Pith. sign in

REVIEW 5 major objections 5 minor 58 references

STAIL: Semantic Text-Anchored Incremental Learning for Medical Imaging via Large Language Models

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read STAIL replaces most raw-image replay buffers with clinical text anchors and reports consistent gains against forgetting across three medical imaging modalities.

desk verdict A genuinely interesting idea—frozen LLM text embeddings as a compressed rehearsal buffer—but the paper's headline claim of consistent, significant gains is not supported by its own error bars. read the letter →

arxiv 2608.05808 v1 pith:S33A4WWH submitted 2026-08-06 cs.CV

Songpan Gao , Yajie Zhang , Guanxing Chen , Jiayu Qian , Zhenzhen Liu , Shijun Li , Xiaowei Zhu , Yao Hu
show 4 more authors
Kay Chen Tan Yu-An Huang Shiqi Wang Zhi-An Huang
This is my paper · ORCID
classification cs.CV
keywords class-incrementallearningcatastrophicforgettingmedicalimagingsemanticconsolidationbufferLLManchoringtext-anchoredfeaturealignmentmultimodalcontrastive
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a medical imaging model can keep learning new diagnostic tasks without forgetting old ones by rehearsing text instead of mostly rehearsing raw images. Its Semantic Consolidation Buffer keeps only a handful of image exemplars per class and pairs them with many clinical text descriptions, which are tens to thousands of times cheaper in bytes. A frozen large language model turns those descriptions into stationary semantic anchors, and three anchoring losses hold the evolving visual feature space close to that fixed semantic space while still allowing new classes to separate. Plugged into four existing rehearsal-based baselines, the module improves average, anytime, and backward-transfer AUC on fundus, ultrasound, and chest X-ray datasets, with average gains of 2.24% in AAA-AUC and 3.55% in BWT-AUC across benchmarks. The reader should care because this points to a memory-efficient and privacy-relevant alternative to pixel-level rehearsal, if paired text is available.

What carries the argument

The load-bearing object is the asymmetric Semantic Consolidation Buffer (SCB) combined with the LLM-derived Semantic Anchoring Mechanism (LSAM). The SCB keeps a per-class visual core of 5–20 images selected by Herding and a much larger text set of 25–100 descriptions per class, made of the paired reports of the kept images plus randomly sampled extra reports; because a 224×224 image costs hundreds to thousands of times more bytes than a short report, this reconstructs old-task semantics at a fraction of pixel-rehearsal memory. The LSAM then uses a frozen biomedical LLM as a stable coordinate system: a learned projection maps image features into the LLM text space, and the three losses $L_{\mathrm{DSA}}$, $L_{\mathrm{EPA}}$, and $L_{\mathrm{CSE}}$ enforce alignment, contrastive plasticity, and replay of previous geometric constraints. The theoretical argument is that a bounded residual alignment error between image and text features keeps the distance between two visual representations close to the distance between their semantic anchors, so the frozen text space acts as a limit on feature drift.

What would settle it

A decisive test would be to shuffle the text descriptions across images within each class before training STAIL: if the gains over the image-only baseline persist under corrupted alignment, the reported improvement is not caused by semantic anchoring, and if they disappear, the frozen text anchors are doing the work. A complementary check is to give the pixel-rehearsal baseline the same total byte budget but let it spend the text portions on additional real images, then compare across three random class orders.

Watch

Extended reading notes

Core claim

The central claim is that catastrophic forgetting in medical class-incremental learning can be reduced by using frozen LLM text embeddings as a stationary coordinate system for the visual feature space. Instead of storing many raw images, STAIL stores a small mean-matching image core plus a large set of paired and randomly sampled reports; during training, a learned projection maps image features into the frozen text space, and the three LSAM losses — developmental semantic anchoring, evolutionary plasticity adaptation, and consolidated stability enforcement — keep old-class geometry from drifting while refining new-class separability. The reported result is that STAIL works as a plug-and-play module: it improves Replay, iCaRL, WA, and TagFex on ODIR-5K, US-DATA, and MS-CXR, with the largest single improvements of about +8.10 percentage points in Avg-AUC on US-DATA for Replay and +12.00 percentage points in BWT-AUC on ODIR-5K for WA, while using less memory than the image-heavy buffers it replaces. The paper also argues theoretically that bounded alignment error between visual and semantic features bounds the distortion of inter-class distances, which is the mechanism behind the stability it observes.

Load-bearing premise

The method's gains rely on every training image having a paired text description that actually describes the clinically relevant content and that the frozen LLM embeds faithfully; if reports are missing, noisy, or semantically misaligned with the image, the anchoring losses can pull visual features toward the wrong points and the theoretical bound on alignment error no longer holds.

