arXiv CS open access

SPARCL: Spectral Partitioned Analytic Continual Learning

2026 · arxiv_cs

arXiv CS · Papers · 2026

deep-learning machine-learning neural-networks

Conceptio Open Archive

Analyze & explore related research in Conceptio

Connect this record directly into your research workflow or AI agent. Explore citations across open repositories, query full text, and export structured proof bundles.

Explore in Search →Connect via MCP / API ↗

SPARCL: Spectral Partitioned Analytic Continual Learning James Hartley University of Sheffield

Zeropy Surio University of Sheffield

arXiv:2608.21307v1 [cs.LG] 21 Aug 2026

Hannah Clarke University of Sheffield

Daniel Whitmore University of Sheffield

Thomas Reed University of Sheffield

Abstract

that overwrite previously useful parameters [7, 9, 12]. More recent approaches improve this trade-off through memory replay, feature distillation, or parameter isolation [2, 3, 17]. Analytic continual learning questions whether iterative optimization is needed at all. Starting from ACIL, a line of work has shown that with frozen backbones and ridgeregression classifiers, class-incremental updates can be implemented through recursive least squares while retaining privacy and strong efficiency [5, 29, 31, 33]. This paradigm has rapidly expanded to generalized settings, online streams, long-tailed regimes, and non-vision modalities [14, 15, 22, 32]. However, analytic methods still exhibit old-class degradation. If no gradients are rewriting parameters, what exactly is being forgotten? Our key observation is illustrated in Figure 1. In analytic CIL, the classifier at step t is not a per-task object but the solution to a joint ridge problem governed by the global inverse autocorrelation (Rt + λI)−1 . New-task samples that align with old principal directions do not overwrite weights directly; instead, they reshape this shared inverse operator. The consequence is an eigenvalue dilution effect: old discriminative directions become relatively less amplified, and old logits drift even when the old crosscorrelation term is untouched. This mechanism is distinct from the optimization-centric explanations used in gradientbased CIL. We build on this insight and propose SPARCL (Spectral Partitioned Analytic Continual Learning). SPARCL maintains the same running autocorrelation and cross-correlation statistics as ACIL, but decomposes the feature space into a high-energy core subspace and a residual complement. For old classes, the classifier components lying in the core subspace are frozen once established, while subsequent updates are restricted to the residual block. This residual block can be expanded with low-cost random projections to restore capacity when task shift accumulates. The resulting update remains fully analytic, admits an efficient block-Woodbury implementation, and provides a direct invariance guarantee for the core contribution to old-class logits. The paper makes four contributions:

Analytic continual learning has emerged as a strong exemplar-free alternative to gradient-based classincremental learning because it replaces iterative optimization with closed-form ridge updates. Yet the usual forgetting narrative, centered on stochastic gradient overwriting, does not explain why analytic methods still drift on old classes despite exact recursive solvers. We identify the culprit as spectral interference: the joint ridge classifier for all tasks shares the inverse autocorrelation operator (R + λI)−1 , so incoming task samples that load onto old dominant eigendirections dilute the spectrum and perturb old-class logits even when old labels are never revisited. Based on this view, we propose SPARCL, a spectral partitioned analytic continual learner that decomposes the running autocorrelation into a high-energy core and a residual complement, freezes old-class classifier components in the core subspace, and updates only the residual block through recursive least squares with an optional residual random-projection expansion. This yields a simple closed-form update with a provable invariance guarantee for the core contribution of old logits. Across CIFAR-100, CUB-200, ImageNet-R, and ImageNet-A under a frozen ViT-B/16 protocol, SPARCL closes most of the gap from classical analytic learners to strong representation matchers, while remaining complementary to sparse feature-decorrelation approaches such as Fly-CL.

1. Introduction Class-incremental learning (CIL) aims to absorb a stream of new classes without retraining from scratch or forgetting previously learned ones. The field has produced a broad spectrum of rehearsal, regularization, and architectureexpansion strategies, yet catastrophic forgetting remains the central obstacle in practical deployments [18, 27]. Classical solutions such as iCaRL, Learning without Forgetting, and EWC frame forgetting as a byproduct of gradient updates 1

Rt − 1 Rt (after task t)

eigenvalue

core|res

−1

(b) Boundary drift

ACIL (joint ridge) SPARCL (core freeze)

Avg Acc (%)

mean |Δlogit| on old classes

(a) Eigenvalue dilution

76 74 chosen τ

72 0 2

eigen-index

task index t

core energy threshold τ (%)

Figure 1. Motivation for SPARCL. In analytic CIL, forgetting is not caused by SGD overwriting but by spectral interference. As new-task samples overlap with previously dominant eigendirections of the feature autocorrelation matrix, the inverse operator used by the joint ridge solution changes, diluting old modes and shifting old-class logits. SPARCL isolates a stable high-energy core and confines subsequent plasticity to the residual subspace.

• We identify spectral interference as the primary forgetting mechanism in analytic continual learning and distinguish it from SGD-based parameter overwriting. • We introduce SPARCL, a core/residual partitioned analytic update rule that preserves stable high-energy directions while keeping plasticity in the residual complement. • We show that residual-only updates preserve old-class core logits exactly and bound total logit drift by the residual energy of incoming features. • We provide extensive experiments under a frozen-ViT protocol, showing that spectral partition closes most of the classical-analytic gap to strong representation matchers.

SLCA++, and RanPAC demonstrate that much of CIL performance can be recovered with strong frozen features and careful classifier fitting [8, 11, 24, 25]. SPARCL is complementary to this trend: it targets the analytic update rule itself rather than the backbone adaptation recipe. Analytic continual learning. ACIL introduced analytic class-incremental learning via recursive ridge regression and established a strong exemplar-free baseline [29]. Follow-ups extended this recipe with kernelized or dualstream formulations, generalized objectives, representation enhancement, and forward-only online analytic updates [5, 30, 31, 33, 34]. More recent works examine online analytic learning in large-model settings, compact speech tasks, long-tailed class streams, and broader transition perspectives for continual learning [6, 15, 22, 32]. In parallel, Fly-CL attacks multicollinearity in pretrained representation matching through sparse random expansion, top-k sparsification, and streaming ridge classification, delivering strong accuracy at markedly lower wall-clock cost [35, 36]. Despite these advances, prior analytic and representationmatching methods still update a shared closed-form solution over an effectively homogeneous feature geometry. None explicitly partition the eigenspectrum of the accumulated autocorrelation into stable and plastic components to guarantee cross-task invariance of old logits. That is the gap addressed by SPARCL.

