acceptodds
Under review as a conference paper at ICLR 2027

CL-Mamba: Continual Learning Mamba Sub-Networks for Robust Cross-Domain High Efficient Hyperspectral Anomaly Detection

Abstract

Bio-inspired continual learning draws inspiration from biological neural systems that continuously acquire new knowledge while retaining previously learned information. In particular, the Drosophila mushroom body employs parallel and compartmentalized neural pathways to encode complementary information and reduce interference across experiences. Inspired by this mechanism, we investigate whether a collaborative multi-learner architecture can benefit hyperspectral anomaly detection (HAD) in continuously changing scenes, where models must adapt to new background distributions while avoiding catastrophic forgetting and excessive storage of historical samples. Therefore, we propose CL-SSMamba, a bio-inspired continual learning framework built upon multiple Spatial-Spectral Mamba (SSMamba) subnetworks for open-scenario hyperspectral anomaly detection (OHAD). CL-SSMamba employs parallel SSMamba learners to collaboratively reconstruct the hyperspectral background while retaining only model parameters from previous tasks, thereby reducing storage overhead and alleviating catastrophic forgetting. Each learner integrates Global Spatial Mamba (GSM) and Local Spectral Mamba (LSM) to capture long-range spatial dependencies and cross-band spectral correlations, together with orthogonal spectral correlation enhancement to suppress redundant responses. To alleviate background–anomaly imbalance, Simple Linear Iterative Clustering (SLIC) and Local Density Peak Clustering (LDPC) are introduced to enhance spatially coherent background structures and improve background–anomaly separation. In addition, reconstruction, ensemble cooperation, and continual-learning losses are jointly optimized to promote inter-learner consistency and preserve historical knowledge. Experiments on the Abu-Beach-Urban and HAD100 datasets demonstrate that CL-SSMamba effectively adapts to sequential hyperspectral scenes without replaying historical samples, providing a robust solution for OHAD in continuously changing environments.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual learning, Hyperspectral anomaly detection, Mamba

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QH7vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Inspired by Bio-inspiredcontinual learning, the authors attempt to investigate whether a collaborative multi-learner architecture can benefit hyperspectral anomaly detection(HAD) in continuously changing scenes. The authors propose CL-SSMamba,a bio-inspired continual learning framework built on multiple spatial-spectral Mamba subnetworks for open-scenario hyperspectral anomaly detection. The method is evaluated on a six-task ABU sequence and five HAD100 scenes, with comparisons to several continual HAD baselines as well as component and learner-number ablations.

Critical strengths

The paper addresses a practically relevant continual-learning setting in which sequential hyperspectral scenes must be adapted to without retaining historical raw HSIs. The framework based on Mamba subnetworks and parameter regularization provides a meaningful alternative to approaches relying on generative mechanisms or historical-data replay. The overall framework and experimental study are reasonably comprehensive. The method achieves strong final-stage performance on the six-task ABU sequence.

Critical weaknesses

1.The core multi-learner and continual-learning mechanisms are not sufficiently isolated. Increasing K from 1 to 2 also increases model capacity, but no parameter-matched single-learner baseline is provided. L_CL and L_EC are not individually ablated, and there is no direct evidence that the learners develop complementary representations. The update of parameter importance I and the value of λ are also insufficiently specified. The contribution of LSM is similarly unclear, as removing it slightly improves ACC and BWT at the final ABU stage.

2.The current evidence is insufficient to fully support the claims of continual-learning robustness and computational efficiency. Eq. (20) appears to measure mainly the final task transition rather than cumulative forgetting, while multi-seed and task-order evaluations are missing. CL-BioGAN also outperforms CL-SSMamba on all four final-stage HAD100 metrics, and the efficiency analysis is largely limited to parameter count and a reported 15% training-speed improvement over CL-CaGAN.

Questions for authors

1.How is the parameter importance I in L_CL computed and updated? What value of λ is used, and what additional state must be retained after each task?

2.Why is the learner aggregation weight defined as s = 1/T ? Does the method require the total number of tasks T to be known in advance?