Editorial extensions

If this is right

  • Existing rehearsal-based baselines can adopt STAIL without changing their own loss terms; the module only adds a text stream and three weighted anchor losses, and the reported gains hold across four baselines and three imaging modalities.
  • Because a standard 224×224 image costs roughly 462–3,369 times more memory than the average text report in the authors' datasets, the same byte budget can support far denser semantic memory when images are partially replaced by text.
  • The three losses separate the stability and plasticity jobs: $L_{\mathrm{DSA}}$ and $L_{\mathrm{CSE}}$ protect old classes while $L_{\mathrm{EPA}}$ specializes new ones, so practitioners can reweight them for datasets dominated by drift or dominated by new-class confusion.
  • On the long-tailed MS-CXR benchmark, STAIL can reach positive backward transfer (BWT-AUC +2.03% for iCaRL), meaning later tasks can strengthen, not merely preserve, earlier knowledge.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same text-anchored replay recipe should transfer to nonmedical continual learning wherever paired image–text data already exist, such as satellite imagery with captions or inspection logs; running STAIL on a standard class-incremental benchmark with a general-purpose LLM under byte-equal memory would test this directly.
  • The privacy advantage is only partial: trading raw images for free-text radiology reports moves the sensitive data from pixels to patient language, so a deployment would still need de-identification or synthetic text before the privacy claim is complete.
  • Positive backward transfer on MS-CXR hints that frozen text anchors can act as a deliberate review mechanism: scheduling rare classes early and giving them dense text memory might be a better curriculum than the uniform task orders the paper evaluates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes STAIL, a plug-and-play module for class-incremental learning (CIL) in medical imaging. STAIL replaces pixel-level rehearsal with an asymmetric Semantic Consolidation Buffer (SCB) that stores a few image exemplars plus many paired and randomly sampled text descriptions, and it adds an LLM-derived Semantic Anchoring Mechanism (LSAM) with three losses: developmental semantic anchoring (DSA, Eq. 4), evolutionary plasticity adaptation (EPA, Eq. 5), and consolidated stability enforcement (CSE, Eq. 6). Experiments on ODIR-5K, US-DATA, and MS-CXR claim consistent and significant improvements over four rehearsal baselines (Replay, iCaRL, WA, TagFex) across Avg-AUC, AAA-AUC, and BWT-AUC. The paper also reports ablations, an equal-memory comparison, different LLM choices, text selection strategies, hyperparameter sensitivity, and category-order experiments, plus a theoretical analysis in Appendix B and pseudocode in Appendix C.

Significance. The core idea is interesting and timely: using frozen LLM text embeddings as stationary anchors, and compressing historical memory through text, is a plausible route to privacy-friendlier and more memory-efficient medical CIL. The experimental scope is broad (three heterogeneous modalities, multiple baselines, several ablations), and the equal-memory comparison in Fig. 2 is a genuinely useful check: even when the baseline is given slightly more memory and more visual exemplars, STAIL variants show large gains on US-DATA and MS-CXR. The method is plug-and-play, and code availability is promised. The significance is tempered, however, by the fragility of the statistical evidence, the inconsistency of the full LSAM relative to its own subcomponents, and the overclaimed theoretical contribution. If the statistical and memory-fairness issues are resolved, the work could be a solid contribution to medical continual learning.

major comments (5)
  1. [§4.2, Table 1] The headline claim that STAIL 'consistently and significantly' enhances baselines is not supported by the reported error bars. With only 3 seeds, most of the 36 baseline–dataset–metric comparisons have overlapping means and standard deviations. For example, on MS-CXR, Replay+STAIL gains are only 0.37 (Avg-AUC), 0.01 (AAA-AUC), and 2.85 (BWT-AUC) with standard deviations of 2.71, 1.30, and 4.28, respectively; TagFex+STAIL's BWT-AUC moves in the wrong direction (-2.36 vs -1.42), as does iCaRL+STAIL's BWT-AUC on US-DATA (-8.29 vs -8.07). No per-seed significance testing, confidence intervals, or multiple-comparison correction is provided. The authors should either add paired significance tests (e.g., paired t-test or Wilcoxon across seeds) and report effect sizes with confidence intervals, or qualify the claim to identify the specific settings where gains are statistically reliable.
  2. [§4.2, Table 2, ablation] The full LSAM is not consistently better than its own subcomponents, which undercuts the claim that the three losses are complementary. On US-DATA with Replay, DSA-only achieves Avg-AUC 85.09 and AAA-AUC 83.27, while the full DSA+EPA+CSE drops to 78.74 and 81.06; for iCaRL, DSA-only Avg-AUC is 83.49 versus 80.23 for the full model. The paper acknowledges this in the ablation discussion, but it is a load-bearing inconsistency because the final reported model includes all three losses. Please explain why the full model is the deployed configuration, or provide a selection rule, or report the best-performing variant per baseline/dataset.
  3. [§4.1.3, Appendix A, §4.2 equal-memory experiment] The main comparisons in Table 1 are not byte-fair: STAIL stores per-class image exemplars plus text exemplars, while the baselines store only images, so the 'fixed per-class memory budget' statement is ambiguous unless text storage is explicitly counted. The equal-memory experiment in Fig. 2 covers only US-DATA and MS-CXR, and for ODIR-5K no byte-level accounting is given (the average text length for ODIR is not reported). Please provide a byte-level memory table for all three datasets and show the main Table 1 results under equal total bytes, or clearly state how the per-class budget was defined and whether the reported improvements change under byte-fair comparison.
  4. [Appendix B, contributions] The theoretical claims are overclaimed relative to the content. Proposition 1 is a direct consequence of the triangle inequality under an assumed alignment bound epsilon; it does not prove that the SCB 'reconstructs the data manifold of historical tasks,' as stated in the contribution list. Proposition 2 is an informal statement with no formal statement or proof. Because the paper explicitly highlights this theoretical proof as a contribution, the authors should either provide a genuine formalization (e.g., a coverage or reconstruction guarantee under concrete assumptions on the text distribution) or substantially soften the theoretical claims.
  5. [§4.1.1, Algorithm 1 lines 27-30, Appendix B] The method depends on paired, diagnosis-relevant text for every training image and on the quality of freely sampled extra texts; however, the paper does not analyze robustness to text noise or absence. ODIR-5K was filtered to retain only samples with text, which changes the benchmark, and the random extra texts in Eq. (7) are not quality-filtered. The theoretical guarantee in Appendix B assumes a bounded alignment error epsilon, which may not hold with noisy or semantically misaligned reports. Please add an ablation with corrupted or missing text (e.g., replacing reports with class-level generic descriptions) and report the size and difficulty of the filtered ODIR subset relative to the original.
minor comments (5)
  1. [Table 4] No standard deviations or seed counts are reported in the category-order experiment; please state whether these are single-run results and add error bars or at least note the number of seeds.
  2. [Fig. 2] The caption says results are averaged over three runs, but the figure does not specify whether error bars/shaded regions are standard deviation or confidence intervals; please clarify.
  3. [Eq. (6)] The notation for the CSE loss reuses B_I and P(i) without stating that the text batch is drawn from the SCB; make the data source explicit in the equation or its surrounding text.
  4. [Algorithm 1, line 29, Eq. (8)] Line 29 samples extra texts of size '≤ N_text - N_img', but the buffer update in Eq. (8) adds paired texts plus extra texts; it is unclear whether the final per-class text count is exactly N_text or can be smaller. Clarify the budget semantics.
  5. [§4.2] The 'rare phenomenon' of positive backward transfer (BWT-AUC 2.03 for iCaRL+STAIL on MS-CXR) is reported with a standard deviation of 2.82, so it is statistically indistinguishable from zero; please temper the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: STAIL's empirical gains rest on external frozen LLM text embeddings and a new training-time multimodal regularizer, not on a fitted parameter renamed as a prediction or on a load-bearing self-citation chain.

