Skip to content
Computational Pathology Reading Roadmap

Image foundations and recognition

Image recognition compresses a spatial measurement into a prediction. Before choosing a backbone, understand how the measurement was formed, sampled, transformed, and split. Otherwise a model may learn resampling artifacts, background, or capture conditions instead of the object.

Latest literature discovery audit: 2 October 2026. Existing paper reviews retain their individual source-check dates.

ECCV 2026 conference additions and metadata checked through 2026-09-12; the broader literature audit boundary above is unchanged.

Papers remain organized by mechanism: image formation and matching, learned convolutional recognition, efficient scaling, and transfer.

Prerequisites and objectives

Section titled “Prerequisites and objectives”

You should know vectors, matrices, derivatives, probability, convolution, gradient descent, and train/validation/test separation. By the end, you should be able to:

  • trace coordinates and values from a scene through sampling into a tensor;
  • compute convolutional output shapes and receptive fields;
  • explain invariance versus equivariance and when augmentation changes the label;
  • compare depth, width, resolution, connectivity, and efficiency independently;
  • design a controlled scratch, frozen-transfer, and full-fine-tuning comparison.

From a scene to a tensor

Section titled “From a scene to a tensor”

A pinhole camera maps a 3D point X to a homogeneous image coordinate x̃ ∼ K[R | t]X̃. Exposure, optics, sensor response, demosaicing, color conversion, compression, and resizing then alter the sampled values. A digital image is consequently evidence produced by a pipeline, not the scene itself.

For a 2D convolution with kernel k, stride s, padding p, and dilation d, one output dimension is

floor((n + 2p − d(k − 1) − 1) / s + 1)

Repeated stride or pooling increases the effective receptive field but discards spatial samples. Downsampling without a suitable low-pass operation creates aliasing. Color conversion and JPEG quality can become unintended dataset labels.

Useful open references are the Szeliski textbook, MIT Vision Book, and CS231n notes.

Classical detection and correspondence

Section titled “Classical detection and correspondence”

Classical vision makes the engineered assumptions visible. Edge filters respond to local derivatives; scale-space descriptors seek repeatable points; RANSAC separates a geometrically consistent hypothesis from outliers.

Stage Primary reading · year · venue/status Read it for
A1 Random Sample Consensus — 1981 · Communications of the ACM, peer reviewed robust hypothesis fitting, minimal samples, inliers, and reprojection error
A2 A Computational Approach to Edge Detection — 1986 · IEEE TPAMI, peer reviewed smoothing–localization trade-offs, non-maximum suppression, and hysteresis
A3 Distinctive Image Features from Scale-Invariant Keypoints — 2004 · IJCV, peer reviewed scale-space extrema, orientation normalization, descriptors, and matching
A4 Histograms of Oriented Gradients for Human Detection — 2005 · CVPR, peer reviewed local gradient statistics and the lineage from features to sliding-window recognition
A5 ImageNet: A Large-Scale Hierarchical Image Database — 2009 · CVPR, peer reviewed dataset scale, label hierarchy, benchmark construction, and dataset-induced progress

Do not treat hand-crafted and learned features as mutually exclusive eras. Classical matching remains a strong diagnostic for registration and geometry; engineered baselines expose whether a deep model is using color, texture, or shape.

Convolutional recognition

Section titled “Convolutional recognition”

A CNN shares a kernel over positions:

y_(o,i,j) = σ(b_o + Σ_c Σ_(u,v) w_(o,c,u,v) x_(c,i+u,j+v))

Weight sharing is translation equivariance before boundary effects, stride, pooling, and global aggregation. Classification invariance is produced by the whole training and pooling system, not by convolution alone.

