acceptodds
Withdrawn from ICLR 2027

SN-ABSTOPK: AN UNROLLED SPATIAL-NORMATIVE SPARSE AUTOENCODER FOR MEDICAL FOUNDATION MODELS

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse autoencoders (SAEs) decompose foundation model representations into monosemantic latents. However, applying independent per-token sparsity across image grids ignores spatial correlations, which causes feature fragmentation in medical imaging. We introduce Spatial-Normative AbsTopK (SN-AbsTopK), an unrolled sparse autoencoder that integrates spatial group support containment and normative reference centering into signed-magnitude dictionary learning. Rather than evaluating tokens independently, SN-AbsTopK operates via a two-stage se- lection mechanism across spatial token groups. This structure enables effective processing across continuous regions, such as contiguous histology tiles or vol- umetric MRI slice stacks. In the first stage, the model aggregates pre-activation column energies across all group tokens to rank dictionary atoms and establish a shared group candidate support. In the second stage, each individual token se- lects its active signed features strictly from within this shared group candidate pool. This guarantees that all token-level active features remain strict subsets of the group candidate set. To isolate disease-specific variations, an exponential moving average reference buffer first centers the input embeddings. This pro- cess forces the sparse latent codes to represent residual departures from base- line reference states. Evaluated across histopathology (BCSS with frozen UNI) and radiology (CheXpert and Duke Breast MRI with frozen BiomedCLIP), SN- AbsTopK substantially enhances structural continuity, increasing spatial autocor- relation by 2.9× on BCSS (Moran’s 0.244 vs. 0.084) and 3D adjacent-slice sup- port persistence by 60.7% on Duke MRI (Jaccard 0.614 vs. 0.382). Although strict candidate bounds increase reconstruction error and cause feature lock-out ( 80% dead atoms), raw reconstruction error proves decoupled from clinical de- codability. Consequently, SN-AbsTopK yields superior downstream linear probe performance (+0.085 macro-AUROC on CheXpert; +0.033 on Duke MRI).

Primary area:
interpretability and explainable AI
Keywords:
Sparse Autoencoders, Spatial Group Sparsity, Sparse Coding, Representation Learning, Medical Imaging

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ZzfyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SN-AbsTopK, a sparse autoencoder for medical foundation-model representations. Its main contribution is a group-level support selection, where spatially related tokens share a candidate dictionary before individual AbsTopK selection, together with normative centring using an EMA reference.

Critical strengths

Addresses spatial fragmentation in conventional SAE representations with a simple group-support mechanism. Ablations indicate that group selection substantially improves spatial consistency.

Critical weaknesses

1 The overall figure has poor layout/resolution and is difficult to follow; figure sizes are also inconsistent.

2 Experimental discussions do not always clearly reference the corresponding tables, making results difficult to trace.

3 Improved Moran’s I and Jaccard is partly expected because spatial coherence is explicitly imposed by group support. Stronger evidence of medical interpretability is needed. How this benifit the downstream task and reconstruction?

4 The motivation and benefit of normative residual centring remain unclear.

5 The substantial reduction in alive feature raises concerns about dictionary utilisation and capacity. How it affects representation diversity, reconstruction, and interpretability?

6 Results are reported over only two seeds, raising questions about statistical robustness.

Questions for authors

1 Can the authors better demonstrate that increased spatial coherence corresponds to meaningful anatomical/pathological concepts?

2 What is the impact of the substantial reduction in alive atoms on representation capacity, reconstruction and interpretability?

3 Does “two seeds” mean only two independent training runs for the reported mean and standard deviation?

4 Clearer table references and a better organised, higher-resolution overall framework figure would make the relationships among the different components easier to follow..

5 Please provide clearer visualisations comparing grouped and non-grouped representations across contiguous histology regions and MRI slices. This would better demonstrate the actual effect of group-level support selection.

Reviewer_LMTvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents SN-AbsTopK, a type of sparse autoencoder designed for the features of medical foundation models. First, a common group of candidate atoms is selected for each spatial group based on energy, and then each token performs a signed TopK operation within that group. Furthermore, an EMA reference anchor and a TV term are employed. The authors achieve improvements in Moran's I, the adjacent-slice Jaccard, and the linear-probe AUROC when evaluated on BCSS (using UNI features) and on CheXpert and Duke Breast MRI (using BiomedCLIP features).

