When Blocks Outlive Tasks: Three Safeguards for Task-ID-Free Routing in Growing Continual Learners

Abstract Growing continual learners add blocks for new data and freeze old ones. Without task labels, task-identity predictors pick the answering block but assume one block per task. Learners that grow when a block saturates break that assumption: a block learns several data periods and keeps changing while the router relies on it. For task-identity-free routing over such blocks' complete predictions we state three safeguards for replay-based growers, each tested by removal: the active block replays its earlier periods until it freezes; the router's records are written at freeze from final weights; and later training never overwrites the router's state for a frozen block, kept per block or as exact running statistics. In a pre-declared, Holm-corrected confirmation on ten independent seeds, 19 of 20 tests (four hypotheses on five benchmarks) held, except safeguard~2 on Permuted-MNIST. On CLEAR-10's natural drift one block covers nearly every year, so block choice matters little. A freeze-time competence record, compared among blocks claiming the same data, picks a revised block's competent version, but its centring lets version pairs capture other blocks' inputs, so banks keeping old versions route better without it. A bank/router loss split finds selection loss large but router- and stream-dependent: on Split CIFAR-100 the blocks hold 0.80; a per-block router reaches 0.51, and an analytic shared router using every training row 0.62, above a nearest-class-mean baseline (0.58) but not RanPAC (0.75), both non-growing, on the same features. Under per-period testing, RanPAC beats every routed bank on every frozen-feature stream.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11559005/v1
Primary Topic
Domain Adaptation and Few-Shot Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

When Blocks Outlive Tasks: Three Safeguards for Task-ID-Free Routing in Growing Continual Learners

Saurabh Srivastava, Nishant Kumar
Research Square
Domain Adaptation and Few-Shot Learning
preprint

When Blocks Outlive Tasks: Three Safeguards for Task-ID-Free Routing in Growing Continual Learners

Saurabh Srivastava, Nishant Kumar
preprint en

Abstract

Abstract Growing continual learners add blocks for new data and freeze old ones. Without task labels, task-identity predictors pick the answering block but assume one block per task. Learners that grow when a block saturates break that assumption: a block learns several data periods and keeps changing while the router relies on it. For task-identity-free routing over such blocks' complete predictions we state three safeguards for replay-based growers, each tested by removal: the active block replays its earlier periods until it freezes; the router's records are written at freeze from final weights; and later training never overwrites the router's state for a frozen block, kept per block or as exact running statistics. In a pre-declared, Holm-corrected confirmation on ten independent seeds, 19 of 20 tests (four hypotheses on five benchmarks) held, except safeguard~2 on Permuted-MNIST. On CLEAR-10's natural drift one block covers nearly every year, so block choice matters little. A freeze-time competence record, compared among blocks claiming the same data, picks a revised block's competent version, but its centring lets version pairs capture other blocks' inputs, so banks keeping old versions route better without it. A bank/router loss split finds selection loss large but router- and stream-dependent: on Split CIFAR-100 the blocks hold 0.80; a per-block router reaches 0.51, and an analytic shared router using every training row 0.62, above a nearest-class-mean baseline (0.58) but not RanPAC (0.75), both non-growing, on the same features. Under per-period testing, RanPAC beats every routed bank on every frozen-feature stream.

Research Square
Indian Institute of Technology Dhanbad (IN)
Domain Adaptation and Few-Shot Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon