About
I categorize, annotate and write comments for all research papers I read (410+ papers since 2018).
In June 2023, I wrote the blog post The How and Why of Reading 300 Papers in 5 Years (why I think it’s important to read a lot of papers + how I organize my reading + paper statistics + a list of 30 particularly interesting papers).
Categories:
[Uncertainty Estimation], [Ensembling], [Stochastic Gradient MCMC], [Variational Inference], [Out-of-Distribution Detection], [Theoretical Properties of Deep Learning], [VAEs], [Normalizing Flows], [ML for Medicine/Healthcare], [Object Detection], [3D Object Detection], [3D Multi-Object Tracking], [3D Human Pose Estimation], [Visual Tracking], [Sequence Modeling], [Reinforcement Learning], [Energy-Based Models], [Neural Processes], [Neural ODEs], [Transformers], [Implicit Neural Representations], [Distribution Shifts], [Social Consequences of ML], [Diffusion Models], [Graph Neural Networks], [Selective Prediction], [NLP], [Representation Learning], [Vision-Language Models], [Image Restoration], [Computational Pathology], [Survival Analysis], [Miscellaneous].
Papers:
- Papers Read in 2024
- Papers Read in 2023
- Papers Read in 2022
- Papers Read in 2021
- Papers Read in 2020
- Papers Read in 2019
- Papers Read in 2018
Papers Read in 2024:
[24-06-23] [paper411]
- Multimodal Prototyping for Cancer Survival Prediction [pdf] [annotated pdf]
ICML 2024
- [Computational Pathology], [Survival Analysis]
Well-written and quite interesting paper. Basically, they apply the prototype-based slide representation from "Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology" to the survival analysis model in "Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction" (SurvPath), two CVPR 2024 papers from the same group. The main thing is that the compact prototype-based slide representation now allows them to use standard attention without any approximations, and also to train the survival model using the Cox partial log-likelihood loss - instead of the discrete NLL with batch size = 1 used in SurvPath. In fact, if they use the NLL loss with batch size = 1, they even get slightly worse performance than SurvPath (0.621 vs 0.629)? I.e., this seems to be the main/only thing that improves over SurvPath? Still a very solid paper though, and it makes a lot of sense to combine their two previous papers.
[24-06-22] [paper410]
- Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction [pdf] [annotated pdf]
CVPR 2024
- [Computational Pathology], [Survival Analysis]
Well-written and quite interesting paper, I enjoyed reading it. The method makes sense overall, they describe it well. The results actually seem quite impressive, relevant baselines and they do get a relatively clear bump in performance. Difficult for me to judge how actionable the interpretability results in Section 4.5 actually are though.
[24-03-03] [paper409]
- Diffusion Models for Out-of-Distribution Detection in Digital Pathology [pdf] [annotated pdf]
Medical Image Analysis, 2024
- [Computational Pathology], [Out-of-Distribution Detection], [Diffusion Models]
Interesting paper overall, but I got a bit lost in all the details. Not my favourite type of paper perhaps (I might also have been a bit too tired when reading). The overall idea of using diffusion models for reconstruction-based OOD detection is definitely interesting though.
[24-02-18] [paper408]
- Artificial Intelligence to Identify Genetic Alterations in Conventional Histopathology [pdf] [annotated pdf]
Journal of Pathology, 2022
- [Computational Pathology]
Well-written and interesting paper, I enjoyed reading it. Gives a very good background on various biomarkers. I definitely didn't understand all the medical details, but still found it interesting and useful.
[24-02-11] [paper407]
- Transcriptome-Wide Prediction of Prostate Cancer Gene Expression From Histopathology Images Using Co-Expression-Based Convolutional Neural Networks [pdf] [annotated pdf]
Bioinformatics, 2022
- [Computational Pathology]
Well written and quite interesting paper. Helped me understand the general problem a bit better, good background material for me. I definitely didn't understand all the medical stuff, but still found e.g. everything in the Discussion interesting and useful.
[24-02-04] [paper406]
- Assessing and Enhancing Robustness of Deep Learning Models with Corruption Emulation in Digital Pathology [pdf] [annotated pdf]
arxiv, 2023-10
- [Computational Pathology]
Interesting and quite well-written paper. Quite short, and basically no implementation details are given for the various corruptions. It does seem potentially useful though, for both benchmarking and augmentation. The results in Table 3 seem quite impressive.
[24-02-03] [paper405]
- Uncertainty Sets for Image Classifiers using Conformal Prediction [pdf] [annotated pdf]
ICLR 2021
- [Uncertainty Estimation]
Quite interesting and well-written paper. The proposed method mostly makes sense, and it does indeed seem to produce smaller prediction sets than the baselines. Not sure how relevant this is for me though.
