Continual learning

A digit reader learns handwriting, then trains on printed digits. Will it still read handwriting after the update? In an invented test, it gets 18 of 20 handwritten digits right before the update and 14 afterward, while getting 16 of 20 printed digits right. New learning and lost old skill are different measurements. Continual learning asks how to acquire new abilities while keeping useful earlier ones, with limits on stored examples, model size and computation. A method that stores and repeatedly studies its entire history avoids one resource constraint by spending more resources.[1]

How can a model learn a sequence of changing tasks or data populations while retaining useful earlier performance under a fixed resource budget? Consider the illustrative handwriting-then-print reader. On the same 20 held-out handwritten examples, accuracy changes from $18/20=90\%$ to $14/20=70\%$: a loss of 20 percentage points. Its $16/20=80\%$ printed-digit score measures a different skill. Catastrophic forgetting refers to a severe loss of earlier performance after new training; the toy numbers illustrate the measurement, not a finding about a real model. Replay stores old examples for later updates, so its storage and reuse must be counted. Growth in the number of adjustable model values and training operations must also be counted. Guarantees need an explicit stream model and an evaluation rule; low error on old training examples alone does not establish retention on unseen old examples. The cited 2026 analysis distinguishes training and test forgetting for a particular neural-network architecture and structured noisy task data.[1] [2]

Methods work in particular settings; dependable retention and adaptation across specified changes remain incomplete.[1]

The 2026 gradient-descent study analyzes a restricted network and task-distribution model. Its theory and experiments do not establish universal forgetting-free learning.[1] [2]

A reader that learns new fonts but loses familiar handwriting makes each update a tradeoff for its users. Keeping every past example and retraining from the beginning also becomes costly as the history grows. Understanding retention would let an updating system accumulate useful skills within a stated budget, and reveal when an apparent improvement simply spends more memory or computation.[1]

Reliable retention would let the digit reader add new styles while continuing to serve users who write in older styles, without retaining and retraining on every historical example. A model’s adjustable values, or parameters, are shared across tasks: improving later performance can change decisions that earlier skills relied on. Understanding this interference would explain when resources can support cumulative knowledge rather than a succession of replacements. It also connects retention to generalization: keeping stored examples correct alone does not establish useful performance on unseen examples from an earlier population.[1] [2]