full rationale

The paper's derivation chain is self-contained relative to its inputs. The frozen BioMistral-7B text encoder is external to the training loop and is not trained on the medical incremental benchmarks; the visual encoder is trained with standard baseline losses plus the additive regularizers in Eqs. (4)-(6), combined in Eq. (2). The reported improvements in Table 1 are comparisons against independently implemented baselines (Replay, iCaRL, WA, TagFex, plus EWC, DER, MEMO, L2P, DualPrompt), and the equal-memory comparison in Fig. 2 gives the baseline more image exemplars while STAIL still wins, which makes the empirical claim less dependent on a favorable memory accounting. No fitted parameter from one dataset is renamed as a prediction on a closely related quantity, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. The paper's self-citations (e.g., refs. 14, 28, 52, 53) appear only in background and future-work sentences and are not load-bearing. One caveat worth noting is that Appendix B's Proposition 2 is more a heuristic restatement than a formal manifold-reconstruction proof: it asserts that adding semantic descriptions 'provides a richer representation,' which is close to definitional, but this passage does not generate any empirical number and does not make the central claim circular. The legitimate weaknesses of the paper are evidential (e.g., overlap of error bars in Table 1 for several comparisons, dependence on availability and quality of paired text) rather than circularity of the derivation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The method relies on two main assumptions: (1) frozen LLM text space is a useful fixed semantic anchor, and (2) text reports capture the visual distribution well enough for reconstruction. The free parameters are all loss weights and buffer sizes, explicitly tuned on the datasets. The positive BWT-AUC phenomenon is attributed to 'review effect' from the paper's own buffer strategy, an internal explanation that has not been compared against an external baseline in the paper.