Stage Primary reading · year · venue/status Architectural question
B1 ImageNet Classification with Deep Convolutional Neural Networks — 2012 · NeurIPS, peer reviewed which gains came from learned depth, scale, nonlinearity, augmentation, and compute?
B2 Very Deep Convolutional Networks for Large-Scale Image Recognition — 2015 · ICLR, peer reviewed what does repeated small-kernel depth buy, and what is its compute cost?
B3 Going Deeper with Convolutions — 2015 · CVPR, peer reviewed how do multi-branch, multi-scale modules trade representation for efficiency?
B4 Deep Residual Learning for Image Recognition — 2016 · CVPR, peer reviewed why does learning a residual path improve optimization at depth?
B5 Densely Connected Convolutional Networks — 2017 · CVPR, peer reviewed how do feature reuse and concatenation differ from residual addition?
B6 EfficientNet — 2019 · ICML, peer reviewed why must depth, width, and resolution be scaled together under a resource budget?
B7 An Image is Worth 16×16 Words — 2021 · ICLR, peer reviewed when can patch tokens and global attention replace convolutional inductive bias?
B8 Training Data-Efficient Image Transformers — 2021 · ICML, peer reviewed how do augmentation, regularization, and distillation change a transformer’s data requirement?
B9 Swin Transformer — 2021 · ICCV, peer reviewed how do shifted local windows restore hierarchy and multiscale feature maps?
B10 A ConvNet for the 2020s — 2022 · CVPR, peer reviewed which modern training and architectural choices—not attention alone—explain backbone performance?

Training is part of the model

Section titled “Training is part of the model”

Augmentation defines desired invariance. Normalization couples or separates examples. The optimizer, learning-rate schedule, regularization, label smoothing, class sampling, resolution, and pretraining can change conclusions more than a block.

Long-tail teacher bias is an objective problem

Section titled “Long-tail teacher bias is an objective problem”
Primary reading · year · venue/status Controlled question
Distilling Balanced Knowledge from a Biased Teacher — 2026 · CVPR, peer reviewed separate cross-group and within-group teacher bias across head, medium, and tail classes; compare against reweighting, margin, calibration, and non-distilled controls at the same data and capacity

Report per-frequency-group accuracy and calibration alongside the overall score. A student should not inherit a teacher’s head-class confidence and call the transfer successful merely because mean accuracy increases.

Use mixup, CutMix, and RandAugment to ask what information a transformation preserves. Do not stack recipes until the source of a gain becomes unidentifiable.

Transfer modes

Section titled “Transfer modes”
Mode Trainable components Question answered
nearest neighbour none is local geometry in the frozen space useful?
linear probe one affine head are classes linearly accessible?
shallow adapter a small head or selected blocks can limited adaptation expose the representation?
full fine-tuning entire model what can the initialization reach with this target data and budget?
scratch control entire randomly initialized model how much does pretraining contribute under the same recipe?

These are different experiments. A full-tuned score does not prove the frozen representation was strong, and a frozen result does not bound full adaptation.

Controlled lab

Section titled “Controlled lab”

Use CIFAR-100, a licensed ImageNet subset, or another natural-image dataset whose split and class definitions you can publish.

  1. Inspect duplicates, class frequency, resolution, color statistics, and capture sources before training.
  2. Implement Sobel/Canny and SIFT-or-HOG baselines for a small subset.
  3. Compare a ResNet and ConvNeXt with the same input, optimizer family, epoch or update budget, augmentations, and parameter range.
  4. Add a frozen pretrained backbone with a linear head; then fine-tune it.
  5. Report top-1, macro F1, per-class recall, Brier score or calibration curve, parameters, peak memory, measured batch-one latency, and throughput.
  6. Slice errors by class frequency, image size, background, corruption, and source.

The experiment log must state checkpoint, transform order, interpolation, crop policy, normalization, seed, hardware, numerical precision, and selection rule.

Failure modes

Section titled “Failure modes”
  • Background shortcut: the object is removed but the prediction remains.
  • Texture shortcut: style or texture transfer changes the class despite preserved shape.
  • Split leakage: near-duplicates or frames from one source cross the split.
  • Resolution confound: one backbone receives more pixels or tokens.
  • recipe confound: the newer model also gets stronger augmentation and longer training.
  • long-tail blindness: mean accuracy hides failure on rare classes.
  • miscalibration: the correct ranking coexists with unsafe confidence.
  • pretraining overlap: a “zero-shot” example may already occur in web-scale data.