3.Under a parameter-matched setting, does K=2 still outperform a widened single learner? How do removing L_CL and L_EC affect cumulative BWT and the full task-performance matrix?

4.Why does the benefit of LSM appear task-dependent? Since removing LSM slightly improves ACC and BWT at the final ABU stage, could the authors clarify under what conditions LSM is actually beneficial?

5.Could the authors provide a more complete efficiency evaluation, including training/inference time.

6.Appendix A.4.5 describes Fig. 13 as visualization results on the five sequential HAD100 tasks, whereas the figure caption refers to “five ABU datasets” and lists ABU scenes such as Los Angeles, Bay Champagne, San Diego, and Gulfport. Could the authors clarify which dataset and tasks Fig. 13 actually corresponds to and correct the inconsistency between the text and the figure caption?

Reviewer_ZQkmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CL-SSMamba for continual hyperspectral anomaly detection. Its central idea is to combine several Mamba-based background reconstruction networks, encourage agreement between them, and constrain changes to parameters considered important for previous tasks. Anomalies are detected from reconstruction residuals, without storing historical images for replay. The main question is whether this collaborative design preserves earlier knowledge better than a single reconstruction network.

Critical strengths
  • The method addresses a meaningful practical problem. Hyperspectral detectors need to adapt to changing scenes, while storing and revisiting previous images can be expensive. I find the focus on learning without historical-image replay useful, particularly for a system that receives new observations over time.
  • The ABU results support the quality of the background model. Figure 4 shows clearer anomaly responses and less background interference than several competing methods. The substantial deterioration when GSM is removed in Table 2 also provides convincing evidence that global spatial context contributes to detection.
Critical weaknesses
  • The differences in performance across datasets are not sufficiently explained. The method performs well on ABU, but CL-BioGAN achieves better detection accuracy and background suppression at the final stage on HAD100. The appendix nevertheless describes CL-BioGAN’s detection accuracy as unsatisfactory, which is not supported by the reported results. CL-AE, which achieves competitive detection accuracy on ABU, is also omitted from the main comparison table. Including it would make the improvement over a simpler baseline easier to assess.
  • The benefit of the multi-learner design is not clearly explained. I acknowledge that two learners perform more steadily during the early stages in Table 8. However, these results do not establish whether the improvement comes from the claimed complementary memories, additional model capacity, or parameter regularization. The ablations do not separate the contributions of the cooperation loss and the continual-learning loss, leaving their respective roles in knowledge retention unclear.
Questions for authors
  1. How do the authors explain the different relative performance on ABU and HAD100? Under what conditions should CL-SSMamba be preferred, and could CL-AE be included in the main comparison table?

  2. What evidence supports the claim that different learners retain complementary background knowledge? How do the cooperation loss and parameter-importance regularization each contribute to knowledge retention? This distinction would substantially affect my assessment of the multi-learner contribution.

  3. Does the reported training-speed improvement include SLIC and LDPC preprocessing? Please clarify the timing scope so that readers can assess whether this improvement reflects the computational cost of the complete pipeline.

Reviewer_SYkrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CL-SSMamba, a bio-inspired continual learning framework for open-scenario hyperspectral anomaly detection (OHAD). Inspired by the parallel and compartmentalized information-processing mechanism of the Drosophila mushroom body, the authors construct KK parallel spatial-spectral Mamba (SSMamba) learners, each sequentially comprising Global Spatial Mamba (GSM), Local Spectral Mamba (LSM), Spectral Correlation Enhancement (SCE), and a reconstruction head, to collaboratively reconstruct the hyperspectral background. During sequential learning, only the model parameters from the previous task are retained, without storing historical HSIs, thereby reducing storage overhead and alleviating catastrophic forgetting. To handle the severe background-anomaly imbalance, the authors introduce a two-stage SLIC + LDPC preprocessing scheme. The model is jointly optimized with a reconstruction loss, an ensemble cooperation loss (KL divergence enforcing inter-learner consistency), and a continual learning loss (an EWC-like importance-weighted parameter constraint). At inference, an RX detector is applied to the reconstruction residuals to obtain anomaly scores. Experiments are conducted on the six-task ABU sequence and HAD100, comparing with CL-AE, CL-TAEF, CL-GT, CL-CaGAN, and CL-BioGAN, and reporting metrics including ACC, BWT, FWT, AUC(D,F), and AUC(BS), along with module ablations and an analysis of the number of learners K.