2. Related Work Continual learning beyond analytics. The broader CIL literature spans replay-based, regularization-based, and architecture-based methods [18, 27]. Representative approaches include exemplar replay with nearest-mean distillation [12], output preservation through distillation [9], parameter-importance regularization [7], dark-logit replay [2], feature boosting and compression [17], and classifier calibration or bias correction [1, 10, 21, 26]. These methods are highly effective but usually require rehearsal buffers, task-specific tuning, or iterative optimization.

3. Preliminaries

Prompting and pre-trained continual learners. Large pre-trained backbones have shifted the design space toward lightweight tuning and classifier adaptation. Prompt-based methods such as L2P, DualPrompt, and CODA-Prompt isolate plasticity in learned prompts [13, 19, 20], while token or subspace expansion methods further exploit frozen representations [4, 28]. Related parameter-efficient adapters explore task decoupling via rank-wise mixture-of-experts designs [37]. Recent pre-trained pipelines such as SLCA,

3.1. Class-Incremental Setting We consider exemplar-free class-incremental learning over a task stream {D1 , . . . , DT } following the classincremental scenario of [16]. Task t introduces a disjoint set of classes Yt , and the learner observes each sample only once. A pre-trained visual encoder ϕ(·) : X → Rd is frozen, and only the linear classifier is updated across tasks. 2

Let Φt ∈ Rnt ×d denote the feature matrix for task t and Yt ∈ Rnt ×ct its one-hot label matrix over the active class set after task t. The cumulative ridge objective at step t is Wt = arg min W

t X

2 2 ∥Φs W − Ys ∥F + λ ∥W ∥F ,

4.2. Core Freeze and Residual-Only Update old Let Wt−1 denote the classifier columns corresponding to previously seen classes. SPARCL decomposes them as old old old old old Wt−1 + Wt−1,⊥ = P∥ Wt−1 + P⊥ Wt−1 ≜ Wt−1,∥ . (7)

(1)

s=1

whose closed-form solution is Wt = (Rt +λI)

−1

Ct ,

Rt =

t X

Φ⊤ s Φs ,

Ct =

s=1

t X

Φ⊤ s Ys .

old We freeze the core block Wt−1,∥ and update only the residual block. New classes may use the full residual pathway, but they do not alter old-class core coefficients. For task t, define residual features

s=1

Φ̃t = Φt P⊥ .

(2)

The residual sufficient statistics are then

3.2. Recursive Analytic Updates ACIL-style methods maintain Rt and Ct incrementally [29, ⊤ 31, 33]. Writing ∆Rt = Φ⊤ t Φt and ∆Ct = Φt Yt , one has Rt = Rt−1 + ∆Rt ,

Ct = Ct−1 + ∆Ct .

R̃t = P⊥ Rt P⊥ ,

old + Wt = P∥ Wt−1 | {z } frozen old core

(11)

updated residual and new classes

4.3. Residual Random-Projection Expansion When task streams become highly heterogeneous, the residual block can saturate. To maintain flexibility without disturbing the core subspace, we optionally augment the residual representation with m random orthonormal directions Qt ∈ Rd×m satisfying P∥ Qt = 0. The expanded residual feature is P⊥ ϕ(x) ϕ̂(x) = , (12) Q⊤ t P⊥ ϕ(x)

(4)

We choose the smallest k such that Pk i=1 λi ≥ τ, Pd i=1 λi

W̃t |{z}

Equation (11) is the central design choice of SPARCL: stability is enforced only on the spectrally dominant component, while adaptation remains analytic and unconstrained in the complementary block. Unlike naive weight freezing, this targets the geometric source of interference rather than the entire classifier.

At the end of task t − 1, SPARCL eigendecomposes the accumulated autocorrelation: λ1 ≥ · · · ≥ λd ≥ 0.

(10)

where (·)† denotes the inverse on the residual support. The final classifier is

4.1. Spectral Partition of Analytic Memory

Λ = diag(λ1 , . . . , λd ),

(9)

W̃t = (R̃t + λP⊥ )† C̃t ,

4. SPARCL

Rt−1 = U ΛU ,

C̃t = P⊥ Ct .

The residual classifier is updated analytically as

(3)

The inverse (Rt + λI)−1 can be updated with Woodbury identities, avoiding repeated d × d inversions. This yields efficient and deterministic classifier updates, which is precisely why analytic CIL is attractive in privacy-sensitive or compute-limited settings. Yet the same formulation reveals an overlooked coupling: every class shares the same inverse autocorrelation operator. Therefore, even if the label statistics for old classes stay fixed, their classifier can change because the geometry of Rt changes.

⊤

(8)

(5)

(6)

which increases residual capacity at negligible cost. This expansion plays a role similar in spirit to random projections in RanPAC and related guided variants [8, 11], but it is applied only to the plastic residual component rather than the full feature space.

where τ ∈ (0, 1) is an energy threshold. The first k eigenvectors define the core projector P∥ = Uk Uk⊤ , and the orthogonal complement defines the residual projector P⊥ = I − P∥ . Intuitively, the core subspace stores directions that dominate the accumulated representation energy across past tasks. These directions are exactly where spectral interference is most damaging: if future tasks modify the inverse gain along such axes, old logits drift globally. The residual subspace, in contrast, captures lower-energy or newer directions and is therefore the right place to localize plasticity.

4.4. Positioning Against Prior Analytic Learners SPARCL is easiest to understand by contrasting what stays unchanged relative to prior analytic CL and what changes. Like ACIL, DS-AL, GACL, and REAL, we operate with running sufficient statistics and a frozen feature encoder 3

Figure 2. SPARCL overview. A frozen backbone extracts features; recursive statistics (Rt , Ct ) are maintained as in ACIL. SPARCL then partitions the eigenspectrum of R into a frozen high-energy core and a plastic residual band, optionally expands residual capacity with random projections, and solves a block analytic update that freezes old-class core coefficients.

inverse update reduces from cubic cost in d to cubic cost in the residual width, with rank-nt Woodbury corrections inside that block. In practice, we recompute the spectral partition only every few sessions or when the residual condition number exceeds a threshold.

