Mammo-SAE: Interpreting Breast Cancer Concept Learning with Sparse Autoencoders

Krishna Kanth Nakka

REVIEW 3 major objections 5 minor 22 references

Mammo-SAE: Interpreting Breast Cancer Concept Learning with Sparse Autoencoders

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A sparse autoencoder reveals that just ten latent neurons carry most of the breast-cancer signal in Mammo-CLIP.

desk verdict A useful first transfer of SAEs to mammography with two concrete neuron-level findings, but the headline intervention result needs a random-neuron baseline and a real train/test split before it can be trusted. read the letter →

arxiv 2507.15227 v2 pith:XRK6OT45 submitted 2025-07-21 cs.CV

Krishna Kanth Nakka This is my paper

classification cs.CV

keywords SparseAutoencodersExplainableAIBreastCancerImagingMammographyConceptProbingLatentNeuronInterventionVision-LanguageFoundationModel

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

This paper tries to show that a sparse autoencoder inserted into the final layer of the Mammo-CLIP breast-imaging model can expose which internal neurons actually encode clinically relevant concepts such as mass and suspicious calcification. The central claim is that a very small set of latent neurons, often as few as ten, captures most of the signal the model uses for these predictions, since keeping only those neurons nearly restores the original AUC-ROC and deleting them makes accuracy collapse. The paper further claims that top class-level neurons frequently align with ground-truth lesion regions after finetuning, while a few persistent background neurons reveal confounds the model continues to rely on. If true, this gives clinicians a concrete way to see and control what a black-box breast-cancer model is looking at.

What carries the argument

The load-bearing object is the Mammo-SAE: an encoder-decoder sparse autoencoder trained with a reconstruction loss plus an $\ell^1$ sparsity penalty, mapping patch-level CLIP features $x_j \in \mathbb{R}^{2048}$ to a sparse latent $z \in \mathbb{R}^{16384}$ through a ReLU encoder. Class-level relevance scores rank latent neurons by their mean activation over the positive class, and group interventions (top-k activated keeps only the selected neurons; top-k deactivated zeroes them out) turn correlation into a causal test by passing edited latents through the decoder to the classifier. This machinery lets the paper identify, localize, and functionally ablate concept neurons without retraining the vision model.

What would settle it

Train a separate sparse autoencoder on final-layer features from the frozen pretrained Mammo-CLIP model and another on the finetuned model, then repeat the class-separation plots and the top-10 activation and deactivation curves; if the finetuned-versus-pretrained separation and the ten-neuron recovery pattern disappear when each model is measured with its own autoencoder, the conclusions are artifacts of a single SAE specialized to the finetuned feature distribution rather than properties of the models.

Watch

Extended reading notes

Core claim

Trained on final-layer patch features of the finetuned Mammo-CLIP suspicious-calcification classifier, a sparse autoencoder with 16,384 latent neurons learns representations whose top class-specific neurons are functionally decisive. Activating only the top-10 neurons for the two classes almost reproduces the model's full AUC-ROC in several settings, and zeroing those neurons out causes a sharp drop, indicating that the model relies on a compact, concept-aligned subset rather than a diffuse feature spread. For suspicious calcification, seven of the top ten latent neurons activate inside the annotated region; for mass, alignment is weaker, consistent with the model's lower AUC. Finetuning the backbone enlarges the separation between class-mean latent activation vectors and yields higher localization mean average precision, which the paper attributes to the model learning more discriminative, spatially meaningful features while still retaining at least one background-sensitive neuron that fires across classes and models.

Load-bearing premise

The conclusions about finetuning and about the mass concept rest on a single sparse autoencoder trained only on final-layer activations of the finetuned suspicious-calcification model being an unbiased measuring instrument for all four model-and-concept settings.