Self-check and exit task

Section titled “Self-check and exit task”

Before leaving this chapter, answer:

  1. What is the difference between equivariance and invariance?
  2. Where can aliasing enter a standard image pipeline?
  3. Why does a residual addition differ from dense concatenation?
  4. Which transfer setting tests representation quality most directly?
  5. Why are FLOPs and parameter count insufficient measures of deployment cost?

Exit task: produce a two-page model record for the controlled lab. Include a receptive-field derivation, an equal-budget table, calibration, failure clusters, and one counterfactual background or texture test.

Pathology bridge

Section titled “Pathology bridge”

The same image pipeline questions become physical in pathology: pixel size, magnification, focus, scanner response, compression, tissue detection, stain, and tile sampling define the evidence. Natural-image augmentation is not automatically label preserving. Carry the measurement record into WSI engineering and compare natural controls against pathology foundation encoders under identical tiles, transforms, splits, and adaptation budgets.

Illuminant coordinates and camera color calibration

CCM-MLP conditions a compact color matrix on two-dimensional illuminant chromaticity, relaxing the traditional one-dimensional color-temperature assumption. Read it as image formation and camera calibration: the key change is the conditioning space, not a generic larger network.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Off the Planckian Locus: Using 2D Chromaticity to Improve In-Camera ColorSaiKiran Tedla, Joshua Little, Hakki C. Karaimer, and Michael S. Brown2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #1413. Publisher proceedings/version of record was not verified on 2026-09-12.CCM-MLP replaces one-dimensional color-temperature interpolation with a two-dimensional illuminant-chromaticity-conditioned color correction matrix. A tiny MLP maps a white point to eight free matrix entries, enabling an ordinary camera pipeline to handle LED illuminants away from the Planckian locus. This is calibrated color mapping, not a new white-balance estimator.ECCV Spotlight program · Conference manuscript · Author project (paper/program link)

Reading question: How does off-Planckian illumination expose the limits of temperature-only interpolation, and who estimates the white point?

Evidence boundary: White-point estimation error and metamerism remain. Angular-error gains can differ from perceptual color-error rankings and do not establish stain-normalization results.

Teacher-class mixtures and privileged residual routing

MoE-KD reformulates classification through latent teacher-class experts and EM. MARS uses full-minus-partial features to teach a deployable router. Specify whether knowledge resides in label mixtures, prototypes, routing decisions or a residual target.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
MoE-KD: Your Teacher Model is Worth Mixture-of-Experts for Knowledge DistillationKuo Shi, Wenjie Zhu, Bo Peng2026 · ECCV · acceptedECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. The EM convergence discussion assumes an improving M-step; it is not a guarantee of global optimum for stochastic neural training. Table 3 values take precedence over inconsistent prose deltas. No runnable official implementation was verified in this pass.MoE-KD treats the teacher’s class as a latent variable in the student classification model. Teacher-derived feature prototypes define class-conditioned experts, while a reused teacher classifier gates projected student features. Alternating posterior estimation and evidence-lower-bound optimization couples these pieces instead of imposing one fixed teacher soft-label distribution alongside the ground-truth target.Official ECCV poster record
Residual-Guided Expert Specialization for Incomplete Multimodal LearningSeunghun Baek, Jihwan Park, Jaeyoon Sim, Minjae Jeong, Hoseok Lee, Won Hwa Kim2026 · ECCV · acceptedECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. MARS distills a privileged residual router into an available-feature router; it is not a cascade of residual-error student networks. Full modalities are assumed available during training. Food-101 full-modality accuracy is below SimMLM despite higher missing-modality and average scores.MARS uses the difference between complete- and partial-modality representations to teach expert routing. A privileged residual router learns which experts address each missingness pattern; a second router learns to imitate those decisions from available features alone. Thus the residual is a training signal, while deployment uses a practical feature router without the missing modalities.Official ECCV poster record · Official code

Reading question: Which privileged information is available only during training, and exactly what remains in the student?

Evidence boundary: Residual-guided routing is not a residual-error student cascade; conditional optimization arguments do not certify best accuracy.

