family: hybrid
components:
- optimization
- architecture
introduced-in: '[[liang2023-api]]'
stores:
- layer input-subspace bases (via DualGPM)
- per-task expanded channel parameters
requires-boundaries: 'yes'
requires-task-id: train+test
memory-growth: sublinear
tags:
- method
API (Adaptive Plasticity Improvement)
What it does
A from-scratch continual learner that couples orthogonal-projection stability with measured, demand-driven expansion. Old tasks are protected by projecting each new-task gradient onto the complement of the accumulated old-task subspace (maintained by DualGPM). Because that projection steadily consumes the space left for new tasks, API measures how much of the raw new-task gradient survives the projection — a per-layer gradient-retention ratio (GRR) it shows correlates with new-task accuracy — and, before training each task, expands only the layers whose retention has fallen below a target, adding input channels in proportion to the deficit. So stability comes from the projection and plasticity is restored exactly where the projection has depleted it.
Aspects
- Stores: the DualGPM old-task subspace bases (see [[dualgpm]]) plus the per-task expanded channel parameters Bₗ,ₜ. No raw samples, no stored gradients. Expanded parameters (API-Param) grow monotonically but stay small relative to the basis store (Split CIFAR100: total 2.0M, capacity 105%).
- Update rule: for each new task t, (1) compute per-layer GRR(l,t) = E‖ĝₗ,ₜ‖/‖gₗ,ₜ‖ (surviving projected-gradient norm over raw), (2) set added input dimension dₗ ← dₗ + max(⌊K(ρ − GRR(l,t)) + 0.5⌋, 0) so low-retention layers grow and layers with GRR ≥ ρ do not, (3) train the (now wider) network with DualGPM-projected SGD; the newly added channels' transform Bₗ,ₜ is trained for the current task while the portions of Bₗ,ₜ belonging to previous tasks are frozen. K=10, ρ=0.5 across all experiments.
- Requires: task boundaries at training; task identity at both train and test — inference selects the task-specific expanded weight Wₗ,ₜ by task id (multi-head task-IL). From-scratch; per-layer ε_th matched to GPM; K, ρ tuned for the memory/accuracy trade-off.
- Costs: memory sublinear in tasks (DualGPM store rises then falls; expanded params grow slowly). Compute: one GRR estimation pass + DualGPM SVDs per task; training projects every mini-batch gradient. Inference is a single forward pass through the task-selected weights.
Documented behavior
- [[liang2023-api]] (introduces): task-IL from-scratch, 20-task streams — best ACC on all four benchmarks (Split CIFAR100 81.4 / BWT −0.8; Split Mini-Imagenet 65.9 / −0.3; 5-Datasets 91.1 / −0.5; CIFAR100-sup 60.2) above GPM (78.9 / 61.2), TRGP (80.5 / 62.5) and expansion baselines, at lower memory than GPM and lower capacity (105%) than DEN/RCL/APD (130–191%). Ablations: the GRR-driven expansion beats a fixed-per-task "Equal" channel budget at matched accuracy; DualGPM-vs-GPM backing changes memory not accuracy; API's gain over GPM widens with task index as GPM's plasticity falls. Authors flag the method is confined to the task-IL setting (task id at test) and leave class-IL / task-free extension open.
Links
combines:: [[dualgpm]] (supplies the old-task gradient subspace and the orthogonal projection that provides stability; API adds the measured expansion on top) uses:: [[orthogonal-update-projection]] (new-task gradients projected onto the old-task subspace complement each step — DualGPM's constraint, providing the stability half) uses:: [[retention-gated-channel-expansion]] (per-layer input channels added before each task in proportion to the gradient-retention deficit ⌊K(ρ − GRR(l,t))+0.5⌋, so only projection-choked layers grow — GRR read directly off the projection as a data-free plasticity gauge)
Papers engaging this method
LIST FROM [[]] AND ("papers" OR "influences")