Editorial extensions

If this is right

Keeping only the top-k class-specific SAE neurons nearly recovers the original AUC-ROC, so a compact set of latent features is sufficient for mass and calcification prediction.
Zeroing those top-k neurons sharply drops AUC-ROC, showing the model causally depends on those specific neurons.
In the finetuned model, top-10 latent neurons achieve higher localization mAP than in the pretrained model, indicating that finetuning aligns concept features with ground-truth regions.
Finetuning makes the class-mean latent vectors more separated, offering an explanation for its AUC gain and improved interpretability.
A persistent background-sensitive neuron remains active across both classes and both model variants, showing that confounding reliance survives finetuning.

Reading between the lines

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

The paper reuses one SAE trained only on the finetuned suspicious-calcification model for all comparisons; I would not trust the finetuned-versus-pretrained separation until it is reproduced with an SAE trained separately on each model.
If the top-k intervention generalizes, one could build a concept-bottleneck classifier from a few dozen SAE neurons and get nearly the same accuracy, which would make high-stakes screening models auditable by radiologists.
A natural testable extension is to deactivate the persistent background neuron at inference and measure whether robustness or specificity improves; the paper's data suggest it is spurious but do not run that experiment.
Applying the same probe to earlier layers and to other concepts such as nipple retraction or skin thickening would test whether monosemantic concept neurons exist throughout the backbone rather than only at the final layer.

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.

Desk Editor's Note

read the letter

Sarah,

This is a genuine first: applying sparse autoencoders to a breast imaging foundation model, with two specific observations that are worth keeping—neuron 13867 tracks suspicious calcification after finetuning, and neuron 15699 is a persistent background confound. The reconstruction sanity check (Table 1) is a good idea, and the <2% AUC drop suggests the SAE preserves task-relevant information. For the finetuned calcification model, several of the top-10 heatmaps do land on annotated regions. That is real evidence of monosemantic-like behavior in a medical model.

The soft spots are concentrated in Section 3.1, and they are load-bearing. The top-k neurons are ranked by class-mean activation, presumably on training images, but the paper never states that the post-intervention AUC is computed on a disjoint set. If ranking and evaluation share data, the "10 neurons recover AUC" result is close to mechanical: you have selected the dimensions that best separate the labels on the evaluation set itself. The absence of any random-neuron control makes that interpretation impossible to rule out. This is not a nitpick.

The shared-SAE issue is real but secondary. Training one SAE on finetuned suspicious-calcification features and reusing it for the pretrained model and mass concept could bias the finetuning-versus-pretrained comparisons in Figures 4 and 5. The paper states the cost rationale, which is fair, but it never checks whether the SAE is an unbiased instrument for both feature distributions.

Table 2 is another weak spot: the mAP values are low for mass and for the pretrained model, and the "7/10 neurons align" claim comes only from finetuned calcification. The abstract's "often tend to align" overstates the general pattern. Error bars are absent throughout, which matters for the intervention curves.

All of this is fixable. Add a random-neuron control, describe and enforce a split between neuron selection and evaluation, train per-model SAEs or at least show matched reconstruction quality, and temper the abstract. The core finding is plausible and the domain transfer is valuable, so I would not desk-reject it. Send it to reviewers who know SAE evaluation practices and ask them to focus on the intervention baseline and the data split. It deserves a serious referee.

Best, [Your name]

Referee Report

3 major / 5 minor

Summary. The paper introduces Mammo-SAE, a sparse autoencoder trained on patch-level final-layer activations of the Mammo-CLIP breast imaging foundation model, and uses it to identify latent neurons associated with the concepts of mass and suspicious calcification. The authors report that top class-level neurons align with ground-truth regions, that interventions on a small number of top-k neurons can largely recover or destroy classification AUC, and that finetuning produces stronger class separation in the SAE latent space. The main evidence is the intervention analysis in Section 3.1 and Figure 2, the localization mAP in Table 2, and the latent-vector comparisons in Section 3.3.