Critical strengths

• The motivation is valid since medical image features have spatial structure and selecting supports one token at a time loses that structure. The method is easy to implement and, when singleton groups are used together with a group budget of k, it is exactly equivalent to AbsTopK, which is a straightforward property. • The experiments involve pathology, X-ray and MRI studies and include ablations of components. Some of the BCSS latents clearly correspond to tissue types. Critical Weaknesses

Critical weaknesses

• The figures and the figures given do not agree with one another. While the main table says that there is a 4x dictionary with an L0 of 32, the appendix states 16x and k=24 for BiomedCLIP. The values given in the 'Full' row under BCSS in the ablation section do not match those reconstructed from the main table. The baseline figures mentioned in the text are different from those in the tables. Some of the bold entries are incorrect; for example, on Duke, Mat. BatchTopK has a better nMSE and EV yet the bold mark is assigned to the proposed method. • The fact that the dead-atom rate is presented as an advantage is noteworthy. When it says "fewest alive features", it means that about eighty per cent of the atoms are dead, a term which the authors use themselves to describe lock-out. On each of the datasets the number of alive atoms is less than the input dimension, so the dictionary is no longer overcomplete, which goes against the idea of an SAE. • The fact that coherence is achieved is mainly due to the design of the system; since all the tokens in a group choose from the same small set of candidates, the Jaccard and Moran's I values between neighboring tokens increase almost automatically. This does not mean that more meaningful structure has been learned. • The evidence regarding the downstream results is meagre and inconsistent; the downstream tables appear as 'Table ??' and therefore cannot be verified. Only AbsTopK is compared, with no investigation carried out on the raw FM embeddings. Although the macro-AUROC in the main table for CheXpert AUROC-5 is basically identical to that of AbsTopK, the text states there is a clear improvement in macro-AUROC over AbsTopK. Performance on ER and Grade in the Duke dataset has declined. • The innovation is minimal; at its heart the approach involves TopK with a group-shared mask. The EMA anchor actually achieves very little with respect to standardized features, as the authors themselves admit. There is nothing in the method that corresponds to 'unrolled' as stated in the title; the appendix explains that a single forward pass is carried out at each step. The lemmas are simply well-known results.

Questions for authors
  1. What dictionary size and value of k were actually used? Why is the entry for Full in the ablation study different from the one in the main table? 2. Does the test-time encoding make use of other samples from the same group? What will occur if the baselines experience the same pooling? 3. How can the increase in clear CheXpert macro-AUROC and the identical AUROC-5 shown in the main table be reconciled? What does an investigation of the raw embeddings reveal?
Reviewer_WJCfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Spatial-Normative AbsTopK (SN-AbsTopK), an unrolled sparse autoencoder that incorporates spatial group support constraints and normative reference centering into signed-magnitude dictionary learning. The method employs a two-stage feature selection mechanism to capture spatially coherent representations across contiguous regions, such as histopathology tiles and volumetric MRI slices. Specifically, the first stage aggregates pre-activation energies across tokens within each spatial group to identify a shared candidate set of dictionary atoms. The second stage performs token-wise signed-magnitude selection restricted to this candidate set, ensuring that token-level active features remain within the group-level support. Additionally, an exponential moving average (EMA) reference buffer is introduced to center input embeddings, encouraging sparse representations to capture deviations from normative reference states. Experiments across histopathology and radiology foundation models demonstrate the potential effectiveness of the proposed approach.

Critical strengths
  1. The paper introduces a two-stage sparse feature selection mechanism that combines group-level candidate support construction with token-wise signed-magnitude selection. The integration of an EMA-based normative reference further enables residual sparse coding, providing a structured approach to learning spatially coherent representations.
  2. The paper provides a relatively comprehensive evaluation across histopathology and radiology foundation model representations, covering reconstruction quality, sparsity, spatial consistency, downstream decodability, and component ablations. These experiments help characterize the contributions of group support selection and normative centering.
Critical weaknesses
  1. The manuscript contains several presentation and formatting issues that should be carefully addressed. In Lines 351–352, unresolved cross-references appear as "?". In Lines 502, 543, and 558, citation placeholders marked as "[VERIFY]" remain in the manuscript, which is unacceptable for a submitted paper. Equation (7) also contains an unexpected leading "]", making the mathematical expression difficult to interpret. Additionally, the appendix and author biographies appear to be improperly arranged, with inconsistent page numbering.
  2. The presentation of experimental results lacks sufficient clarity, particularly in the qualitative comparisons. For example, in Figure 3, the authors should explicitly highlight the relevant regions and key visual differences in the chest X-ray images. Without clear annotations or explanations, it is difficult to identify the improvements demonstrated by the proposed method or assess whether the visual comparisons adequately support the authors' claims.
Questions for authors

Please refer to the weaknesses

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

Related papers

Discussion (0)

Sign in to comment.