free parameters (5)
  • lambda_DSA = 1 on ODIR, tuned per dataset elsewhere
    Loss weight for the developmental semantic anchoring loss in Eq. 2. Chosen by grid search on ODIR (0.01-10) and reported to require different values on US-DATA and MS-CXR.
  • lambda_EPA = 1.5 on ODIR, tuned per dataset elsewhere
    Loss weight for the evolutionary plasticity adaptation loss in Eq. 2. Selected by the same sequential grid search; the sensitivity curves in Fig. 6 show strong inverted-U dependence.
  • lambda_CSE = 1 on ODIR, tuned per dataset elsewhere
    Loss weight for the consolidated stability enforcement loss in Eq. 2. Tuned alongside the other two weights; the paper says the optimal combination is not transferable across datasets.
  • N_text (text exemplars per class) = 100 (ODIR, MS-CXR), 25 (US-DATA)
    Architectural budget choice for the asymmetric buffer. The ratio N_text:N_img is set by hand (5:1 for ODIR/MS-CXR, 5:1 for US-DATA with N_img=5), not derived from analysis.
  • temperature tau = 0.07
    InfoNCE temperature in Eqs. 5 and 6, set to the standard CLIP-style value without a sensitivity analysis.
assumptions (4)
  • domain assumption Frozen LLM text embeddings provide a stable, meaningful semantic space for medical concepts (BioMistral embeddings are valid anchors).
    The entire LSAM mechanism treats the LLM output space as a fixed developmental prior (Section 3.4, Eq. 4). If text embedding geometry is not semantically consistent for medical imaging, the alignment loss would pull visual features toward uninformative targets.
  • domain assumption The paired texts (reports) are sufficiently aligned with the visual content of the images (image-text correspondence holds).
    The DSFE and LSAM losses (Eqs. 4-6) assume each image has a semantically paired text. The paper filters ODIR to only samples with text, and does not measure alignment or noise in the reports.
  • ad hoc to paper A bounded alignment error epsilon exists between visual and semantic representations after training (Appendix B).
    Proposition 1 assumes ||z_I - z_S|| <= epsilon and Proposition 2 assumes coverage. These bounds are never estimated or enforced, so the theoretical explanation rests on an unverified premise.
  • domain assumption Herding-selected image exemplars plus randomly sampled texts can 'densely reconstruct' the historical data distribution (Eq. 7-8).
    This is the core efficiency claim of the SCB. The paper provides no quantitative measure of reconstruction error or distribution coverage; it only shows downstream task accuracy.
invented entities (2)
  • Semantic Consolidation Buffer (SCB) independent evidence
    purpose: A new memory structure replacing raw image rehearsal with a small image core plus many stored text descriptions.
    The buffer itself is not an entity with falsifiable predictions, but its outputs are measurable through the benchmark tables and ablation study, which provide evidence of its utility.
  • LLM-derived Semantic Anchoring Mechanism (LSAM) with three losses (DSA, EPA, CSE) independent evidence
    purpose: A new training objective that anchors visual features to frozen text to reduce forgetting.
    Not a physical entity; its contribution is directly tested by the ablation study (Table 2) and hyperparameter sensitivity, which constitute falsifiable handles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STAIL: Semantic Text-Anchored Incremental Learning for Medical Imaging via Large Language Models." pith.science (2026). https://pith.science/paper/S33A4WWH

@misc{pith2026260805808,
  author       = {Pith},
  title        = {Pith review of: STAIL: Semantic Text-Anchored Incremental Learning for Medical Imaging via Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S33A4WWH}},
  note         = {Machine review of arXiv:2608.05808}
}
read the original abstract

Deep learning models applied to medical image analysis suffer from severe catastrophic forgetting when continually adapting to new clinical tasks in dynamic environments. Mainstream incremental learning methods typically mitigate this by rehearsing raw historical images. However, this pixel-level rehearsal incurs significant storage overhead, raises privacy concerns, and fails to adequately capture the true data distribution with sparse exemplars. Inspired by human cognitive mechanisms, we propose a novel framework termed Semantic Text-Anchored Incremental Learning (STAIL) for sequential clinical tasks. To overcome the rehearsal bottleneck, STAIL introduces an asymmetric semantic consolidation buffer (SCB). By incorporating a minimal set of image anchors and extensive textual descriptions, the SCB enables dense semantic reconstruction of old tasks at a minimal storage cost. Furthermore, we design an LLM-derived Semantic Anchoring Mechanism (LSAM) that leverages the stable semantic space of frozen large language models as developmental priors. This mechanism explicitly anchors evolving visual features to textual representations, guiding and constraining plasticity and stability at both macroscopic and microscopic levels. Extensive experiments across three heterogeneous medical datasets, covering fundus, ultrasound, and X-ray imaging, demonstrate that STAIL acts as a highly effective plug-and-play module. It comprehensively enhances the performance of various existing baselines, achieving average gains of 2.24\% in AAA-AUC for sustained performance and 3.55\% in BWT-AUC for reduced forgetting. Code is available.

Figures

Figures reproduced from arXiv: 2608.05808 by the authors.

Figure 1
Figure 1. Schematic illustration of the proposed STAIL framework. The top panel depicts the training process for task 𝑡, where the model accepts current task data combined with historical asymmetric exemplars retrieved from the SCB as input. Following dual-stream feature extraction via a trainable visual encoder and a frozen LLM, the LSAM regularizes the feature space through three losses: 𝐷𝑆𝐴, 𝐸𝑃 𝐴, and 𝐶𝑆𝐸 . The bottom p… view at source ↗
Figure 2
Figure 2. Performance comparison between baseline methods and their STAIL variants under equal buffer memory usage. Results are averaged over three runs. of US-DATA, which contains 41 fine-grained categories across multiple anatomical organs, where strong semantic regularization provides effective guidance for final repre￾sentation optimization. However, due to the heterogeneous and fine-grained characteristics of multi-organ… view at source ↗
Figure 3
Figure 3. Comparison of different LLMs on the ODIR-5K dataset, averaged across three runs. media. Specifically, a single standard 224 × 224 RGB image consumes approximately 147 KB (150, 528 Bytes). In stark contrast, our statistical analysis reveals that the average text description lengths for the US-DATA and MS-CXR datasets are merely 325.39 Bytes and 44.68 Bytes, respectively. This means that a single image requires approx… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Activation maps on a cataract sample from the ODIR-5K dataset. strategy as the default setting for our framework to ensure efficiency and generalization capability. Visualization Results. To explore how the model com￾bats feature drift, we visualized feature activation…
Figure 5
Figure 5. Figure 5: Avg-AUC, AAA-AUC, and BWT-AUC at each incremental step on US-DATA, averaged across three runs [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Sensitivity analysis on hyperparameters: 𝜆𝐷𝑆𝐴, 𝜆𝐸𝑃 𝐴, and 𝜆𝐶𝑆𝐸 . shifts and numerous fine-grained categories, while the MS￾CXR dataset exhibits an extreme long-tail class distribu￾tion, the optimal hyperparameter combinations required by different baseline methods natu…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 48 canonical work pages

