BayesianDeepLearning-Survey: A Maintained Reading List, Not a Library

What this repository actually is: a README that tracks a field

The README opens by describing the project as an updating survey for Bayesian Deep Learning, an extended and continuously updated version of the manuscript A Survey on Bayesian Deep Learning, published in ACM Computing Surveys in 2020. That sentence sets the contract. The repository is a bibliography with commentary by structure, not a software package. There is no source tree described in the material, no build file, no package manifest, and no releases. The primary language is listed as unknown, which is consistent with a repository whose main artifact is a Markdown file plus one image, BDL_Table.png, referenced inline in the Survey section.

The table of contents lists sixteen sections. They run from the base Survey through application areas: recommender systems, domain adaptation and meta learning, healthcare, NLP, computer vision, control and planning, graphs, topic modeling, speech, forecasting, distributed and federated learning, continual learning, AI4Science, a miscellaneous framework section, and a closing section on Bayesian or probabilistic neural networks as building blocks. If you have ever tried to assemble that map yourself from conference proceedings, you know the work is in the grouping, not in the individual links.

The intended reader is someone who needs orientation before committing to a method. A practitioner deciding whether to put a variational treatment into a recommender model, or a student looking for the canonical references behind Bayesian deep learning, can use the section headings as a syllabus. The repository does not claim to be more than that, and the README does not present code of its own.

The section taxonomy is the product, and it is opinionated

Most paper lists are chronological or alphabetical. This one is organized by application domain, and that choice carries information. Placing Collaborative Deep Learning for Recommender Systems in a recommender section rather than in a general Bayesian methods section tells you the maintainers think of BDL as something applied, and it lets a reader jump straight to their own domain.

The entries follow a consistent citation pattern: title, authors and venue, then a bracketed list of links. A typical entry reads: Collaborative Deep Learning for Recommender Systems, by Wang et al., KDD 2015, followed by PDF, Project Page, a 2014 arXiv version, and separate code links for CDL, MXNet, and TensorFlow, plus dataset links for citeulike-a and citeulike-t, a Jupyter notebook, and two slide decks. That is a lot of surface area for one paper, and it is the pattern the maintainers repeat where it exists. Not every entry is that rich. Collaborative Recurrent Autoencoder, from NIPS 2016, has a single PDF link. The unevenness is honest: it reflects what the maintainers had on hand.

The practical effect is that you can filter by artifact type without reading every line. If you need a dataset to reproduce something, the recommender section is the only place in the visible README where dataset links appear. If you need slides for a reading group, the same section has them. If you want the foundational survey itself, the Survey section links the CSUR 2020 PDF, a blog post at wanghao.in/BDL.html, and the 2016 TKDE framework paper.

One caveat on the taxonomy: domain adaptation, domain generalization, and meta learning are folded into a single section, and the heading itself says so with an etc. That grouping mixes problems that are related but not identical, and a reader looking specifically for domain generalization will have to scan past meta learning entries to find Learning to Learn with Variational Information Bottleneck for Domain Generalization. The heading is doing organizational work, but it is not a precise index.

How you use it: there is nothing to install

There is no installation step, and no command to run. The README is the interface. Getting the material means cloning the repository or opening the README on GitHub and following the table of contents anchors, which point at the same file with fragment identifiers such as #bdl-and-recommender-systems. If you want the image that summarizes the framework, it is referenced as ./BDL_Table.png in the Survey section, so a clone gives you the Markdown and the PNG together.

A clone also gives you a snapshot you can search locally. Because the entries are plain Markdown with bracketed links, a grep over the README will find every entry mentioning a venue or an author. That is the only query interface the material supports. There is no search index, no tags, and no metadata file.

The one substantive configuration decision is whether to treat the repository as a pinned reference or as a feed. The README describes itself as constantly updated, and the repository metadata shows a last push of 2026-07-04. Those two facts together mean the file changes over time. If you cite the survey in your own work, cite the underlying ACM Computing Surveys paper and the specific version you read, because the README has no releases and no tags to point at. There is no version number to pin.

If you want to contribute, the only mechanism the material implies is editing the README through a pull request on the master branch, which is the default branch. No contribution guidelines appear in the supplied material.

The lack of a license is the sharpest limitation

The license is listed as unknown. Nothing in the supplied README states licensing terms for the repository contents. That matters more here than it would for a code project, because the repository's value is its selection and arrangement of links, and the links point outward to papers with their own terms. The PDFs are hosted at author pages, ACM, arXiv, JMLR, Springer, NeurIPS proceedings and similar venues. Each of those has its own reuse policy, and none of them is the repository's policy.