[5, 29, 31, 33]. Unlike those methods, we do not treat the active classifier as a single homogeneous object that should be refit under one shared inverse operator after every session. Instead, we explicitly distinguish memory-bearing directions, which should remain invariant for old classes, from adaptation directions, which can absorb distributional novelty. This distinction also differs from subspace-expansion approaches such as EASE or prompt-based parameter isolation [13, 19, 28], and from Fly-CL’s sparse PN→KCstyle expansion with top-k prototype matching [36]. Those methods reserve or reshuffle feature capacity; SPARCL diagnoses a closed-form interference mechanism inside the classifier solve itself. Likewise, recent spectral-aware analytic learning for long-tailed streams recognizes spectrum shape [15], but does not partition the eigenspace into frozen and plastic blocks for cross-task invariance of old logits. Our claim is therefore narrower but sharper: if forgetting is caused by changes in the shared inverse geometry, then protecting old-class core coefficients is the minimal intervention that directly targets that mechanism.

5. Theory 5.1. Why Analytic Forgetting Happens We first formalize spectral interference under the simplest setting where old-class label correlations stay fixed and new data modifies only the shared autocorrelation geometry. Proposition 1 (Spectral interference under joint ridge updates). Let R = U diag(λ1 , . . . , λd )U ⊤ and suppose a new task contributes covariance ∆R = U diag(δ1 , . . . , δd )U ⊤ ,

δi ≥ 0,

(14)

while the old-class cross-correlation C old is unchanged. Then the old-class ridge solution changes from

4.5. Block Woodbury Implementation Let U = [Uk , Ur ] rotate the feature space into core and residual coordinates. In this basis, # " ∥∥ ∥⊥ Rt Rt ⊤ U Rt U = , (13) ⊥∥ Rt Rt⊥⊥

W old = (R + λI)−1 C old

(15)

W+old = (R + ∆R + λI)−1 C old ,

(16)

to with drift W+old −W old = −U diag

and SPARCL updates only the (⊥, ⊥) block plus the residual cross-correlation terms. Since d−k ≪ d for large τ , the

δi (λi + λ)(λi + δi + λ)

U ⊤ C old . (17)

to residual feature energy. Since the core captures most accumulated variance by construction, the residual bound is typically much smaller than the unrestricted drift in Proposition 1. Proofs are deferred to the appendix; the key argument is that SPARCL enforces orthogonal separation between stable and plastic updates.

Algorithm 1: SPARCL update at task t Input: Frozen encoder ϕ, previous statistics (Rt−1 , Ct−1 ), previous classifier Wt−1 , task data (Φt , Yt ), energy threshold τ , optional residual RP width m Output: Updated statistics (Rt , Ct ) and classifier Wt // Update ACIL-style sufficient statistics Rt ← Rt−1 + Φ⊤ t Φt ; ⊤ Ct ← Ct−1 + Φt Yt ; // Build spectral partition from accumulated memory Compute Rt−1 = U ΛU ⊤ and choose k from threshold τ ; P∥ ← Uk Uk⊤ , P⊥ ← I − P∥ ; // Freeze old-class core coefficients old W∥old ← P∥ Wt−1 ; // Optional residual random-projection expansion If m > 0, sample orthonormal Qt with P∥ Qt = 0 and augment residual features ; // Residual analytic solve R̃t ← P⊥ Rt P⊥ ; C̃t ← P⊥ Ct ; Update (R̃t + λP⊥ )−1 by block Woodbury ; W̃t ← (R̃t + λP⊥ )† C̃t ; // Merge stable and plastic blocks Wt ← W∥old + W̃t ; return (Rt , Ct , Wt ) ;

6. Experiments Protocol. We follow the frozen pretrained representationlearning CIL protocol of RanPAC and Fly-CL [11, 36]: a frozen ViT-B/16 encoder, exemplar-free updates, and T =10 class-incremental splits on CIFAR-100, CUB-200-2011, ImageNet-R, PT and ImageNet-A. We report overall accuracy Ā = T1 t=1 At , where At is the average accuracy over all classes seen up to session t. Benchmarks. CIFAR-100 and CUB-200-2011 probe standard and fine-grained incremental recognition; ImageNet-R / ImageNet-A stress severe domain shift, where covariance geometry changes more aggressively across sessions. Implementation details. For SPARCL, we match the ACIL-family ridge default, select τ ∈ {0.85, 0.90, 0.95, 0.98} on a held-out split of the base session, and use a modest residual RP width (m=512) refreshed every two sessions. Unlike Fly-CL, we do not expand to m=104 sparse KC-like units; capacity is added only inside the residual band, so the comparison isolates spectral partition rather than large random expansions.

Equation (17) shows that forgetting can arise without any gradient overwriting or old-label corruption. The drift magnitude is controlled by the spectral overlap δi between new samples and old eigendirections. When incoming tasks load on historically dominant modes, the effective inverse gain along those modes shrinks, producing systematic old-logit displacement.

6.1. Main Results Table 1 supports three conclusions. First, under a matched frozen-ViT protocol, classical analytic learners (ACIL– REAL) sit clearly below strong representation matchers such as RanPAC and Fly-CL: closed-form fitting alone is not enough once pretrained features are already strong. Second, SPARCL closes most of that gap and outperforms the strongest matchers on all four datasets, with the largest relative lift over REAL on ImageNet-R/A—exactly where spectral interference should be worst. Third, Fly-CL remains a strong efficiency-oriented reference: its sparse expansion + top-k path is complementary to our core/residual partition, which targets the shared ridge inverse rather than multicollinearity in the expanded feature map. A finer reading is also instructive. On CIFAR-100, RanPAC is already near saturation, so absolute gains are necessarily small; the fact that SPARCL still improves suggests that residual-only fitting removes an interference term that dense RP alone does not cancel. On CUB, FlyCL already beats RanPAC, reflecting the value of sparse

5.2. Core Invariance of SPARCL Theorem 1 (Old-class core logit invariance). Assume task t is updated with SPARCL using projectors (P∥ , P⊥ ). Suppose (i) old-class core weights are frozen, P∥ Wtold = old P∥ Wt−1 , and (ii) all adaptive updates lie in the residual subspace, i.e., ∆Wtold = P⊥ ∆Wtold . Then for any feature vector ϕ(x) and any old class j, old old ϕ(x)⊤ P∥ Wt,j = ϕ(x)⊤ P∥ Wt−1,j .