  1. [1]

    CCSI: Continual Class-Specific Impression for data-free class incremental learning

    Ayromlou, S., Tsang, T., Abolmaesumi, P., Li, X., 2024. CCSI: Continual Class-Specific Impression for data-free class incremental learning. Medical Image Analysis 97, 103239

    work page 2024

  2. [2]

    Biaspruner: Mitigating bias transfer in continual learning for fair medical image analysis

    Bayasi, N., Fayyad, J., Bissoto, A., Hamarneh, G., Garbi, R., 2025. Biaspruner: Mitigating bias transfer in continual learning for fair medical image analysis. Medical image analysis , 103764

    work page 2025

  3. [3]

    IEEE Transactions on Medical Imaging 43, 3767–3779

    Bayasi, N., Hamarneh, G., Garbi, R., 2024.𝐺𝐶 2: Generalizable Continual Classification of Medical Images. IEEE Transactions on Medical Imaging 43, 3767–3779

    work page 2024

  4. [4]

    Making the Most of Text Semantics to Improve Biomedical Vision–Language Processing, in: European Conference on Computer Vision, Springer

    Boecking, B., Usuyama, N., Bannur, S., Castro, D.C., Schwaighofer, A.,Hyland,S.,Wetscherek,M.,Naumann,T.,Nori,A.,Alvarez-Valle, J., et al., 2022. Making the Most of Text Semantics to Improve Biomedical Vision–Language Processing, in: European Conference on Computer Vision, Springer. pp. 1–21

    work page 2022

  5. [5]

    MS- CXR: Making the Most of Text Semantics to Improve Biomedical Vision-Language Processing

    Boecking, B., Usuyama, N., Bannur, S., Coelho de Castro, D., Schwaighofer, A., Hyland, S., Sharma, H., Wetscherek, M.T., Nau- mann,T.,Nori,A.,AlvarezValle,J.,Poon,H.,Oktay,O.,2024. MS- CXR: Making the Most of Text Semantics to Improve Biomedical Vision-Language Processing. PhysioNet URL:https://doi.org/10. 13026/9g2z-jg61, doi:10.13026/9g2z-jg61. version 1.1.0

    doi:10.13026/9g2z-jg61 2024

  6. [6]

    Medical Image Analysis 114, 104235

    Bui,D.C.,Kwak,J.T.,2026.Welcomenewdoctor:Continuallearning with expert consultation and autoregressive inference for whole slide image analysis. Medical Image Analysis 114, 104235. URL:https: //www.sciencedirect.com/science/article/pii/S136184152600304X, doi:https://doi.org/10.1016/j.media.2026.104235

    work page arXiv 2026

  7. [7]

    Newinsightsonreducingabruptrepresentation changeinonlinecontinuallearning

    Caccia, L., Aljundi, R., Asadi, N., Tuytelaars, T., Pineau, J., Belilovsky,E.,2021. Newinsightsonreducingabruptrepresentation changeinonlinecontinuallearning. arXivpreprintarXiv:2104.05025

    arXiv 2021

  8. [8]

    Toward Bidirectional Adaptability for Few-Shot Class-Incremental LearningWithForward-BackwardKnowledgeTransfer

    Chen, B., Chen, Z., Cai, S., Fang, X., Bennamoun, M., Xie, S., 2026. Toward Bidirectional Adaptability for Few-Shot Class-Incremental LearningWithForward-BackwardKnowledgeTransfer. IEEETrans- actionsonMultimedia28,2337–2351. doi:10.1109/TMM.2026.3651015

    work page arXiv 2026

Show all 58 references
  1. [9]

    Meditron-70b:Scalingmedicalpretrainingforlargelanguagemodels

    Chen, Z., Cano, A.H., Romanou, A., Bonnet, A., Matoba, K., Salvi, F., Pagliardini, M., Fan, S., Köpf, A., Mohtashami, A., et al., 2023. Meditron-70b:Scalingmedicalpretrainingforlargelanguagemodels. arXiv preprint arXiv:2311.16079

    2023 arXiv

  2. [10]

    Distribution-Level Memory Recall for ContinualLearning:PreservingKnowledgeandAvoidingConfusion

    Cheng, S., Geng, K., He, C., Qiu, Z., Xu, L., Qiu, H., Wang, L., Wu, Q., Meng, F., Li, H., 2025. Distribution-Level Memory Recall for ContinualLearning:PreservingKnowledgeandAvoidingConfusion. IEEE Transactions on Multimedia 27, 4151–4166