For an individual reader this is mostly a non-issue. You follow links and read papers under whatever terms those venues set. For anyone who wants to mirror the list, republish the taxonomy, or bundle the README into internal documentation, the absence of a license is a real blocker: without an explicit grant you do not have stated permission to redistribute the compilation. This is not legal advice, and it is not a claim that redistribution is forbidden. It is a statement that the material does not tell you, and that a reader who needs to know should ask the maintainers rather than assume.

The same uncertainty applies to the BDL_Table.png image. It is included in the repository, and no separate attribution or license note accompanies it in the supplied material. If you plan to reuse that figure, the repository does not answer the question for you.

Where it breaks down: freshness, coverage, and the missing code

An updating survey has a maintenance problem that a frozen one does not. The README is the only maintained surface, and there is no changelog, no release, and no dated entry format. Entries do not carry the date they were added. That means a reader cannot tell from the file alone whether a section was refreshed last month or three years ago. The last push timestamp is repository-level, so a single typo fix elsewhere would move it without any content change in the section you care about. If you are using this to judge whether a subfield is current, the README will not tell you.

Coverage is also visibly uneven. The recommender section is dense, with entries spanning 2015 to 2020 and multiple code and dataset links. The healthcare section, in the visible portion, shows a JMLR 2016 paper and then the README is truncated mid-entry at Structured Inference Networks. The supplied material does not show the full contents of that section or of the later ones, so any claim about how complete NLP, speech, or AI4Science coverage is would be a guess. What can be said is that the entries that do appear are anchored to specific venues and years, which makes them checkable.

The third limitation is the one that surprises people who arrive expecting a framework. The repository name and the README's talk of a BDL framework could suggest runnable software. The 2016 TKDE paper is linked as a framework paper, but the repository itself does not ship that framework. Code links exist for individual papers, such as the CDL implementations in MXNet and TensorFlow, and they are separate repositories. If your goal is to train a Bayesian model this week, this repository is a detour, not a starting point.

What to use instead when you need working code

The honest alternative depends on what you actually wanted. If you wanted a maintained software library for Bayesian deep learning, this repository is the wrong tool, and a probabilistic programming framework is the right category. Pyro and NumPyro, built on PyTorch and JAX respectively, give you variational inference and MCMC primitives you can apply to a neural network. TensorFlow Probability plays the same role in the TensorFlow ecosystem. The difference in approach is fundamental: those projects ship inference machinery and documentation for running it, while this repository ships a curated set of paper links and expects you to implement or find the method elsewhere.

If what you wanted was a survey to read rather than a list to browse, the ACM Computing Surveys paper linked in the Survey section is the more appropriate artifact. It is a peer-reviewed manuscript with a fixed structure and a citable version of record, whereas the README is a living document with no version number. The README is the update channel for that paper, which is the relationship the first line describes.

If you wanted a general paper index rather than a Bayesian-specific one, the difference is scope. A broad index covers everything and organizes nothing for you. This repository covers one subfield and organizes it by application domain, which is why the recommender and domain adaptation sections read as coherent reading paths rather than search results. That focus is the reason to prefer it over a general index, and also the reason it will be thin wherever the maintainers have not yet collected entries.

Maintenance cost and who should adopt it

The maintenance cost falls almost entirely on the maintainers, not on you. There is no dependency to upgrade, no API to track, and no breaking change to absorb, because there is no code. Your cost is attention: you have to re-read sections you care about to notice new entries, since the repository gives you no notification mechanism beyond watching it on GitHub. The maintainers' cost is the real one, and it is the cost of any curated list. Every section needs someone to keep scanning venues, and the sections with the fewest entries are the ones most likely to fall behind.

For an engineering team, the sensible use is as an onboarding artifact. Point a new hire at the recommender section and they get the canonical papers plus, in several cases, the code and datasets to reproduce them. That is a concrete, bounded task. Pointing them at the whole README is less useful, because sixteen sections with no priority ordering is a lot of surface for someone who does not yet know which subfield matters to the project.

What to verify before you depend on it: check the license situation with the maintainers if you intend to redistribute anything, confirm that the section you need has entries from the last year or two rather than treating the repository-level push date as evidence, and follow the individual code links to see whether those separate repositories are themselves maintained. The README will not tell you any of that, and the last push date on this repository is the only freshness signal it exposes.

Editorial conclusion

Adopt this as a reading list if you are starting a Bayesian deep learning literature review and want a pre-grouped set of paper links across recommender systems, healthcare, NLP, computer vision, graphs, forecasting, federated learning and AI4Science. Do not adopt it if you need runnable code, versioned dependencies, or a clearly licensed artifact to redistribute: the repository has no detected license and no releases. Before relying on it, check whether the section you care about has been updated recently, since the README is the only maintained surface and its last push date is the only freshness signal you have.

Official sources

1. Issues
2. js05212/BayesianDeepLearning-Survey on GitHub
3. README