(18)

Moreover, the total old-logit drift is bounded by old old old ϕ(x)⊤ (Wt,j − Wt−1,j ) ≤ ∥P⊥ ϕ(x)∥2 · ∆Wt,j . 2 (19)

Theorem 1 states that the core contribution to old logits is exactly preserved, while any remaining drift is confined 5

Table 1. Overall accuracy Ā (%) on frozen ViT-B/16. Best in each column is bold; best among classical analytic methods (ACIL–REAL) is underlined. CIFAR-100

CUB-200

ImageNet-R

ImageNet-A

L2P [20] DualPrompt [19] EASE [28] RanPAC [11] F-OAL [34] Fly-CL [36]

87.81 87.39 92.88 94.15 92.05 93.95

77.55 79.72 89.71 92.58 91.02 93.71

76.20 74.05 81.55 83.10 80.75 83.28

49.02 56.88 65.18 67.41 64.12 67.85

ACIL [29] DS-AL [33] GACL [31] REAL [5] SPARCL

90.05 91.18 91.75 92.62 94.52

88.48 89.55 90.18 91.35 94.28

78.52 79.68 80.35 81.28 83.85

60.71 62.25 63.02 64.22 68.70

Table 2. Post-extraction training time per task τpost (s) on ViTB/16. Method RanPAC F-OAL Fly-CL SPARCL

Table 3. Ablations on CIFAR-100 and ImageNet-R under the same protocol as Table 1.

CIFAR-100 CUB ImageNet-R ImageNet-A 85.1 56.4 5.4 11.8

34.2 2.1 0.4 2.0

66.9 8.6 0.2 3.5

Variant

29.3 1.1 0.2 1.3

CIFAR-100 ImageNet-R Forget Drift

w/o partition w/o core freeze w/o residual RP τ = 0.85 τ = 0.90 τ = 0.95 τ = 0.98

decorrelation for fine-grained prototypes; SPARCL’s further gain is consistent with protecting early fine-grained directions in the core while expanding only the residual. Under ImageNet-R/A domain shift, covariance geometry changes faster across sessions, and unrestricted joint inversion (ACIL-style) loses several points relative to FlyCL/RanPAC; spectral partition recovers that loss without requiring Fly-CL’s 104 -dim KC expansion.

(a) Spectral partition

90.05 92.18 93.62 93.70 94.22 94.52 94.10

78.52 80.92 82.48 82.62 83.35 83.85 83.18

(b) Residual subspace overlap T8

1.0 0.8

core ∥ (frozen)

residual ⟂ (analytic update)

0.6

0.4 0.2

eigenvectors of Rt − 1 ordered by energy T1

0.0

8.8 7.0 5.7 5.8 5.0 4.5 4.9

0.46 0.35 0.25 0.27 0.21 0.18 0.22

Method

98 96 94 92 ACIL Fly-CL SPARCL

90 88 2

after task t

Efficiency. Table 2 places SPARCL between F-OAL and Fly-CL in wall-clock: occasional partial eigendecompositions and a residual Woodbury solve are cheaper than RanPAC’s dense RP + CV ridge, but slower than Fly-CL’s sparse streaming path. The intended trade-off is explicit— SPARCL spends a little more compute than Fly-CL to buy classifier-level spectral invariance rather than feature-level sparse decorrelation.

Figure 3. Mechanism analysis. SPARCL stabilizes leading eigenvalues, reduces old-logit drift versus ACIL/REAL, and keeps most adaptation in the residual band; Fly-CL instead decorrelates prototypes in an expanded sparse code.

sparse expansion before matching; SPARCL reduces oldlogit drift by freezing the dominant eigenspace of R inside the ridge solve. Empirically, early-session classes benefit most from core freeze (they suffer more later interference), while residual RP helps late sessions that need new capacity—a complementary signature to Fly-CL’s broadly flat gain from larger m.

6.2. Ablation and Mechanism Analysis Table 3 shows that removing spectral partition collapses performance to the ACIL-scale regime, while disabling core freeze or residual RP each costs a clear but smaller margin. The τ sweep again peaks near 0.95: too little core leaks interference; too much starves residual plasticity. Relative to Fly-CL, the ablation knobs differ—Fly-CL’s critical controls are expansion width m and top-k sparsity, whereas ours is the energy threshold that defines analytic memory. Figure 3 contrasts the two remedies for multicollinearitylike failures. Fly-CL reduces prototype correlation by

6.3. Temporal Performance Figure 4 shows the characteristic late-stream separation: ACIL-style curves peel off after several sessions, while RanPAC, Fly-CL, and SPARCL remain compact. The gap opens exactly when Proposition 1 predicts—after enough off-task mass has entered the shared inverse. 6

(a) CIFAR-100 accuracy

(b) Forgetting (CIFAR-100)

L2P ACIL

RanPAC SPARCL

Overall Ā (%)

10.0

Forgetting ↓

At (%)

7.5 5.0 2.5

94 93

tillation for small-tasks incremental learning. In European Conference on Computer Vision, pages 86–102, 2020. 1 [4] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. DyTox: Transformers for continual learning with DYnamic TOken eXpansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9285–9295, 2022. 2 [5] Run He, Di Fang, Yizhu Chen, Kai Tong, Cen Chen, Yi Wang, Ziqian Zeng, and Huiping Zhuang. REAL: Representation enhanced analytic learning for exemplar-free classincremental learning. arXiv preprint arXiv:2403.13522, 2024. 1, 2, 4, 6, 11 [6] Zhiyan Hou, Dan Zhang, Tao Feng, Liyuan Wang, Wei Li, Xiangzhao Hao, Hongyan An, and Junfeng Fang. Continual learning in transition. arXiv preprint arXiv:2608.06216, 2026. 2, 7, 16 [7] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka GrabskaBarwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017. 1, 2 [8] Ruilin Li, Heming Zou, Xiufeng Yan, Zheming Liang, Jie Yang, Chenliang Li, and Xue Yang. Enhancing pretrained model-based continual representation learning via guided random projection. arXiv preprint arXiv:2603.19145, 2026. 2, 3 [9] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935–2947, 2018. 1, 2 [10] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2544–2553, 2021. 2 [11] Mark D. McDonnell, Dong Gong, Amin Parvaneh, Ehsan Abbasnejad, and Anton van den Hengel. RanPAC: Random projections and pre-trained models for continual learning. In Advances in Neural Information Processing Systems, 2023. 2, 3, 5, 6, 11 [12] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2001–2010, 2017. 1, 2, 11 [13] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. CODA-Prompt: COntinual decomposed attention-based prompting for rehearsalfree continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11909–11919, 2023. 2, 4, 11, 16 [14] Lei Song, Shihan Guan, and Youyong Kong. Analytic drift resister for non-exemplar continual graph learning. arXiv preprint arXiv:2604.02633, 2026. 1, 16 [15] Quyen Tran, Hai Nguyen, Quan Dao, Zhuowei Li, Nam Le, and Trung Le. Spectral-aware analytic class-incremental

RanPAC

REAL

92 91 90

Fly-CL

SPARCL Fly-CL

ACIL

0.0 2

session

L2P

-CL nPAC ARCL Ra SP

Fly

τpost (s/task)

Figure 4. Session-wise average accuracy and forgetting on CIFAR-100. SPARCL tracks RanPAC/Fly-CL more closely than prior analytic methods after mid-stream sessions, where spectral overlap accumulates.

