. A memory-aware utility makes reinitialisation defer to protected directions, so a reset stops destroying retained structure.
- Direction → unit. Recycled placement makes new units grow out of the protected subspace rather than beside it, which is where transfer should come from if it comes from anywhere.
- Budget coupling. A single recycling rate governs both, under a memory ledger that stays flat as tasks accumulate — no stored samples, ever.
Whether those couplings do anything is precisely what is unresolved.
Predictions, written first
The design fixes a stage ladder and states, in advance, what each stage should do to a five-criterion report card — retention, memory growth, plasticity, freedom from task labels, and forward transfer.
| Stage | Added mechanism | Prediction |
|---|---|---|
| S0 | Continual Backprop baseline | Reproduces the published behaviour: plasticity holds, retention fails |
| S1 | Protected subspace + projection | Retention rises and plasticity collapses — the tension, reproduced inside the method |
| S2 | Memory-aware utility | Retention restored, plasticity partially recovered |
| S3 | Budgeted eviction and recycling | Plasticity returns while retention holds — the core claim |
| S4 | Trust-region transfer | Transfer appears without costing retention |
| S5 | Recycled reinitialisation placement | Transfer improves further |
S1 is a prediction that the method will get worse, and S3 is the one the whole idea stands or falls on. Stating both in advance is the point: a ladder written after the fact can be made to fit whatever happened.
What exists
Implemented and running end to end: permuted-MNIST and split-MNIST streams, an MLP exposing per-layer activations, the Continual Backprop utility and reinitialisation path, the subspace memory with online second-moment estimation, budgeted selection with lease carry, projection with a transfer gate, and the metrics — task accuracy, effective feature rank, dead-unit fraction, prototype bank, memory ledger.
Not established: anything at all about whether the couplings work. The runs to date are single-seed and hand-tuned while the mechanisms were being calibrated. They told me the basis timescale matters and that protecting a classifier head is useless if the features underneath it drift — both of which are facts about the testbed, not findings about the method.
Where it connects
Research 1 to 3 asked how a robot should represent and act on uncertainty that already exists. This asks a question underneath that: what a network may safely change about itself while it keeps learning, and what it must hold still.
That is the same question Research 4 was scaffolded around — what may adapt online and what must stay fixed — approached from the learning-algorithm side rather than the navigation stack. The testbeds here are deliberately small and standard, because the mechanism has to be shown to work somewhere clean before it is worth putting near a robot.