    2025

  3. [11]

    Trends in Cognitive Sciences 21, 618–631

    Gilboa,A.,Marlatte,H.,2017.Neurobiologyofschemasandschema- mediated memory. Trends in Cognitive Sciences 21, 618–631

    2017

  4. [12]

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al- Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al.,

  5. [13]

    Deep residual learning for image recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778

    2016

  6. [14]

    scBIT: Integrating Single-cell Transcriptomic Data into fMRI-based Prediction for Alzheimer’s Disease Diagnosis

    Huang, Y.A., Hu, Y., Li, Y.C., Cao, X., Li, X., Tan, K.C., You, Z.H., Huang, Z.A., 2026. scBIT: Integrating Single-cell Transcriptomic Data into fMRI-based Prediction for Alzheimer’s Disease Diagnosis. IEEE Transactions on Medical Imaging

    2026

  7. [15]

    Overcoming catastrophic forgetting in neural net- works

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G.,Rusu,A.A.,Milan,K.,Quan,J.,Ramalho,T.,Grabska-Barwinska, A., et al., 2017. Overcoming catastrophic forgetting in neural net- works. ProceedingsoftheNationalAcademyofSciences114,3521– 3526

    2017

  8. [16]

    What learning systemsdointelligentagentsneed?Complementarylearningsystems theory updated

    Kumaran, D., Hassabis, D., McClelland, J.L., 2016. What learning systemsdointelligentagentsneed?Complementarylearningsystems theory updated. Trends in Cognitive Sciences 20, 512–534

    2016

  9. [17]

    Continual learning in medical image analysis: A comprehensive review of recent advancements and future prospects

    Kumari, P., Chauhan, J., Bozorgpour, A., Huang, B., Azad, R., Merhof, D., 2025. Continual learning in medical image analysis: A comprehensive review of recent advancements and future prospects. Medical Image Analysis , 103730

    2025

  10. [18]

    Biomistral: A collection of open-source pre- trained large language models for medical domains

    Labrak, Y., Bazoge, A., Morin, E., Gourraud, P.A., Rouvier, M., Dufour, R., 2024. Biomistral: A collection of open-source pre- trained large language models for medical domains. arXiv preprint arXiv:2402.10373

    2024 arXiv

  11. [19]

    Building machines that learn and think like people

    Lake, B.M., Ullman, T.D., Tenenbaum, J.B., Gershman, S.J., 2017. Building machines that learn and think like people. Behavioral and Brain Sciences 40, e253

    2017

  12. [20]

    Turning the Tables: Enabling Backward Transfer via Causal-Aware LoRAinContinualLearning,in:TheThirty-ninthAnnualConference on Neural Information Processing Systems

    Li, C., Ye, R., Qin, J., Cui, J., Wang, L., Hu, N., Liao, Q., 2025a. Turning the Tables: Enabling Backward Transfer via Causal-Aware LoRAinContinualLearning,in:TheThirty-ninthAnnualConference on Neural Information Processing Systems

  13. [21]

    ProgressiveDistillationforIncrementalLearninginCornealConfocal Microscopy Segmentation

    Li, H., Ma, B., Mou, L., Liu, Y., Zheng, Q., Qi, H., Zhao, Y., 2025b. ProgressiveDistillationforIncrementalLearninginCornealConfocal Microscopy Segmentation. IEEE Transactions on Medical Imaging

  14. [22]

    Ultrasound report generation with cross-modality feature alignmentviaunsupervisedguidance

    Li, J., Su, T., Zhao, B., Lv, F., Wang, Q., Navab, N., Hu, Y., Jiang, Z., 2024. Ultrasound report generation with cross-modality feature alignmentviaunsupervisedguidance. IEEETransactionsonMedical Imaging 44, 19–30

    2024

  15. [23]

    Learning without forgetting

    Li, Z., Hoiem, D., 2017. Learning without forgetting. IEEE transac- tions on Pattern Analysis and Machine Intelligence 40, 2935–2947

    2017

  16. [24]

    Beyond not-forgetting: Continual learning with backward knowledge transfer

    Lin, S., Yang, L., Fan, D., Zhang, J., 2022. Beyond not-forgetting: Continual learning with backward knowledge transfer. Advances in Neural Information Processing Systems 35, 16165–16177

    2022

  17. [25]

    Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory

    McClelland, J.L., McNaughton, B.L., O’Reilly, R.C., 1995. Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological Review 102, 419

    1995

  18. [26]

    Catastrophic interference in con- nectionistnetworks:Thesequentiallearningproblem,in:Psychology of Learning and Motivation

    McCloskey, M., Cohen, N.J., 1989. Catastrophic interference in con- nectionistnetworks:Thesequentiallearningproblem,in:Psychology of Learning and Motivation. Elsevier. volume 24, pp. 109–165

    1989

  19. [27]

    Language as a cognitive tool

    Mirolli, M., Parisi, D., 2009. Language as a cognitive tool. Minds and Machines 19, 517–528

    2009

  20. [28]

    CPSR- CLIP: Conditional Prompt-Induced Style Reconstruction for Zero- Shot Domain Adaptation