Spatial-frequency and response-token distillation targets

SFKD selectively aligns wavelet-localized Fourier amplitudes across backbone families. TRAG aligns response-to-vision attention using token-specific entropy-weighted KL. Both teach structure beyond final logits, but use different correspondence assumptions.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
SFKD: Spatial–Frequency Joint-Aware Heterogeneous Knowledge Distillation via Multi-Level Wavelet Spectral InteractionCuipeng Wang, Haipeng Wang2026 · ECCV · acceptedECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. The linked official repository still says source code will be released soon. This is heterogeneous-architecture feature KD, not demonstrated cross-magnification transfer. The selected CIFAR-100 table includes a ViT-S→MobileNetV2 case where FBT is better.SFKD transfers knowledge between different backbone families by separating local spatial structure from global spectral information. It decomposes intermediate maps into Haar-wavelet subbands, refines the student bands through convolutional and Transformer streams, and matches Gaussian-filtered Fourier amplitudes to teacher bands. Reconstruction and logit supervision keep the transfer connected to the classification task.Official ECCV poster record · Official code

Reading question: Are the grids, semantic units and probability supports genuinely aligned before applying the loss?

Evidence boundary: SFKD has architecture-specific regressions; TRAG shares a visual encoder and resolution. Neither establishes cross-magnification performance.

Historical adapter anchors and analytic continual classifiers

Historical LoRA anchors constrain encoder drift while text-guided accumulated statistics update a classifier without raw-example replay. This is a teacher-state and classifier-memory contract, not the residual-error cascade used by other KD methods.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Multi-Anchor Distillation with Text-Guided Analytic Classifier for Continual LearningQier Meng, De Cheng, Jiahao Li, and Cheng Deng2026 · ECCV · acceptedECCV 2026 Spotlight verified via the official related presentation event; the poster identity itself says Accept Poster. No Springer version of record or public runnable implementation was verified. Section 3.5's printed closed-form equation omits a term implied by its objective; the review flags this reproducibility issue without silently repairing the authors' formula.Multi-anchor distillation restrains accumulated CLIP feature drift while a text-guided analytic classifier avoids gradient-based overwriting of classifier weights. Historical LoRA states supply visual anchors; accumulated sufficient statistics and frozen text prototypes supply the classifier. This exemplar-free approach stores model state and statistics rather than old raw images.Official ECCV 2026 Spotlight

Reading question: Which stored states prevent forgetting, and does the analytic argument still apply when the feature encoder changes?

Evidence boundary: Stable-feature equivalence is conditional; check the printed objective/closed-form mismatch and separate average from final accuracy.

Conflicting supervision and adaptive distillation objectives

DTO-KD adapts the balance between task and teacher-transfer objectives using relative-loss optimization. Compare the exact log-gradient construction with the practical amortized update; control the transfer architecture before attributing gains to optimization.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
DTO-KD: Dynamic Trade-off Optimization for Effective Knowledge DistillationZeeshan Hayder, Ali Cheraghian, Lars Petersson, Mehrtash Harandi, Richard Hartley2026 · ICLR · peer-reviewedOfficial ICLR 2026 oral and proceedings manuscript checked. Adaptive task/distillation optimization, not cross-magnification or residual-error KD. Separate the exact two-gradient construction from the amortized implementation; the printed simplex handling and lower-bound argument require qualification, documented in the source-checked review.DTO-KD treats task supervision and teacher imitation as a multi-objective optimization problem. Its central idea is to adapt their contribution according to relative loss improvement instead of choosing one fixed task/distillation weight. The paper distinguishes conflicting gradient directions from unequal gradient magnitudes, but its exact analytical construction and practical amortized algorithm must not be treated as identical.Official proceedings record · OpenReview · Theoretical supplement

Reading question: Which objectives conflict, which update is actually implemented, and what costs and guarantees belong to that implementation?

Evidence boundary: The printed closed form omits simplex boundary handling and the claimed non-collapse lower bound has a counterexample. Natural-image results do not establish cross-magnification or WSI validity.

Which feature directions should distillation preserve?