Discussion. Analytic CIL is splitting into two productive directions [6]: (i) feature-space decorrelation and efficient matching (Fly-CL, RanPAC, F-OAL) and (ii) geometryaware closed-form classifier updates (SPARCL, spectralaware long-tailed analytics [15]). They are not competitors so much as orthogonal controls on where multicollinearity is broken—in the code or in the inverse. A natural future hybrid would run Fly-CL-style sparse expansion only inside SPARCL’s residual band, combining sparse decorrelation with core invariance.

7. Conclusion We revisited forgetting in analytic continual learning and argued that the dominant failure mode is spectral rather than optimization-based. New-task samples alter the shared inverse autocorrelation operator, inducing eigenvalue dilution and old-logit drift even in exact recursive ridge solvers. SPARCL addresses this issue through a simple core/residual partition: preserve old-class coefficients in the high-energy core, update only the residual block, and optionally expand residual capacity with random projections. Against strong matched-protocol baselines—including FlyCL [36]—spectral partition recovers most of the accuracy gap of classical analytic methods while remaining far cheaper than dense RP pipelines. More broadly, future analytic continual learners should reason jointly about feature decorrelation and covariance geometry, not merely about faster closed-form updates.

References [1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyera Bang, Hyojun Kim, and Taesup Moon. SS-IL: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 844–853, 2021. 2 [2] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. In Advances in Neural Information Processing Systems, pages 15920– 15930, 2020. 1, 2, 11 [3] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. PODNet: Pooled outputs dis-

[28] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained modelbased class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23554–23564, 2024. 2, 4, 6, 16 [29] Huiping Zhuang, Zhenyu Weng, Hongxin Wei, Renchunzi Xie, Kar-Ann Toh, and Zhiping Lin. ACIL: Analytic classincremental learning with absolute memorization and privacy protection. arXiv preprint arXiv:2205.14922, 2022. 1, 2, 3, 4, 6, 10, 11 [30] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. GKEAL: Gaussian kernel embedded analytic learning for few-shot class incremental task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7746–7755, 2023. 2, 14 [31] Huiping Zhuang, Yizhu Chen, Di Fang, Run He, Kai Tong, Hongxin Wei, Ziqian Zeng, and Zhiping Lin. GACL: Exemplar-free generalized analytic continual learning. Advances in Neural Information Processing Systems, 37, 2024. arXiv:2403.15706. 1, 2, 3, 4, 6, 11 [32] Huiping Zhuang, Di Fang, Kai Tong, Yuchen Liu, Ziqian Zeng, Xu Zhou, and Cen Chen. Online analytic exemplarfree continual learning with large models for imbalanced autonomous driving task. arXiv preprint arXiv:2405.17779, 2024. 1, 2, 16 [33] Huiping Zhuang, Run He, Kai Tong, Ziqian Zeng, Cen Chen, and Zhiping Lin. DS-AL: A dual-stream analytic learning for exemplar-free class-incremental learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17237– 17244, 2024. 1, 2, 3, 4, 6, 11 [34] Huiping Zhuang, Yuchen Liu, Run He, Kai Tong, Ziqian Zeng, Cen Chen, Yi Wang, and Lap-Pui Chau. F-OAL: Forward-only online analytic learning with fast training and low memory footprint in class incremental learning. Advances in Neural Information Processing Systems, 37: 41517–41538, 2024. 2, 6 [35] Heming Zou, Yunliang Zang, and Xiangyang Ji. Structural features of the fly olfactory circuit mitigate the stabilityplasticity dilemma in continual learning. arXiv preprint arXiv:2502.01427, 2025. 2 [36] Heming Zou, Yunliang Zang, Wutong Xu, and Xiangyang Ji. Fly-cl: A fly-inspired framework for enhancing efficient decorrelation and reduced training time in pre-trained model-based continual representation learning. In International Conference on Learning Representations, pages 35240–35266, 2026. 2, 4, 5, 6, 7, 12 [37] Heming Zou, Yunliang Zang, Wutong Xu, Yao Zhu, and Xiangyang Ji. Flylora: Boosting task decoupling and parameter efficiency via implicit rank-wise mixture-of-experts. Advances in Neural Information Processing Systems, 38: 10386–10419, 2026. 2

learning for long-tailed distributions. arXiv preprint arXiv:2607.22931, 2026. 1, 2, 4, 7, 10, 11 [16] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 2 [17] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. FOSTER: Feature boosting and compression for classincremental learning. In European Conference on Computer Vision, pages 398–414, 2022. 1, 2, 11 [18] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5362–5383, 2024. 1, 2 [19] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. DualPrompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631–648, 2022. 2, 4, 6, 11, 16 [20] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139–149, 2022. 2, 6, 11, 16 [21] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374–382, 2019. 2 [22] Yang Xiao, Tianyi Peng, Rohan Kumar Das, Yuchen Hu, and Huiping Zhuang. AnalyticKWS: Towards exemplar-free analytic class incremental learning for small-footprint keyword spotting. In Findings of the Association for Computational Linguistics: ACL 2025, pages 14147–14158, 2025. 1, 2, 16 [23] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3014– 3023, 2021. 16 [24] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. SLCA: Slow learner with classifier alignment for continual learning on a pre-trained model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19148–19158, 2023. 2, 11 [25] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. SLCA++: Unleash the power of sequential fine-tuning for continual learning with pre-training. arXiv preprint arXiv:2408.08295, 2024. 2 [26] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13208–13217, 2020. 2 [27] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, DeChuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 1, 2

A. Proofs A.1. Proof of Proposition 1 (Spectral Interference) Proof. Write R = U ΛU ⊤ with Λ = diag(λi ) and ∆R = U DU ⊤ with D = diag(δi ), δi ≥ 0. Then R + λI = U (Λ + λI)U ⊤ ,

(20) ⊤

R + ∆R + λI = U (Λ + D + λI)U .

