No replay / Task A
—
baseline after Task B
Retention / plasticity
Learn new.
Keep old.
Train one small classifier on a task, then give it a conflicting task. Replay a few old examples and watch the tradeoff become measurable.
Waiting for an experiment
Run the fixture to compare overwrite and replay.
Task A / before
—
old task accuracy
Task B / new
—
after conflicting updates
Retention
—
Task A points forgotten
Replay / Task A
—
selected retention
Task B
—
selected plasticity
Replay budget
—
old samples / epoch
w1 —w2 —bias —
The completed experiment will explain whether the new task overwrote the old decision boundary.
Interpretation boundary
This is a six-point conflicting fixture. Replay is a useful mechanism to inspect, not proof that continuous learning solves general intelligence, theorem discovery, or adaptation in the wild.
This is a six-point conflicting fixture. Replay is a useful mechanism to inspect, not proof that continuous learning solves general intelligence, theorem discovery, or adaptation in the wild.
Plasticity writes new rules
Task B applies gradients that fit its inverted labels. Without a retention mechanism, those updates can overwrite the Task A boundary.
Replay spends capacity
Old examples preserve memory, but replay competes with the new task. The useful question is which balance your real workload needs.