Significance. If the central intervention claim holds, the paper would be a useful demonstration that sparse autoencoders can localize compact, concept-relevant subspaces in medical vision-language models, with practical implications for auditing confounding behavior and for post-hoc intervention. The reconstruction sanity check in Table 1 is a genuine strength: it shows that inserting the SAE at the final layer costs less than 2% AUC, which supports the faithfulness of the SAE as a probe for the finetuned suspicious-calcification model. The intervention framework itself is appropriate in spirit. However, the claimed quantitative result that 'as few as 10 neurons' capture most of the task-relevant signal is currently undermined by the absence of a clear train/test separation in the neuron-ranking procedure and by the lack of a random-neuron baseline; the cross-model comparison is further weakened by reusing a single SAE trained on one model's features. These issues are load-bearing rather than cosmetic.

major comments (3)

[Section 3.1 and Eq. (2)] The central intervention result may have selection leakage. The top-k neurons are ranked by class-mean activation computed in Eq. (2) over 'training samples,' but Section 3.1 does not state that the images used to rank neurons are disjoint from the images used to compute the post-intervention AUC in Figure 2. If the ranking and evaluation sets overlap, the top-k activated and deactivated curves become a near-mechanical consequence of selecting the dimensions that best separate the labels on the evaluation set itself, rather than evidence that a compact concept-aligned subspace is causally responsible for the model's predictions. The paper should either explicitly state that a held-out split was used for AUC evaluation, or repeat the analysis with neurons ranked on a training split and evaluated on a test split. In addition, the intervention curves need a random-neuron baseline: without it, the reader cannot tell whether selecting any k high-variance directions would produce a similar AUC recovery, which is essential for interpreting the '10 neurons suffice' claim.
[Section 3.3 and Section 3, SAE Training] The conclusion that finetuning sharpens class separation in the SAE latent space (Figures 4a, 4b, and Figure 5) is potentially an artifact of the measurement instrument. The paper trains a single SAE on features from the finetuned suspicious-calcification model (d=2048, h=16384) and then reuses that same SAE to analyze the pretrained model and the mass concept. If the SAE has specialized to the finetuned feature distribution, its latent space may exaggerate structure in the finetuned model's activations while misreconstructing the pretrained model's activations. Table 1 shows only that AUC is preserved when the frozen model is run through the SAE, not that the latent geometry is unbiased across both feature distributions. The authors should either train a separate SAE per model and per concept, or provide evidence—for example, matched reconstruction error and per-neuron consistency across models—that the shared SAE is a neutral measurement instrument.
[Table 2 and Section 3.2] The quantitative localization evidence is weaker than the narrative suggests. Table 2 reports mAP at IoU=0.25 for the top-10 neurons, but the per-neuron mAP values are generally low—for example, finetuned suspicious calcification has values between 0.005 and 0.278, and the mass pretrained row ranges from 0.018 to 0.053. The paper claims that '7 out of the top 10 latent neurons activate strongly within the annotated region' for suspicious calcification, but this claim is not directly backed by the mAP numbers, which are not averaged and are shown for only a small set of representative images. To support the alignment claim, the authors should report the fraction of neurons above a pre-specified mAP threshold over the full evaluation set, and should clarify whether the 7/10 statement refers to qualitative inspection of the figures or to a quantitative criterion.

minor comments (5)

[Eq. (3)] The notation is inconsistent: the text says T(c)k is a subset of {1,...,k}, but the top-k neurons should be a subset of {1,...,h} of cardinality k. Please correct the set definition.
[Section 2.1] There is a typo in 'Further implementaton details' and the term 'calciﬁcation' has a ligature/formatting artifact in the appendices. Please proofread the final typeset version.
[Figure 8 caption] The caption for Figure 8 in the appendix reads 'Mass Calcification,' but the concept under study is 'Mass'; suspicious calcification is a separate concept. This appears to be a copy-paste error from Figure 6 and should be corrected.
[Section 2.2] The notation is mixed: the dataset is described as {x, c} with x an image, but in Eq. (2) xj is a CLIP feature at spatial location j. Please disambiguate image-level and patch-level notation for readability.
[Appendix, Figure 4 and Figure 5] The y-axes of the mean latent activation plots are unlabeled, and the x-axis is labeled only in units of 'Latent Neuron.' Adding axis labels and a caption describing the normalization would make the 'separation' claim easier to evaluate.