SpecMatch balances reconstruction along teacher principal components rather than letting high-variance directions dominate an L2 objective. Low variance need not mean low downstream value.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
The Devil is in the Spectrum Bias: Spectrum-Balanced Feature Matching for Robust Representation DistillationKuniaki Saito and Yoshitaka Ushiku2026 · arXiv · preprintarXiv v1 checked 2 October 2026. No official code/model release or conference acceptance verified. Camelyon17 is patch-level evaluation, not whole-slide cross-magnification validation. PE teacher/student labels and parameter counts differ between main text/tables and Appendix C; this review uses unambiguous DINOv2-to-MobileNet Table 1 rows.SpecMatch changes how feature distillation allocates its learning signal. A fixed teacher PCA basis separates representation directions, and a sublinear function of each direction’s reconstruction error reduces the dominance of high-variance coordinates. The aim is to preserve useful low-variance information, not merely make the total feature MSE small.

Reading question: Does a smaller global feature error hide worse transfer on discriminative but low-variance directions?

Evidence boundary: Patch-level Camelyon17 is not a WSI endpoint. Spectral balancing is not evidence that unobserved high-magnification information can be recovered.

Distillation projectors: dimension matching versus optimization interference

A projector can separate the student’s task representation from the space used to imitate a teacher even when both feature dimensions match. PEFD averages parallel nonlinear projector outputs before direction matching; the extended study adds logit, representation-similarity and calibration analyses.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Understanding the Effects of Projectors in Knowledge DistillationYudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, Brano Kusy, and Zi Huang2026 · IEEE TPAMI · peer-reviewedIEEE publisher metadata identifies TPAMI 48(8), 9078–9092, August 2026. The accessible scientific text checked here is arXiv:2310.17183v1 (26 October 2023); final journal numerical identity was not established. This extends the authors’ NeurIPS 2022 Projector Ensemble Feature Distillation work, rather than originating PEFD in 2026.A distillation projector can do more than resize a feature vector: it can buffer the conflict between learning the label task and matching a teacher. This study tests projectors even when dimensions already agree, extends the idea to logits, and analyzes a parallel projector ensemble. The scientific summary follows the accessible 2023 manuscript; the catalog also identifies the 2026 journal publication.Inspected author manuscript (v1) · Author code and earlier PEFD implementation

Reading question: Does the alignment head protect useful student features, or merely fit the teacher target more closely?

Evidence boundary: PEFD originated in 2022. The accessible 2023 manuscript supports the reviewed scientific details; final 2026 journal numerical identity was not verified. More projectors or greater depth do not guarantee improvement.

Complete maintained paper route

Section titled “Complete maintained paper route”

41 canonical readings · 1981–2026· route reviewed through October 2, 2026.

This is the complete structured list for this chapter, ordered chronologically from earlier foundations toward current work. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.

Show all 41 readings in maintained order

  20. #202021IEEE Transactions on Image ProcessingPeer reviewedL3

    Resolution-Aware Knowledge Distillation for Efficient Inference

    Zhanxiang Feng, Jianhuang Lai, and Xiaohua Xie

    Peer-reviewed journal article, volume 30, pages 6985–6996. The paper calls its framework RKD, meaning resolution-aware—not relational—knowledge distillation. Teacher and student use the same architecture but paired high- and low-resolution inputs; only the low-resolution student is needed at inference.

    Official artifacts: PubMed record

    Metadata checked

  21. #212021NeurocomputingPeer reviewedL3

    Residual error based knowledge distillation

    Mengya Gao, Yujun Wang, and Liang Wan

    Published in Neurocomputing 433 (14 April 2021), pages 154–161, DOI 10.1016/j.neucom.2020.10.113. The differently authored arXiv:2002.09168 is recorded only as a related precursor, not as this journal version.

    Metadata checked

  33. #332026Pattern RecognitionPeer reviewedL3

    On learning denoisable student logits

    Diqi Chen, Yang Li, Jiajun Liu, Brano Kusy, Jun Zhou, and Yongsheng Gao

    Peer-reviewed article 113684 in Pattern Recognition 179; available online 10 April 2026 and assigned to the November 2026 issue.

    Metadata checked

  34. #342026ECCVAccepted · proceedings pendingL3

    Multi-Anchor Distillation with Text-Guided Analytic Classifier for Continual Learning

    Qier Meng, De Cheng, Jiahao Li, and Cheng Deng

    ECCV 2026 Spotlight verified via the official related presentation event; the poster identity itself says Accept Poster. No Springer version of record or public runnable implementation was verified. Section 3.5's printed closed-form equation omits a term implied by its objective; the review flags this reproducibility issue without silently repairing the authors' formula.