    Qian, J., Lu, Y., Xie, W., Lai, Z., Wang, M., Li, X., 2025. CPSR- CLIP: Conditional Prompt-Induced Style Reconstruction for Zero- Shot Domain Adaptation. IEEE Transactions on Multimedia

    2025

  21. [29]

    Learning Transferable Visual Models From Natural Language Supervision, in: International Conference on Machine Learning, PMLR

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S.,Sastry,G.,Askell,A.,Mishkin,P.,Clark,J.,etal.,2021. Learning Transferable Visual Models From Natural Language Supervision, in: International Conference on Machine Learning, PMLR. pp. 8748– 8763

    2021

  22. [30]

    Rahmani,S.,Chatterjee,R.,Etemad,A.,Hashemi,J.,2025. Dynamic Prototype Rehearsal for Continual ECG Arrhythmia Detection, in: ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Songpan Gao et al.:Preprint submitted to ElsevierPage 14 of 15 STAIL Speech and Signal Pro...

    2025

  23. [31]

    iCaRL: Incremental Classifier and Representation Learning, in: Proceedings oftheIEEEConferenceonComputerVisionandPatternRecognition, pp

    Rebuffi,S.A.,Kolesnikov,A.,Sperl,G.,Lampert,C.H.,2017. iCaRL: Incremental Classifier and Representation Learning, in: Proceedings oftheIEEEConferenceonComputerVisionandPatternRecognition, pp. 2001–2010

    2017

  24. [32]

    Grad-CAM: Visual Explanations from Deep Net- works via Gradient-based Localization, in: Proceedings of the IEEE International Conference on Computer Vision, pp

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D., 2017. Grad-CAM: Visual Explanations from Deep Net- works via Gradient-based Localization, in: Proceedings of the IEEE International Conference on Computer Vision, pp. 618–626

    2017

  25. [33]

    Large-scale and Fine-grained Vision-Language Pre-training for Enhanced CT Image Understand- ing, in: International Conference on Learning Representations, pp

    Shui, Z., Zhang, J., Cao, W., Wang, S., Guo, R., Lu, L., Yang, L., Ye, X., Liang, T., Zhang, Q., et al., 2025. Large-scale and Fine-grained Vision-Language Pre-training for Enhanced CT Image Understand- ing, in: International Conference on Learning Representations, pp. 24094–24107

    2025

  26. [34]

    Artificial life 11, 13–29

    Smith,L.,Gasser,M.,2005.Thedevelopmentofembodiedcognition: Six lessons from babies. Artificial life 11, 13–29

    2005

  27. [35]

    Qwen2.5: A Party of Foundation Models

    Team, Q., 2024. Qwen2.5: A Party of Foundation Models. URL: https://qwenlm.github.io/blog/qwen2.5/

    2024

  28. [36]

    Multi-Scale Feature Alignment for Continual Learning of Unlabeled Domains

    Thandiackal, K., Piccinelli, L., Gupta, R., Pati, P., Goksel, O., 2024. Multi-Scale Feature Alignment for Continual Learning of Unlabeled Domains. IEEE Transactions on Medical Imaging 43, 2599–2609

    2024

  29. [37]

    In- ternational Competition on Ocular Disease Intelligent Recognition

    University, P., Shanggong Medical Technology Co., L., 2019. In- ternational Competition on Ocular Disease Intelligent Recognition. https://odir2019.grand-challenge.org

    2019

  30. [38]

    Wang, E., Li, Q., Liu, D., Yang, B., 2025a. Enhancing Continual Learning for Medical Imaging: Efficient Knowledge Transfer and Multi-DiseasePrediction,in:ICASSP2025-2025IEEEInternational Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 1–5

  31. [39]

    Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion, in: The eleventh International Conference on Learning Representations

    Wang, F.Y., Zhou, D.W., Liu, L., Ye, H.J., Bian, Y., Zhan, D.C., Zhao, P., 2022a. Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion, in: The eleventh International Conference on Learning Representations

  32. [40]

    Foster: Feature boosting and compression for class-incremental learning, in: European Conference on Computer Vision, Springer

    Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C., 2022b. Foster: Feature boosting and compression for class-incremental learning, in: European Conference on Computer Vision, Springer. pp. 398–414

  33. [41]

    Rethinking Class-Incremental Learning From a Dynamic ImbalancedLearningPerspective

    Wang, L., Xiang, L., Wang, Y., Wu, H., Yang, H., Liu, J., He, Z., 2025b. Rethinking Class-Incremental Learning From a Dynamic ImbalancedLearningPerspective. IEEETransactionsonMultimedia 28, 825–836

  34. [42]

    Rehearsal-free modular and compositional continual learning for language models

    Wang, M., Adel, H., Lange, L., Strötgen, J., Schütze, H., 2024. Rehearsal-free modular and compositional continual learning for language models. arXiv preprint arXiv:2404.00790

    2024 arXiv

  35. [43]

    Cross-Domain Invariant Feature Absorption and Domain-Specific Feature Retention for Domain Incremental Chest X-Ray Classifica- tion

    Wang, M., He, Y., Peng, L., Song, X., Dong, S., Gong, Y., 2025c. Cross-Domain Invariant Feature Absorption and Domain-Specific Feature Retention for Domain Incremental Chest X-Ray Classifica- tion. IEEE Transactions on Medical Imaging 44, 2041–2055

