Map of deep learning and notes from papers.
Switch branches/tags
Nothing to show
Clone or download


Map of deep learning and notes from papers.

Math is rendered in [KaTeX-friendly version on GitLab]. ([GitHub version])

Note (13 June 2018): I am revamping the lists to structure them around more meaningful questions. The lists (previous version) can be found in the directory previous-lists.

README contents:

  1. Contents of this Repo
  2. Vision for the Deep Learning Map
  3. Summaries
  4. Topics to cover (in the DL map)
  5. Recommended Deep Learning Resources
  6. Existing repositories with summaries of papers on machine learning
  7. Newsletters
  8. Other Resources
  9. Development notes

0. Contents of this Repo

  1. Paper summaries: summaries/
    • See 'Summaries' section in this README for details.
  2. Glossaries
      • all works in progress
      • version with LaTeX equations rendered: basics-glossary.ipynb
  3. Lists
    • NOTE: Many of these comprise only items I've come across in my reading since Dec 2017, so these lists don't represent my view of e.g. 'the most important datasets'. Though I try to include only items I think are significant.
    • As of 13 June 2018: Many of these may be in the previous-lists directory.
    • problem-areas
      • '`
    • tools
      • environments for training DL algorithms.
    • Some papers I'm interested in.
    • List of components of neural networks.
      • A first step towards drawing a 'map' of Deep Learning.
    • Misc:
  4. Implementations
    • implementations/neural-networks: implementations of deep learning algorithms (early stages, currently have 2D MLP working)

1. Vision for the Deep Learning Map

The idea is to write (or link to) paper summaries, blog posts or articles that will help people (especially those starting out):

  • Understand what different models or terms mean
    • Know what the state-of-the-art results are in each domain
    • Be able to look up known advantages and disadvantages of key models and approaches
  • See how different concepts connect with each other

It is thus crucial that

  • these summaries are presented in a way that makes the relationships between different concepts clear (hence this being a 'map'), and that
  • materials are chosen selectively so as not to overwhelm the reader.

This is still in early stages - at the moment this is more a collection of paper summaries and definitions of terms than a map. It also contains material outside of deep learning, mostly in machine learning or neuroscience.

Let me know if you'd like to contribute or have suggestions.

2. Summaries

3. Topics to cover

  • DQN

    • Deep convolutional Q-learning
  • A3C (Asynchoronous Advantage Actor-Critic)

    • A2C
  • Policy gradient methods

    • Trust Region Policy Optimisation (TRPO)
    • Proximal Policy Optimisation (PPO)
  • Hierarchical networks

    • Feudal networks
  • Auxiliary tasks

    • UNReAL
  • Dilated convolutions

  • Dilated LSTMs

  • Quasi-recurrent NNs

  • Hierarchical RNNs

  • Capsule Networks

  • AI Safety

4. Recommended Deep Learning Resources

See also my effective deep learning resources shortlist.

5. Existing repositories with summaries of papers on machine learning

Have added a ⭐️ to the ones I find particularly helpful.

  • ⭐️ Alexander Jung
    • Summaries:
      • What, How, Results summary. Easy to digest.
      • Bullet points with images (result graphs, architectures).
      • Links to related resources (paper website, video).
    • Lists by date added, with brief tags such as 'self-driving cars', 'segmentation','gan'.
    • Around 80 papers added from 2016-17, last active Dec 2017.
    • Starred because summaries are easy to digest and understand.
  • ⭐️ Denny Britz
    • Summaries:
      • Short TLDR; section that gives authors' argument, method and high-level findings.
      • Sometimes contain key points (usually model, experiment or result details) and thoughts sections.
    • Lists summaries by date reviewed, list also includes many other papers without summaries (are these key papers or ones he's read?).
    • About 100 summaries. Last active Nov 2017.
    • Starred because summaries are short but to the point, and because there are many summaries.
    • I also recommend his weekly newsletter 'The Wild Week in AI' and his blog WildML which has some great tutorials and a deep learning glossary.
      • In some way the glossary has done part of what I'd like to do with this map. Haha great!
      • Oh and did I mention his repo of implementations of RL algorithms? :O I haven't even gone through this properly yet.
  • Dibyatanoy Bhattacharjee (Yale)
    • Summaries
      • Longer summaries that outline paper content and someotimes include definitions.
      • Paragraphs or bullet points with images (architecture diagrams)
      • Points to Ponder sections (sometimes)
    • Lists all summaries (titles AND summary content) on one webpage. Interesting because you can scroll through all of them in one go and see what catches your attention.
    • 7 summaries. Last active May 2017.
    • Mostly papers on RNNs / memory / translation.
  • Abhishek Das (CS PhD student at Georgia Tech)
    • Summaries:
      • Provides brief summary of model in paper. Does not seem to include results.
      • Gives opinion on the strengths and weaknesses of the paper.
      • Seems to be text only.
    • Lists paper summaries by paper year.
    • Around 40 papers from 2012-2017, Last active August 2017
  • yunjey
    • Summaries:
      • Short, high-level summaries
      • Short list of contributions of each paper.
      • Sometimes includes opinions on similarities of papers to existing models.
    • Lists by topic (yes!) with author, publication month/year and conference (e.g. NIPS, ICLR) if applicable.
      • List is unfortunately not linked to summaries.
    • 6 summaries but many more papers listed. Last active c. Dec 2016.
  • Patrick Emami (CS PhD student at U of Florida)
    • Summaries:
      • Paragraph-based summary of paper: what authors propose, model outline, experiment outline.
      • 'My Notes' section with opinions on main contributions, weaknesses/questions and areas for further research.
        • Nice.
    • Lists papers by topic.
    • About 40 summaries. Last active August 2017.

6. Newsletters

Including these here because they contain fantastic summaries of what's going on in industry and in research. They take some time to go through though.

  • ImportAI by Jack Clark
    • Weekly email newsletter
    • Noteworthy items with headlines and more detailed descriptions.
      • May take more time to go through but the summaries are of high quality and are often structured (e.g. briefly describing testing methods, results).
    • Hilarious tech fiction at the end of the newsletter. :)
  • The Wild Week in AI by Denny Britz
    • Weekly email newsletter
    • Sections:
      • News,
      • Posts, Articles, Tutorials
      • Code, Projects & Data
    • Links with brief summaries (and occasionally helpful context or opinions).

7. Other resources

8. Development Notes

9 Dec 2017

  • Add brainstormed list of topics to cover. The idea is to write (or link to) paper summaries, good blog posts or articles that will help people with limited experience get a better idea of what is going on in the field. This means:
    • Understanding what different models or terms mean
      • Knowing what the state-of-the-art results are in each domain
      • Being able to look up known advantages and disadvantages of key models and approaches
    • Seeing how different concepts connect with each other
  • The target audience is not currently experienced researchers, but the hope is that researchers will eventually benefit from this as well.
  • I will also be going through Goodfellow et. al's book 'Deep Learning' and may add insights or summaries from the book (referencing those appropriately.)
  • Difficulties: it is hard to know how to connect concepts with each other initially, so I will first
    • (1) write paper summaries,
    • (2) write a list of summaries of key terminology, and
    • (2) build a spreadsheet trying to list connections in parallel.
      • The spreadsheet is important: I believe that it will be greatly beneficial to have a visual map and not just a list of papers because the latter is much harder to digest.
  • Everything will likely be scattered at first, but I hope the pieces will start coming together after the first month.