    Official artifacts: Official ECCV 2026 Spotlight

    Metadata checked

  35. #352026ECCVAccepted · proceedings pendingL3

    Off the Planckian Locus: Using 2D Chromaticity to Improve In-Camera Color

    SaiKiran Tedla, Joshua Little, Hakki C. Karaimer, and Michael S. Brown

    ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #1413. Publisher proceedings/version of record was not verified on 2026-09-12.

    Official artifacts: ECCV Spotlight program · Conference manuscript · Author project (paper/program link)

    Metadata checked

  36. #362026ECCVAccepted · proceedings pendingL3

    Residual-Guided Expert Specialization for Incomplete Multimodal Learning

    Seunghun Baek, Jihwan Park, Jaeyoon Sim, Minjae Jeong, Hoseok Lee, Won Hwa Kim

    ECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. MARS distills a privileged residual router into an available-feature router; it is not a cascade of residual-error student networks. Full modalities are assumed available during training. Food-101 full-modality accuracy is below SimMLM despite higher missing-modality and average scores.

    Official artifacts: Official ECCV poster record · Official code

    Metadata checked

  37. #372026ECCVAccepted · proceedings pendingL3

    MoE-KD: Your Teacher Model is Worth Mixture-of-Experts for Knowledge Distillation

    Kuo Shi, Wenjie Zhu, Bo Peng

    ECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. The EM convergence discussion assumes an improving M-step; it is not a guarantee of global optimum for stochastic neural training. Table 3 values take precedence over inconsistent prose deltas. No runnable official implementation was verified in this pass.

    Official artifacts: Official ECCV poster record

    Metadata checked

  38. #382026ECCVAccepted · proceedings pendingL3

    SFKD: Spatial–Frequency Joint-Aware Heterogeneous Knowledge Distillation via Multi-Level Wavelet Spectral Interaction

    Cuipeng Wang, Haipeng Wang

    ECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. The linked official repository still says source code will be released soon. This is heterogeneous-architecture feature KD, not demonstrated cross-magnification transfer. The selected CIFAR-100 table includes a ViT-S→MobileNetV2 case where FBT is better.

    Official artifacts: Official ECCV poster record · Official code

    Metadata checked

  39. #392026ICLRPeer reviewedL3

    DTO-KD: Dynamic Trade-off Optimization for Effective Knowledge Distillation

    Zeeshan Hayder, Ali Cheraghian, Lars Petersson, Mehrtash Harandi, Richard Hartley

    Official ICLR 2026 oral and proceedings manuscript checked. Adaptive task/distillation optimization, not cross-magnification or residual-error KD. Separate the exact two-gradient construction from the amortized implementation; the printed simplex handling and lower-bound argument require qualification, documented in the source-checked review.

    Official artifacts: Official proceedings record · OpenReview · Theoretical supplement

    Metadata checked

  40. #402026arXivPreprintL3

    The Devil is in the Spectrum Bias: Spectrum-Balanced Feature Matching for Robust Representation Distillation

    Kuniaki Saito and Yoshitaka Ushiku

    arXiv v1 checked 2 October 2026. No official code/model release or conference acceptance verified. Camelyon17 is patch-level evaluation, not whole-slide cross-magnification validation. PE teacher/student labels and parameter counts differ between main text/tables and Appendix C; this review uses unambiguous DINOv2-to-MobileNet Table 1 rows.

    Metadata checked

  41. #412026IEEE TPAMIPeer reviewedL3

    Understanding the Effects of Projectors in Knowledge Distillation

    Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, Brano Kusy, and Zi Huang

    IEEE publisher metadata identifies TPAMI 48(8), 9078–9092, August 2026. The accessible scientific text checked here is arXiv:2310.17183v1 (26 October 2023); final journal numerical identity was not established. This extends the authors’ NeurIPS 2022 Projector Ensemble Feature Distillation work, rather than originating PEFD in 2026.

    Official artifacts: Inspected author manuscript (v1) · Author code and earlier PEFD implementation

    Metadata checked