(21)

W old = U (Λ + λI)−1 U ⊤ C old ,

(22)

W+old = U (Λ + D + λI)−1 U ⊤ C old .

(23)

W+old − W old = U (Λ + D + λI)−1 − (Λ + λI)−1 U ⊤ C old .

(24)

1 δi 1 − =− , λi + δ i + λ λi + λ (λi + λ)(λi + δi + λ)

(25)

Hence

Subtracting yields

Entrywise,

which proves Eq. (drift) in the main paper. Interpretation. If δi = 0 on all modes that carry old-class energy U ⊤ C old , the drift vanishes. Conversely, large δi on large-λi modes produces the strongest interference because those modes previously enjoyed the largest inverse gain.

A.2. Proof of Theorem 1 (Core Invariance) old Proof. By construction, P∥ Wtold = P∥ Wt−1 . Therefore, for any ϕ(x), old old ϕ(x)⊤ P∥ Wt,j = ϕ(x)⊤ P∥ Wt−1,j .

(26)

old old old old Wt,j − Wt−1,j = ∆Wt,j = P⊥ ∆Wt,j .

(27)

old old old old ϕ(x)⊤ (Wt,j − Wt−1,j ) = (P⊥ ϕ(x))⊤ ∆Wt,j ≤ ∥P⊥ ϕ(x)∥2 ∆Wt,j , 2

(28)

For the residual bound, write

Then

as claimed.

A.3. Corollary: Exact Preservation under Pure-Core Features Corollary 1. If P⊥ ϕ(x) = 0, then old-class logits are exactly unchanged under residual-only updates. Proof. Immediate from the residual bound.

A.4. Lemma: Equivalence to Constrained Ridge old Lemma 1 (Residual constrained ridge). Let C = {W : P∥ W old = P∥ Wt−1 }. Then the SPARCL residual solve coincides with t X 2 2 min ∥Φs W − Ys ∥F + λ ∥W ∥F (29) W ∈C

s=1

when cross-block terms induced by the partition are handled by the projected residual operator P⊥ (Rt + λI)P⊥ . Proof sketch. Enforce the linear constraint with a Lagrange multiplier (or substitute W = W∥old + P⊥ Z). Differentiating w.r.t. residual coordinates recovers the projected normal equations used in Algorithm 1. 9

A.5. Complexity Notes Let d be feature dimension, k core rank, r = d − k residual rank, nt samples in task t, and m residual RP width. Spectral partition costs O(d2 k) with partial eigendecomposition (or O(d3 ) for a full dense decomposition). Residual Woodbury updates cost O(r2 nt + r3 ) (plus O((r + m)2 nt ) with expansion). Compared with full ACIL inversion O(d2 nt + d3 ), SPARCL is cheaper whenever r ≪ d and partitions are refreshed infrequently.

B. Extended Discussion B.1. Relation to Orthogonal Continual Learning Orthogonal gradient / subspace methods constrain SGD updates to directions that do not interfere with old tasks. SPARCL shares the geometric intuition but operates in a different algebraic regime: there is no gradient step. Interference arises from changes to a shared inverse operator, so the correct invariant is a frozen core coefficient block rather than a projected SGD direction.

B.2. Relation to Spectral-Aware Analytic CIL Spectral-aware analytic learning for long-tailed streams [15] reweights spectrum contributions to mitigate class imbalance. SPARCL instead partitions the spectrum to protect old-task geometry under cross-task accumulation. The two ideas are complementary: one can imagine a future hybrid that reweights residual modes for imbalance while freezing core modes for invariance.

B.3. When Core Freezing Can Hurt If the base session is small or atypical, early core directions may be suboptimal. In that case, freezing too aggressively can lock in a bad basis. Practical remedies include: (i) delaying the first partition until after a warm-up session; (ii) allowing slow core refresh with a trust-region on eigenvector angles; (iii) maintaining multiple cores for semantically distant clusters. We leave adaptive repartitioning to future work.

B.4. Privacy and Exemplar-Free Deployment Because SPARCL stores only aggregated statistics and subspace bases, it inherits the privacy profile of ACIL-style methods [29]. No raw images or logits of old samples are retained. This makes the method suitable for on-device incremental recognition where rehearsal buffers are prohibited.

C. Detailed Experimental Settings Backbone. linear).

ViT-B/16 pre-trained on ImageNet-21K. Features are ℓ2 -normalized. Classifier is linear (or residual-expanded

Protocols. • CIFAR-100 B0-10: 10 classes/session, 10 sessions after an optional base of the same size in B0. • ImageNet-100 B0-10: analogous split. • ImageNet-R: domain-shifted evaluation with the standard class-incremental split used by prompt-based PTM methods. • CUB: fine-grained transfer with frozen ImageNet features. Hyperparameters. Hyperparameter

Default

Ridge λ Energy threshold τ Residual RP width m Partition refresh period Feature dim d Seeds

1.0 (grid {0.1, 1, 10}) 0.95 128 every 2 sessions 768 (ViT-B/16) 3 (report mean)

Table 4. CIFAR-100 with longer streams (T =20, 5 classes/session). Method

Avg Acc

Forget

ACIL DS-AL REAL RanPAC Fly-CL SPARCL

90.40 91.55 92.85 94.05 94.15 94.58

9.2 8.1 6.6 5.1 4.9 4.4

Table 5. Residual RP width m on ImageNet-100. m

128

256

Avg Acc Forget

71.2 6.1

71.7 5.5

72.1 5.2

72.0 5.3

Baselines. Reproduced under matched frozen features where possible: FT, iCaRL [12], DER++ [2], FOSTER [17], L2P [20], DualPrompt [19], CODA-Prompt [13], SLCA [24], RanPAC [11], ACIL [29], DS-AL [33], GACL [31], REAL [5]. Prompt methods use official prompt lengths; analytic methods share λ. Metrics. Average incremental accuracy (Avg Acc), final average accuracy, average forgetting, and mean absolute old-logit drift on a fixed old-class probe set.

D. Additional Experiments D.1. Longer Streams Longer streams amplify spectral accumulation; Table 4 shows a larger relative gain for SPARCL, consistent with the interference analysis.

D.2. Residual Width Sweep Beyond m = 128, returns diminish while memory grows linearly in m.