Circularity Check

1 steps flagged · score 7.0 of 10

Section 3.1 intervention claim reduces to in-sample top-k selection as written; no disjoint evaluation set or random-neuron control is reported.

fitted input called prediction [Section 2.2 Eq. (2); Section 2.3 Eqs. (3)-(4); Section 3.1 Fig. 2]
"we compute the class-wise mean latent activation vector ¯z(c) ∈ Rh by averaging over all spatial locations and all training samples in that class ... Dc is the set of training images with class label c ... Latent neurons are ranked in descending order of s(c)t ... we compute the relevance score of each latent neuron with respect to two classes ... identify the top-k neurons per class. We then perform targeted interventions ..."

The top-k set is defined by Eq. (2) as the k latent coordinates with the largest class-mean activation computed on labeled 'training samples' for each class. The Fig. 2 AUC curves are then measured after keeping exactly those coordinates (Eq. (3)) or zeroing them (Eq. (4)). The paper reports no split between the Dc used in Eq. (2) and the images over which post-intervention AUC-ROC is computed. On the literal reading, the 'top-k' coordinates are the k directions that best separate the two classes on the evaluation set itself, so retaining only those directions preserves most of the linear class signal and removing them destroys it.

full rationale

The paper's central quantitative claim is the Section 3.1 intervention experiment. As written, the procedure selects top-k neurons by Eq. (2) on labeled training samples and then evaluates AUC after masking to those same neurons (Eqs. (3)-(4)) without stating that the AUC evaluation uses a disjoint set. Under that reading, the 'small subset of neurons captures the signal' result is an in-sample feature-selection artifact: the chosen coordinates are precisely those that maximize class separation on the evaluation set. This warrants a high circularity score. I did not find other load-bearing circularity: the paper does not rely on author self-citations, uniqueness theorems, or renamed prior results; the shared-SAE design (Section 3) is a measurement-comparability concern rather than a circular reduction, because the SAE is trained with an unsupervised reconstruction loss, not on the labels used for ranking. If a proper train/validation/test split were reported and the AUC curves were computed on a held-out set, the intervention result would be a legitimate supervised feature-selection finding rather than circular; however, the text as it stands does not document such a split.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard SAE assumptions plus two paper-specific choices: a single SAE reused across models and concepts, and an unstated train/test separation for neuron selection. The free parameters are hyperparameters and thresholds that shape which neurons are called important.

free parameters (5)

Sparsity penalty lambda = 3e-5
Set during SAE training (Section 3); controls how many latent neurons are active and therefore which neurons are considered interpretable.
Latent dimension expansion factor = 8 (h=16384)
Architecture choice (Section 3); determines the number of candidate concept neurons.
Top-k neuron count for interventions and localization = 10
The headline finding that '10 neurons suffice' and the mAP table both depend on this hand-chosen k; scanning k does not remove the dependence.
Heatmap threshold for bounding boxes = 95th percentile
Used to convert latent activation heatmaps to predicted boxes for mAP (Section 3.2); changing it changes all localization numbers.
SAE training schedule = lr=3e-4, batch=4096, 200 epochs
Standard training hyperparameters; listed for completeness, not central to the scientific claim.

assumptions (5)

domain assumption Mammo-CLIP final-layer patch features are a sufficient substrate for concept-level interpretability.
The SAE is inserted only at the final layer of the EfficientNet-B5 backbone, and the paper assumes concept information is accessible there (Section 2.1, Limitations).
domain assumption VinDr-Mammo annotations for mass and suspicious calcification are accurate and define the target concepts.
Class-wise mean activations and mAP are computed against these labels; no annotation noise or reader variability is modeled.
ad hoc to paper A single SAE trained on finetuned suspicious-calcification features transfers without bias to pretrained and mass features.
Reuse is explicit in Section 3 ('train a single SAE once and reuse it across all experiments') and underpins the pretrained vs finetuned comparison in Section 3.3; this is a paper-specific design choice, not a validated general property.
domain assumption ReLU SAE with L1 sparsity yields monosemantic features in this medical domain.
Assumed from the LLM/VLM SAE literature; the paper provides no quantitative monosemanticity metric, only visual inspection of heatmaps.
domain assumption Class-level neuron ranking and evaluation are performed on disjoint data.
No dataset split is described anywhere; all AUC and mAP results are single numbers. The validity of the intervention and localization claims depends on this unstated assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mammo-SAE: Interpreting Breast Cancer Concept Learning with Sparse Autoencoders." pith.science (2026). https://pith.science/paper/XRK6OT45