  36. [44]

    Dual-AttentionTransformersforClass-IncrementalLearning:ATale ofTwoMemories

    Wang, S., Wang, W., Sun, Y., Wang, Z., Wang, B., Yin, B., 2025d. Dual-AttentionTransformersforClass-IncrementalLearning:ATale ofTwoMemories. IEEETransactionsonMultimedia27,8763–8775. doi:10.1109/TMM.2025.3607800

    2025

  37. [45]

    DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning, in:EuropeanConferenceonComputerVision,Springer.pp.631–648

    Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al., 2022c. DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning, in:EuropeanConferenceonComputerVision,Springer.pp.631–648

  38. [46]

    Learning to Prompt for Continual Learning,in:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition, pp

    Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T., 2022d. Learning to Prompt for Continual Learning,in:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition, pp. 139–149

  39. [47]

    Herding dynamical weights to learn, in: Pro- ceedings of the 26th Annual International Conference on Machine Learning, pp

    Welling, M., 2009. Herding dynamical weights to learn, in: Pro- ceedings of the 26th Annual International Conference on Machine Learning, pp. 1121–1128

    2009

  40. [48]

    PMC-LLaMA: toward building open-source language models for medicine

    Wu, C., Lin, W., Zhang, X., Zhang, Y., Xie, W., Wang, Y., 2024a. PMC-LLaMA: toward building open-source language models for medicine. Journal of the American Medical Informatics Association 31, 1833–1843

  41. [49]

    Meta continual learning revisited: Implicitly enhancing online hessian ap- proximation via variance reduction, in: The Twelfth International Conference on Learning Representations

    Wu, Y., Huang, L.K., Wang, R., Meng, D., Wei, Y., 2024b. Meta continual learning revisited: Implicitly enhancing online hessian ap- proximation via variance reduction, in: The Twelfth International Conference on Learning Representations

  42. [50]

    DER: Dynamically expandable representation for class incremental learning, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition, pp

    Yan, S., Xie, J., He, X., 2021. DER: Dynamically expandable representation for class incremental learning, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition, pp. 3014–3023

    2021

  43. [51]

    Continual learning through synapticintelligence,in:InternationalConferenceonMachineLearn- ing, PMLR

    Zenke, F., Poole, B., Ganguli, S., 2017. Continual learning through synapticintelligence,in:InternationalConferenceonMachineLearn- ing, PMLR. pp. 3987–3995

    2017

  44. [52]

    Anti-Confounding Hashing: Enhancing Radiological Image Retrieval via Debiased Weighting and Counterfactual Reasoning

    Zhang, Y., Hu, Y., Cai, C., Huang, Y.A., Huang, Z.A., Chen Tan, K., 2025a. Anti-Confounding Hashing: Enhancing Radiological Image Retrieval via Debiased Weighting and Counterfactual Reasoning. IEEE Transactions on Neural Networks and Learning Systems 36, 15055–15069. doi:10.11...

  45. [53]

    CausalMixNet:Amixed-attentionframeworkforcausal intervention in robust medical image diagnosis

    Zhang, Y., Huang, Y.A., Hu, Y., Liu, R., Wu, J., Huang, Z.A., Tan, K.C.,2025b. CausalMixNet:Amixed-attentionframeworkforcausal intervention in robust medical image diagnosis. Medical Image Analysis 103, 103581. doi:https://doi.org/10.1016/j.media.2025. 103581

    doi

  46. [54]

    Maintaining discriminationandfairnessinclassincrementallearning,in:Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T., 2020. Maintaining discriminationandfairnessinclassincrementallearning,in:Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13208–13217

    2020

  47. [55]

    Task-Agnostic Guided Feature Expansion for Class-Incremental Learning, in: Pro- ceedings of the Computer Vision and Pattern Recognition Confer- ence, pp

    Zheng, B., Zhou, D.W., Ye, H.J., Zhan, D.C., 2025. Task-Agnostic Guided Feature Expansion for Class-Incremental Learning, in: Pro- ceedings of the Computer Vision and Pattern Recognition Confer- ence, pp. 10099–10109

    2025

  48. [56]

    PyCIL: a python toolbox for class-incremental learning

    Zhou, D.W., Wang, F.Y., Ye, H.J., Zhan, D.C., 2023. PyCIL: a python toolbox for class-incremental learning. SCIENCE CHINA Information Sciences 66, 197101. doi:https://doi.org/10.1007/ s11432-022-3600-y

    2023

  49. [57]

    AModelor603 Exemplars: Towards Memory-Efficient Class-Incremental Learning, in: The Eleventh International Conference on Learning Representa- tions

    Zhou,D.W.,Wang,Q.W.,Ye,H.J.,Zhan,D.C.,2022. AModelor603 Exemplars: Towards Memory-Efficient Class-Incremental Learning, in: The Eleventh International Conference on Learning Representa- tions. Songpan Gao et al.:Preprint submitted to ElsevierPage 15 of 15

    2022

  50. [2024]

    arXiv preprint arXiv:2407.21783

    The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.