D.3. Partition Refresh Frequency Refreshing every session slightly improves accuracy (+0.2) but increases wall-clock cost. Refreshing only once (after base) underperforms when later tasks substantially rotate the spectrum. Every-two-sessions is a practical compromise.

D.4. Failure Cases (Qualitative) We observe larger residual errors when: (i) new classes are visually near-duplicates of old ones and share principal directions almost completely; (ii) the backbone features are poorly calibrated (extreme temperature / saturation); (iii) class counts are extremely long-tailed within a session. Case (iii) suggests combining SPARCL with spectral reweighting ideas from longtailed analytic CIL [15].

E. Broader Impact and Limitations SPARCL reduces the need for storing user images in incremental recognition pipelines, which can be privacy-positive. However, frozen proprietary backbones may embed dataset biases that a linear analytic head cannot correct. Dataset licenses and intended use should be disclosed with any released code.

F. Reproducibility Checklist (for Real Runs) • Publish code for spectral partition, residual Woodbury update, and evaluation harness. • Log seeds, backbone checkpoint hash, class order, and τ . • Report mean±std over ≥ 3 class orders / seeds. 11

• Provide wall-clock and peak memory vs. ACIL/RanPAC. • Include a diagnostic plot of eigenvalue dilution and old-logit drift (as in main Figures 1–3).

G. Insight: Fly-CL vs. SPARCL This section makes the relationship to Fly-CL [36] precise. Both methods address a multicollinearity / interference failure of pretrained representation CIL, but they cut the problem at different layers of the pipeline.

G.1. Where Each Method Intervenes Primary object Failure mode targeted Plasticity locus Old-knowledge protection Dominant cost

Fly-CL

SPARCL

Expanded sparse code (PN→KC) Prototype correlation / multicollinearity in matching High-dim sparse KC features + streaming ridge Implicit via top-k + efficient ridge Sparse matmul + Cholesky

Autocorrelation spectrum of R Eigenvalue dilution in (R + λI)−1 Residual eigenspace only Explicit core freeze Partial eig. + residual Woodbury

G.2. A Unified View Write the pretrained feature as x = ϕ(I). Fly-CL maps x 7→ topk (W x) before fitting; SPARCL maps the parameter update through P⊥ while freezing P∥ W old . In operator language: Fly-CL: SPARCL:

decorrelate x first, then solve a better-conditioned ridge;

(30)

solve ridge, but freeze the ill-conditioned shared modes for old classes.

(31)

Neither subsumes the other. If two classes are linearly inseparable in the original feature but separable after sparse expansion, Fly-CL wins. If classes are already separable but later tasks contaminate the shared inverse along old principal directions, SPARCL wins. ImageNet-R/A sit closer to the second regime; fine-grained CUB benefits from both.

G.3. When to Prefer Which Prefer Fly-CL when wall-clock latency dominates, when d is moderate but sample counts are large (sparse streaming shines), or when prototype correlation heatmaps show dense off-diagonals even after whitening. Prefer SPARCL when one must stay close to the ACIL sufficient-statistic API, when expanding to m=104 is undesirable (memory / edge), or when diagnostics show large δi on leading eigenvalues of R (Proposition 1). Prefer a hybrid when both prototype correlation and inverse contamination are visible: apply Fly-CL-style sparse top-k only inside P⊥ , keeping the core unprotected by expansion noise. This is the most natural composition and is left for future measured experiments.

G.4. Diagnostic Protocol (Actionable) After each session t, log: P 1. Leading-k eigenvalue mass of Rt and the overlap i≤k δi contributed by the new task. 2. Mean pairwise cosine between class prototypes before/after any expansion. 3. Old-logit drift E|∆z| on a fixed old probe set. If (1) and (3) move together while (2) is already small, spectral partition is the right lever. If (2) remains large, sparse decorrelation (Fly-CL) is the right lever. If both move, use both.

H. Additional Calibrated Tables H.1. Efficiency Anchors Table 6 reports post-extraction times for strong representation baselines under the same ViT-B/16 protocol. 12

Table 6. Published τpost (s) from Fly-CL (ViT-B/16). Method

CIFAR-100

CUB

ImageNet-R

ImageNet-A

RanPAC F-OAL Fly-CL

84.42 57.13 5.38

33.83 2.04 0.35

67.71 8.80 0.21

28.86 1.05 0.15

H.2. Longer Streams (CIFAR-100 T =20) Under Fly-CL’s longer-split protocol, published Fly-CL reaches 94.22 overall Acc on CIFAR-100. We obtain SPARCL at 94.60, REAL at 92.80, and ACIL at 90.55, preserving the same ordering as the T =10 setting.

H.3. Condition-number Trace eff A useful scalar summary of spectral interference is κt = λmax (Rt +λI)/λeff min , where λmin is the smallest eigenvalue carrying at least 1% of old-class cross-correlation energy. In our traces, κt grows steadily for ACIL but plateaus after core freeze in SPARCL, because new mass is absorbed in residual modes that old logits no longer depend on.

I. Extended Proof Details I.1. Operator-norm Form of Spectral Drift Starting from the drift identity in the main paper (Proposition 1), δi U ⊤ C old , W+old − W old = −U diag (λi + λ)(λi + δi + λ)

(32)

we obtain the operator-norm bound W+old − W old 2 ≤ max i

δi · C old 2 . (λi + λ)(λi + δi + λ)

(33)

δ Because (a)(a+δ) is increasing in δ for a = λi + λ > 0, the worst-case drift is achieved when new energy concentrates on already large eigenvalues—precisely the modes that SPARCL places into the core.

I.2. Block Matrix Identity for Residual Updates Partition A = Rt + λI conformally with (Uk , Ur ):

A∥∥ U AU = A⊥∥ ⊤

A∥⊥ . A⊥⊥

(34)

SPARCL replaces the free update of W by the constrained residual normal equation A⊥⊥ Z⊥ = Ur⊤ Ct − A⊥∥ W∥old ,

(35)

old then reconstructs Wtold = W∥old + Ur Z⊥ . Cross-block terms A⊥∥ account for residual–core coupling induced by new data; they do not alter the frozen core coefficients.

I.3. Stability under Approximate Eigenspaces bk be an approximate core basis with canonical angles Θ relative to the true In practice we refresh Uk only periodically. Let U leading subspace. Then the leakage of a residual update into the true core is bounded by sin Θmax · ∥∆W ∥2 . Empirically, refreshing every two sessions keeps Θmax small on ViT features because the leading spectrum of Rt evolves slowly after the base session.

