annotated_research_papers: A PDF Reading Log for Computer Vision and Self-Supervised Learning

The Problem It Solves: A Second Pass Over Hard Papers

Reading a dense machine learning paper alone is slow. The README states the motivation directly: the author spends a lot of time reading papers as part of ML work, and the repository exists to share the thought process behind that reading. The stated audience is anyone who finds research papers intimidating or who wants annotated versions that are easier to understand. The note about being a pen-and-paper reader is the key detail. The annotations are a substitute for the printed page, not a substitute for the paper. The author also warns that the order of papers will not strictly follow arXiv timelines, and that some papers are put on hold and read later. That matters for anyone expecting a feed. This is a personal reading log published in public, and the selection reflects one person's interests in computer vision and self-supervised learning rather than a curriculum. If your goal is to understand a specific architecture from the inside, a marked-up PDF from someone who has already worked through it can save hours. If your goal is a survey of a field, this is the wrong shape of resource.

What Is Actually in the Repository

The table of contents is organised by field and category. Computer Vision is split into supervised work and self-supervised work, with a segmentation directory appearing in the file paths. Under supervised, the listed papers include Adaptive Risk Minimization, Axial DeepLab, ConvNeXt, EfficientNetV2, Flow-edge Guided Video Completion, Is Batch Norm Unique?, a knowledge distillation paper titled A good teacher is patient and consistent, RandConv, Polyloss, Scaling Down Deep Learning, Segment Anything, Supervised Contrastive Learning, Vision Transformer, and Gaze-LLE. Under self-supervised, the table lists Are all negatives created equal in contrastive instance discrimination?, Towards Domain-Agnostic Contrastive Learning, Emerging Properties in Self-Supervised Vision Transformers (DINO), Decoder Denoising Pretraining, Masked Autoencoders, SwAV, What Should Not Be Contrastive in Contrastive Learning, Vision Transformers need Registers, and NEPA. Each row links to a PDF inside the repository, and most rows also link to an official code repository and an abstract page on arXiv or OpenReview. The file paths are flat and descriptive: ./supervised/convnexts.pdf, ./self-supervised-learning/masked_autoencoder.pdf, ./segmentation/axial_deeplab.pdf. The README table contains several empty rows, which is consistent with the author's note that not every paper read will be annotated.

How the Annotations Are Delivered

There is no application, no build step, and no annotation format. The mechanism is PDF files committed to a git repository, with a Markdown table acting as the index. That is the entire architecture. The consequence is that annotations are not machine-readable: you cannot query them, diff them, or extract them into a training set. You open a PDF in a viewer and read the marks alongside the text. Because the repository has no homepage and no releases, there is no rendered version to browse either. The README embeds a GIF at static/papers.gif, which is the only visual preview of what the annotations look like before you clone anything. If you want to judge the annotation style, that GIF and a single PDF download are the fastest path. The trade-off is deliberate: PDFs preserve the original layout, figures, and equations exactly as published, which a re-typeset HTML version would not. The cost is that everything is opaque to tooling.

Getting the PDFs: Clone or Fetch a Single File

The README does not document installation, because there is nothing to install. The practical commands are the standard git ones. To take the whole collection: git clone https://github.com/AakashKumarNain/annotated_research_papers.git. To avoid cloning every PDF, fetch a single file from the default branch, which the repository metadata identifies as master. For example, the ConvNeXt annotations sit at ./supervised/convnexts.pdf on that branch, so a raw fetch would target that path under the master branch. There are no configuration keys, no environment variables, and no dependencies, since the primary language field is unknown and the content is documents. The only configuration decision is your PDF viewer: if the annotations are handwritten or drawn, you may need a viewer that renders layered annotations correctly rather than flattening them. The README does not state which tool was used to annotate, so that is something to check on the first file you open.

The Coverage Gap Is the Main Limitation

The README is unusually honest about scope. It says the author cannot annotate all the papers read, and that only interesting papers can be expected to appear. Combined with the non-chronological ordering, this means the repository has no completeness guarantee in any direction: not by date, not by venue, not by subfield. The table lists two fields, computer vision and self-supervised learning, and the file paths add segmentation as a directory rather than a top-level category. If you work on NLP, reinforcement learning, or systems, there is nothing here for you. A second limitation is freshness relative to the papers themselves. The table includes recent work such as Gaze-LLE and NEPA alongside older entries, but the annotations describe the version the author read. If a paper has been revised on arXiv since, the annotations may refer to a figure or claim that has moved. A third issue is that PDFs in a git repository are binary blobs. Every added or updated annotation increases clone size permanently, and there is no release artifact to pin, since no releases were retrieved. There is also no stated review process, so accuracy rests on one reader. Treat the annotations as a study aid to compare against your own reading, not as an authoritative commentary.

How It Differs from Papers with Code and alphaXiv

Papers with Code indexes papers and attaches implementations, datasets, and leaderboards, with structured metadata that supports search and filtering across fields. This repository does none of that. It is a hand-curated list where the annotation, not the link, is the product; the code and abstract links are conveniences attached to rows. If you need to find every paper on a topic, Papers with Code is built for that and this repository is not. alphaXiv takes a different route: it hosts papers with an overlay for public discussion, so commentary accumulates from many readers in a web interface. Here the commentary is one person's, stored in the PDF, and readable offline. The distinction that matters is persistence and portability. A cloned PDF works on a plane, in a locked-down environment, or ten years from now, provided you keep the file. A discussion platform depends on the platform. The cost of that portability is that you cannot see how many people found a particular annotation useful, and you cannot add your own marks without editing the PDF or forking the repository.

Maintenance, Licensing, and What to Verify

The repository is MIT licensed and not archived, and the last push recorded in the metadata is 2026-08-29. That licence covers the repository contents, but the annotated PDFs are derivative works of papers owned by their publishers or authors. The MIT grant cannot extend rights the author does not hold, so redistribution or reuse of the underlying papers is a separate question from the licence file. That is a factual boundary, not legal advice; if you plan to redistribute the PDFs, check the terms attached to each paper. On maintenance: there are no releases, no CI, and no dependency graph, so the ongoing cost is limited to pulling new commits. The upgrade risk is the opposite of a software project. Nothing breaks, but nothing is versioned either, so an updated annotation silently replaces the old one in your working copy. If you cite a specific annotation, record the commit hash. Before adopting the collection for a study group, verify three things: that the annotations render in your viewer, that the linked code repository for your target paper still corresponds to the version annotated, and that the paper you need is actually present, since the table's empty rows and the author's own caveat mean absence is expected rather than an oversight.

Editorial conclusion

Adopt this if you are working through computer vision or self-supervised learning and want a second reader's marks on a specific paper already in the table, such as ConvNeXt, MAE, or DINO. Do not adopt it if you need coverage guarantees, structured metadata, or papers outside those two subfields; the table is explicitly incomplete and the order is not chronological. Before relying on it, open the target PDF and confirm the annotations are legible and that the linked code repository still matches the paper version you are reading.

Official sources

1. AakashKumarNain/annotated_research_papers on GitHub
2. Issues
3. License: MIT
4. README