@misc{pith2026250715227,
  author       = {Pith},
  title        = {Pith review of: Mammo-SAE: Interpreting Breast Cancer Concept Learning with Sparse Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XRK6OT45}},
  note         = {Machine review of arXiv:2507.15227}
}

read the original abstract

Interpretability is critical in high-stakes domains such as medical imaging, where understanding model decisions is essential for clinical adoption. In this work, we introduce Sparse Autoencoder (SAE)-based interpretability to breast imaging by analyzing {Mammo-CLIP}, a vision--language foundation model pretrained on large-scale mammogram image--report pairs. We train a patch-level \texttt{Mammo-SAE} on Mammo-CLIP to identify and probe latent features associated with clinically relevant breast concepts such as \textit{mass} and \textit{suspicious calcification}. Our findings reveal that top activated class level latent neurons in the SAE latent space often tend to align with ground truth regions, and also uncover several confounding factors influencing the model's decision-making process. Additionally, we analyze which latent neurons the model relies on during downstream finetuning for improving the breast concept prediction. This study highlights the promise of interpretable SAE latent representations in providing deeper insight into the internal workings of foundation models at every layer for breast imaging. The code will be released at https://krishnakanthnakka.github.io/MammoSAE/

Figures

Figures reproduced from arXiv: 2507.15227 by the authors.

**Figure 1.** Mammo-SAE Framework. The SAE is first trained on patch-level CLIP features xj ∈ Rd at any given layer, projecting them into a highdimensional, interpretable sparse latent space z ∈ Rh , and decoding them back for reconstruction. Once trained, the SAE is used to analyze which latent neurons are activated and what semantic information they encode. We also perform targeted interventions in the latent neuron space to … view at source ↗

**Figure 2.** Intervention on class-level latent neurons. Left: Top-k activated intervention—only the top-k class-specific neurons are retained, and all others are zeroed out. Right: Top-k deactivated intervention—the top-k neurons are zeroed out while the rest remain unchanged [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Visualization of the top-10 class-level latent neurons of class [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Mean latent activation vectors z¯ (c) for each class (c = 1 indicates the presence of the concept) in the pretrained model (left) and finetuned model (right) for the suspicious calcification concept. 4 Conclusion In this paper, we introduced Mammo-SAE, a framework for …

**Figure 5.** Figure 5: Mean latent activation vectors z¯ (c) for each class (c = 1 indicates the presence of the concept) in the pretrained model (left) and finetuned model (right) for the mass concept [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

**Figure 6.** Figure 6: Visualization of Top-10 Latent Neurons of class c = 1 for Suspicious Calcification (Finetuned Model). Red boxes indicate ground-truth annotations [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

**Figure 7.** Figure 7: Visualization of Top-10 Latent Neurons of class c = 1 for Suspicious Calcification (Pretrained Model). Red boxes indicate ground-truth annotations [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

**Figure 8.** Figure 8: Visualization of Top-10 Latent Neurons of class c = 1 for Mass Calcification (Finetuned Model). Red boxes indicate ground-truth annotations [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

**Figure 9.** Figure 9: Visualization of Top-10 Latent Neurons of class c = 1 for Mass Calcification (Pretrained Model). Red boxes indicate ground-truth annotations [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 11 canonical work pages

[1]

Scien- tific Reports 14(1), 10341 (2024)