Critical strengths
  1. The replay-free multi-learner continual HAD design has practical value. Compared with CL-CaGAN/CL-BioGAN, which still rely on replay or auxiliary generative mechanisms, CL-SSMamba retains only parameters and does not store historical HSIs, directly targeting the long-term storage bottleneck of high-dimensional HSI; on ABU, ACC 0.9620 and BWT −0.0012 indicate that it effectively mitigates forgetting over the six-task sequence, which is attractive for practical remote-sensing deployment.

  2. The combination of spatial–spectral Mamba and region-level preprocessing forms a relatively complete design. GSM (bidirectional Mamba + overlapping windows) and LSM (LCSF + LSG + RSDC) handle long-range spatial dependencies and cross-band correlations, respectively; SCE reduces redundancy, and SLIC + LDPC performs region-level enhancement targeting the background–anomaly imbalance. Ablations (Tables 2/9/10) show that removing GSM causes the largest performance drop AUC(BS) 0.9124 → 0.7241), and removing SCE/SLIC/LCSF also consistently degrades performance, indicating that the modules make complementary contributions rather than being a mere stacking of components.

Critical weaknesses
  1. The key details of the continual learning mechanism are unclear, which weakens the verifiability of the method's claims. The paper's core selling point is replay-free continual learning that "retains only the previous task's parameters and stores no historical samples," yet the main text does not clearly explain how the KK learners are allocated/reused/expanded across the TT tasks: if KK is fixed at 2 while the number of tasks is 6, then the claim that "each learner serves as an independent memory compartment" is inconsistent with the implementation; if a new learner is added for each task, the parameter count grows linearly with TT, conflicting with the claim of "reduced storage overhead." In addition, s=1/Ts=1/T requires knowing the total number of tasks in advance, which does not conform to the online continual learning setting.

  2. The fairness of the experimental comparisons and the strength of the conclusions are insufficient. On HAD100, CL-SSMamba's AUC(BS)(0.8817) is lower than that of CL-BioGAN (0.9332), and its ACC is also slightly lower, yet the authors do not explain this discrepancy; on ABU, some baselines (CL-TAEF, CL-GT) perform abnormally poorly, and insufficient hyperparameter tuning needs to be ruled out; the parameter count is significantly larger than that of CL-AE/CL-TAEF, yet no same-budget comparison is provided; and no statistics from multiple runs are reported. In addition, the mixed use of "final-stage" tables and "per-task" tables further weakens the persuasiveness of the "robust solution" claim.

Questions for authors

1.Please clarify the correspondence between the KK learners and the TT tasks: is a new learner added for each task (so that the parameter count grows with TT), or is K=2K=2 fixed and reused across tasks? If the latter, how is the claim that "each learner serves as an independent memory compartment" guaranteed? Please provide the specific parameter update procedure for the case T>KT>K.

2.On HAD100, AUC(BS)is lower than that of CL-BioGAN and ACC is also slightly lower. What do the authors consider to be the reason? Could the authors provide per-task curves to illustrate the source of the gap?

3.In Table 5, CL-TAEF's ACC drops to 0.2114 at Task 5, and CL-GT exhibits obvious grid-like artifacts. Were these baselines subjected to the same degree of hyperparameter search as the proposed method? Could the authors report the best configurations for these baselines?

4.Were all results obtained from a single run? Since continual learning results are sensitive to task order, could the authors supplement the mean and standard deviation over 3–5 random seeds (or different task orders)?

5.The parameter count is significantly higher than that of CL-AE (77,450) and CL-TAEF (240,106). Could the authors provide same-budget comparisons of inference time, FLOPs, and GPU memory usage? Under what batch size and resolution was the reported “15% training speed improvement relative to CL-CaGAN” measured?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.