[24-02-02] [paper404]
- Estimating Diagnostic Uncertainty in Artificial Intelligence Assisted Pathology Using Conformal Prediction [pdf] [annotated pdf]
Nature Communications, 2022
- [Computational Pathology], [Uncertainty Estimation]
Interesting paper, but I found it quite difficult to understand. I was confused by the employed conformal prediction method, it did not make sense to me that it could output empty predictions for some inputs. "efficiency, defined as the fraction of all predictions resulting in a correct single-label prediction" is not something I've seen before either. The dataset setup is neat though, with test sets from the same scanner/lab, from a different scanner, and from a different scanner and lab. Figure 2 is interesting, shows that the model becomes significantly overconfident on test set 3-5.
[24-01-21] [paper403]
- Improving Trustworthiness of AI Disease Severity Rating in Medical Imaging with Ordinal Conformal Prediction Sets [pdf] [annotated pdf]
MICCAI 2022
- [Uncertainty Estimation], [ML for Medicine/Healthcare]
Quite well-written and interesting paper, not overly impressed. I don't fully understand how the method is implemented in practice, Algorithm 1 makes sense, but how does one compute the value of lambda?
[24-01-20] [paper402]
- Hierarchical Vision Transformers for Context-Aware Prostate Cancer Grading in Whole Slide Images [pdf] [annotated pdf]
NeurIPS Workshops 2023
- [Computational Pathology], [Transformers]
Workshop paper, just ~3 pages. The appendix also contains some useful info. Well-written and interesting paper. The method/model makes intuitive sense. Seemingly strong performance without any ad hoc modifications. Qiute interesting that the local model version (fine-tuning also the region-level transformer) gave such a clear performance boost.
[24-01-19] [paper401]
- Conformal Prediction Sets for Ordinal Classification [pdf] [annotated pdf]
NeurIPS 2023
- [Uncertainty Estimation]
Quite well-written paper, interesting overall. The propsed approach is actually very simple in the end, just modify the DNN output before the softmax layer according to eq. (5), train it using standard cross-entropy, and then apply the existing conformal prediction method APS (or LAC) at test-time to output prediction sets (if I understood everything correclty). The results seem reasonable. Tables and figures could be made to look a bit better.
[24-01-19] [paper400]
- Artificial Intelligence for Diagnosis and Gleason Grading of Prostate Cancer: The PANDA Challenge [pdf] [annotated pdf]
Nature Medicine, 2022
- [Computational Pathology]
Well-written and interesting paper. Neat/cool/impressive study/challenge design, it makes sense to evaluate the top-performing mehtods on external data afterwards.
[24-06-16] [paper399]
- Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology [pdf] [annotated pdf]
CVPR 2024
- [Computational Pathology]
Interesting and very well-written paper, I enjoyed reading it. Figure 2 gives a great overview of their approach. The visualizations in Fig 1, 3, S1 - S4 are really neat.
[24-06-12] [paper398]
- Prediction of Recurrence Risk in Endometrial Cancer with Multimodal Deep Learning [pdf] [annotated pdf]
Nature Medicine, 2024
- [Computational Pathology]
Well-written and somewhat interesting paper. Not quite my type of paper, would probably need a bit stronger medical background. The method sort of seems unnecessarily complicated to me, using a second frozen model, embedding layers etc. Yes, they see some gains in ablations, but still feels like doing something simpler also could work well. The experiment on adjuvant chemotherapy response prediction is interesting.
[24-06-11] [paper397]
- A Whole-Slide Foundation Model for Digital Pathology from Real-World Data [pdf] [annotated pdf]
Nature, 2024
- [Computational Pathology]
Well-written and interesting paper, I quite enjoyed reading it. ~1.3 billion 256x256 patches, from ~170k WSIs, from ~30k patients. 45% of the slides are from lung tissue, 30% from bowel, 9% from CNS/brain, 3% from breast (Suppl. Fig 1). I didn't look too too carefully at the results, I was mostly just interested in the data and method, but seems reasonable. The vision-language experiments are interesting, the fact that they to do this contrastive alignment on the slide level, they use ~17k WSI - pathology report pairs.
[24-06-09] [paper396]
- Multistep Distillation of Diffusion Models via Moment Matching [pdf] [annotated pdf]
arxiv, 2024-06
- [Diffusion Models]
Fairly interesting and well-written paper. I should probably have read a more basic paper about diffusion model distillation instead, don't think I was able to fully appreciate the details here. Struggled to properly follow everything in Section 3.2. The proposed method seems somewhat ad hoc, but also seems to work well in practice. Quite interesting to see an example of a SOTA distillation method at least.
[24-06-08] [paper395]
- Flow Matching for Generative Modeling [pdf] [annotated pdf]
ICLR 2023
- [Diffusion Models]
Well-written and interesting paper, I've been meaning to read this for quite some time now. I struggled a bit to follow Section 3 and 4, but overall the method makes sense I think. Would need to read more, and probably also discuss this with someone, to properly understand and appreciate all the details.
[24-05-29] [paper394]
- No "Zero-Shot" Without Exponential Data: Pretraining Concept Frequency Determines Multimodal Model Performance [pdf] [[annotated