Choukali, M.A., Amirani, M.C., Valizadeh, M., Abbasi, A., Komeili, M.: Pseudo- class part prototype networks for interpretable breast cancer classification. Scien- tific Reports 14(1), 10341 (2024)

work page 2024
[2]

arXiv preprint arXiv:2309.08600 (2023)

Cunningham, H., Ewart, A., Riggs, L., Huben, R., Sharkey, L.: Sparse autoen- coders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600 (2023)

arXiv 2023
[3]

In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

Dreyer, M., Purelku, E., Vielhaben, J., Samek, W., Lapuschkin, S.: Pure: Turning polysemantic neurons into pure features by identifying relevant circuits. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 8212–8217 (2024)

work page 2024
[4]

arXiv preprint arXiv:2406.04093 (2024)

Gao, L., la Tour, T.D., Tillman, H., Goh, G., Troll, R., Radford, A., Sutskever, I., Leike, J., Wu, J.: Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093 (2024)

arXiv 2024
[5]

In: International conference on medical image computing and computer- assisted intervention

Ghosh, S., Poynton, C.B., Visweswaran, S., Batmanghelich, K.: Mammo-clip: A vi- sion language foundation model to enhance data efficiency and robustness in mam- mography. In: International conference on medical image computing and computer- assisted intervention. pp. 632–642. Springer (2024)

work page 2024
[6]

arXiv preprint arXiv:2410.20526 (2024)

He, Z., Shu, W., Ge, X., Chen, L., Wang, J., Zhou, Y., Liu, F., Guo, Q., Huang, X., Wu, Z., et al.: Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders. arXiv preprint arXiv:2410.20526 (2024)

arXiv 2024
[7]

In: Proceedings of the IEEE/CVF international conference on computer vision

Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3942–3951 (2021)

2021
[8]

Kajala, A., Jaiswal, S., Kumar, R.: ‘breaking the black box: Heatmapdriven trans- parency to breast cancer detection with efficientnet and grad cam. Educ. Admin., Theory Pract 30(5), 4999–5009 (2024)

work page 2024

Show all 22 references

[9]

In: Convolutional neural networks with swift for Tensor- flow: image recognition and dataset categorization, pp

Koonce, B.: Efficientnet. In: Convolutional neural networks with swift for Tensor- flow: image recognition and dataset categorization, pp. 109–123. Springer (2021) 1 https://centrescancer.chuv.ch/equipe/rencontrer-lequipe-du-centre-du-sein/ Mammo-SAE: Interpreting Breast Cance...

2021
[10]

arXiv preprint arXiv:2408.05147 (2024)

Lieberum, T., Rajamanoharan, S., Conmy, A., Smith, L., Sonnerat, N., Varma, V., Kramár, J., Dragan, A., Shah, R., Nanda, N.: Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2. arXiv preprint arXiv:2408.05147 (2024)

2024 arXiv
[11]

arXiv preprint arXiv:2412.05276 (2024)

Lim, H., Choi, J., Choo, J., Schneider, S.: Sparse autoencoders reveal selective remapping of visual concepts during adaptation. arXiv preprint arXiv:2412.05276 (2024)

2024 arXiv
[12]

IEEE Access (2024)

Liu, S., Himel, G.M.S., Wang, J.: Breast cancer classification with enhanced inter- pretability: Dalaresnet50 and dt grad-cam. IEEE Access (2024)

2024
[13]

arXiv preprint arXiv:2405.08366 (2024)

Makelov, A., Lange, G., Nanda, N.: Towards principled evaluations of sparse au- toencoders for interpretability and control. arXiv preprint arXiv:2405.08366 (2024)

2024 arXiv
[14]

arXiv preprint arXiv:2401.17975 (2024)

Marshall, S.C., Kirchner, J.H.: Understanding polysemanticity in neural networks through coding theory. arXiv preprint arXiv:2401.17975 (2024)

2024 arXiv
[15]

Scientific Data10(1), 277 (2023)

Nguyen, H.T., Nguyen, H.Q., Pham, H.H., Lam, K., Le, L.T., Dao, M., Vu, V.: Vindr-mammo: A large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography. Scientific Data10(1), 277 (2023)