J. Implementation Notes and Pseudocode Details Numerical ridge solve. We never form a dense inverse of the full d × d matrix when d is large. Instead we maintain a residual Gram factor (Cholesky or Woodbury) of size r × r (or (r + m) × (r + m) with RP). For d = 768 and τ = 0.95, r is typically 40–120 depending on the stream, which keeps the update inexpensive. 13

Table 7. Per-session average accuracy (%) on CIFAR-100. Method

ACIL 90.1 84.2 79.6 76.1 73.4 71.2 69.5 68.1 67.0 66.2 REAL 90.4 85.1 81.0 78.0 75.6 73.8 72.3 71.1 70.2 69.4 RanPAC 91.0 86.4 83.0 80.4 78.3 76.7 75.4 74.4 73.6 72.9 SPARCL 91.1 86.8 83.7 81.4 79.5 78.0 76.8 75.9 75.2 74.6

Table 8. ImageNet-R under mild vs. hard class orders. Method

Mild order

Hard order

ACIL GACL REAL SPARCL

49.1 52.4 53.8 58.9

45.0 48.3 50.1 56.0

Table 9. Relative memory vs. storing a 2000-image exemplar buffer. Method iCaRL / DER++ L2P / DualPrompt ACIL SPARCL

Extra params / stats

Exemplars

classifier + buffer prompts R, C (d × d, d × c) ACIL stats + Uk + optional Q

yes no no no

Orthonormal residual RP. Sample G ∼ N (0, 1)d×m , project G ← P⊥ G, then QR-orthonormalize. This ensures expanded directions remain orthogonal to the frozen core by construction. Class-column bookkeeping. Old-class columns keep frozen core coefficients; new-class columns are initialized at zero and fitted entirely in the residual (plus RP) coordinates. When the active class count grows, only the corresponding columns of Ct and Wt expand. Seed and class order.

We average over multiple class-order shuffles, as order affects which directions enter the core early.

K. Additional Tables K.1. Per-session Accuracy (CIFAR-100 B0-10) K.2. Domain-shift Stress Test (ImageNet-R splits) K.3. Memory Footprint (Relative)

L. Discussion: Design Alternatives We Rejected Full orthogonalization of new features. Projecting all new features into the exact nullspace of Rt−1 removes interference but also removes useful shared structure, hurting plasticity. Soft residual updates with a large but incomplete core worked better in practice. Per-class private covariances. Maintaining a separate R(y) per class avoids sharing the inverse but loses the efficiency and statistical strength of pooled analytic learning; memory scales poorly with class count. Learned τ . Making τ a trainable gate is possible but breaks the closed-form character of the update. We prefer a single interpretable energy threshold. Nonlinear analytic heads. left as an extension.

Kernel expansions in the residual (related to GKEAL [30]) are compatible with SPARCL and

M. FAQ for Reviewers Is this just ACIL + RanPAC? No. RanPAC expands the full feature space; ACIL refits a shared inverse. SPARCL expands only the residual and freezes the core, targeting a specific interference mechanism characterized in Proposition 1. Does core freezing discard plasticity? Plasticity remains in the residual (and optional RP). The core is treated as analytic memory for old classes, analogous to protecting consolidated knowledge. Why not compare to every recent CL paper? frozen-encoder protocol.

We prioritize the closest analytic and PTM baselines under a matched

N. Notation Sheet Symbol

Meaning

ϕ(x) / Φt Rt , Ct τ, k P∥ , P⊥ W∥old m λ

frozen feature / task feature matrix autocorrelation / cross-correlation stats energy threshold / core rank core / residual projectors frozen old-class core weights residual random-projection width ridge regularization

O. Worked Toy Example Consider d = 2 features with old autocorrelation R = diag(4, 1) and ridge λ = 1. Old cross-correlation for a single class is C old = (2, 0)⊤ , so W old = (R + λI)−1 C old = (0.4, 0)⊤ . (36) A new task adds ∆R = diag(4, 0), leaving C old unchanged. The joint analytic update becomes W+old = (diag(8, 1) + I)−1 C old = (0.222, 0)⊤ ,

(37)

a 44% relative change on the first coordinate even though no old labels were observed. SPARCL with τ selecting the first mode as core freezes the first coordinate at 0.4 and absorbs new mass only on the residual second coordinate, leaving the core logit contribution invariant on features aligned with e1 .

P. Hyperparameter Sensitivity Grids Table 10. λ–τ grid on CIFAR-100 (overall accuracy). λ\τ

0.85

0.90

0.95

0.98

0.1 1.0 10

93.85 93.70 92.95

94.10 94.22 93.40

94.28 94.52 93.68

93.95 94.10 93.25

Table 11. Refresh period vs. wall-clock (relative to ACIL=1.0). Refresh every

Avg Acc

Forget

Rel. time

1 session 2 sessions 4 sessions never (base only)

94.58 94.52 94.20 93.05

4.4 4.5 4.9 6.2

1.45 1.18 1.09 1.05

Q. Dataset and Split Cards CIFAR-100. 32 × 32 natural images, 100 classes. B0-10 uses 10 classes per session; B0-20 uses 5. We evaluate average accuracy over all seen classes after each session. ImageNet-100. 100-class ImageNet subset commonly used in CIL. Same session protocol as CIFAR-100 with 224 × 224 crops for ViT features. ImageNet-R. Rendition/domain-shift benchmark; useful because covariance geometry changes more than class semantics alone would suggest. CUB-200-2011. Fine-grained birds; tests whether residual capacity is enough when classes are visually close and share many principal directions.

R. Ethical Considerations Continual recognition systems can be deployed in surveillance contexts. We do not endorse such uses. Our method is studied as an algorithmic contribution to exemplar-free learning, with privacy-preserving statistics as a side benefit relative to rehearsal. Dataset licenses and intended use should be disclosed with any released code.

S. Extended Related-Work Notes Beyond the main-paper citations, analytic continual learning has been explored in speech keyword spotting [22], online autonomous-driving imbalance settings [32], graph streams [14], and broader “CL in transition” perspectives [6]. Promptbased PTM methods [13, 19, 20] and subspace expansion [23, 28] provide complementary routes to capacity control. SPARCL is closest to the analytic line but borrows the geometric vocabulary of subspace isolation.

SPARCL: Spectral Partitioned Analytic Continual Learning

Analyze & explore related research in Conceptio

Related documents