2023
[16]

O’Mahony, L., Andrearczyk, V., Müller, H., Graziani, M.: Disentangling neuron representationswithconceptvectors.In:ProceedingsoftheIEEE/CVFConference on Computer Vision and Pattern Recognition. pp. 3770–3775 (2023)

2023
[17]

In: World Conference on Explainable Artificial Intelligence

Pathak, S., Schlötterer, J., Veltman, J., Geerdink, J., van Keulen, M., Seifert, C.: Prototype-based interpretable breast cancer prediction models: Analysis and challenges. In: World Conference on Explainable Artificial Intelligence. pp. 21–42. Springer (2024)

2024
[18]

Multimedia Tools and Applications 83(19), 57551–57578 (2024)

Raghavan, K., B, S., v, K.: Attention guided grad-cam: an improved explainable artificial intelligence model for infrared breast cancer detection. Multimedia Tools and Applications 83(19), 57551–57578 (2024)

2024
[19]

In: European Conference on Computer Vision

Rao, S., Mahajan, S., Böhle, M., Schiele, B.: Discover-then-name: Task-agnostic concept bottlenecks via automated concept discovery. In: European Conference on Computer Vision. pp. 444–461. Springer (2024)

2024
[20]

arXiv preprint arXiv:2403.08295 (2024)

Team, G., Mesnard, T., Hardin, C., Dadashi, R., Bhupatiraju, S., Pathak, S., Sifre, L., Rivière, M., Kale, M.S., Love, J., et al.: Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295 (2024)

2024 arXiv
[21]

arXiv preprint arXiv:2302.13971 (2023)

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

2023 arXiv
[22]

In: Proceedings of the Conference on Empirical Methods in Natural Language Processing

Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from un- paired medical images and text. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing. vol. 2022, p. 3...

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Scien- tific Reports 14(1), 10341 (2024)

Choukali, M.A., Amirani, M.C., Valizadeh, M., Abbasi, A., Komeili, M.: Pseudo- class part prototype networks for interpretable breast cancer classification. Scien- tific Reports 14(1), 10341 (2024)

work page 2024

[2] [2]

arXiv preprint arXiv:2309.08600 (2023)

Cunningham, H., Ewart, A., Riggs, L., Huben, R., Sharkey, L.: Sparse autoen- coders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600 (2023)

arXiv 2023

[3] [3]

In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

Dreyer, M., Purelku, E., Vielhaben, J., Samek, W., Lapuschkin, S.: Pure: Turning polysemantic neurons into pure features by identifying relevant circuits. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 8212–8217 (2024)

work page 2024

[4] [4]

arXiv preprint arXiv:2406.04093 (2024)

Gao, L., la Tour, T.D., Tillman, H., Goh, G., Troll, R., Radford, A., Sutskever, I., Leike, J., Wu, J.: Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093 (2024)

arXiv 2024

[5] [5]

In: International conference on medical image computing and computer- assisted intervention

Ghosh, S., Poynton, C.B., Visweswaran, S., Batmanghelich, K.: Mammo-clip: A vi- sion language foundation model to enhance data efficiency and robustness in mam- mography. In: International conference on medical image computing and computer- assisted intervention. pp. 632–642. Springer (2024)

work page 2024

[6] [6]

arXiv preprint arXiv:2410.20526 (2024)

He, Z., Shu, W., Ge, X., Chen, L., Wang, J., Zhou, Y., Liu, F., Guo, Q., Huang, X., Wu, Z., et al.: Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders. arXiv preprint arXiv:2410.20526 (2024)

arXiv 2024

[7] [7]

In: Proceedings of the IEEE/CVF international conference on computer vision

Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3942–3951 (2021)

2021

[8] [8]

Kajala, A., Jaiswal, S., Kumar, R.: ‘breaking the black box: Heatmapdriven trans- parency to breast cancer detection with efficientnet and grad cam. Educ. Admin., Theory Pract 30(5), 4999–5009 (2024)

work page 2024