Skip to content

Plans.md

Latest commit

 

History

History
171 lines (119 loc) · 6.63 KB

Plans.md

File metadata and controls

171 lines (119 loc) · 6.63 KB

AI group plans

Also see Calendar.

Objectives

  • learn Julia

  • implement machine learning algorithms in it so as to learn how modern AI works

  • see how to use AI on selected problems: see Project Ideas

Julia (while a general purpose programming language) is really "made" for scientific computing and should thus be a very good match for us to implement AI algorithms in (it's likely always possible to get an edge on performance by directly coding in CUDA/OpenCL/C/Rust/... but investing effort there is probably not sensible as we probably won't beat existing libraries anyway, and Julia is going to be much easier to handle and already very fast (and when accessing existing libraries in C etc. then it doesn't matter anyway).)

We didn't know what sources to follow for learning about AI, and so I had just fired up a search for online courses and found the Stanford one, and given that one of the two teachers is the rather well-known Peter Norvig, I figured that will be a good choice. I think the videos are somewhat slow-going, and at first they didn't appear very exciting to me, but my attitude towards them has changed a bit since now I think they are valuable to get an overview on what the AI (research) community is doing, and to get a hold on their vocabulary. Also, those algorithms like the path search do seem like nice examples to base learning Julia on.

Status

  • "Intro to AI": watched videos 1-34 on Youtube (see Teaching). Have implemented the first two path search algorithms presented in the videos in Scheme, as well as the first in Julia (also others in progress)

  • Borg:

    • still runs Ubuntu as the host (now with SSH on port 1022), but has a chroot with Debian testing on the default SSH port. TODO: upgrade Ubuntu release (security upgrades have finished?), LDAP in chroot (talk to Jasper).

    • user chris has successfully installed libraries for Jupyter. TODO: move to other user account, access etc.

  • GPU: we've got a computer with a relatively good (if somewhat aged) GPU now.

Historic notes

  • May 8: decision to have at most one meeting per week. Look out for possibilities of working individually or in subgroups and then present the results, to progress a bit faster.

  • Meeting to start learning Julia on Fri May 13 (7pm (or really 8pm, Christian and the Borg were late and the TV showed remarkable resistance, too)). Bob did a short presentation, then we did some hello world (check the history using gitk) level programming. Due to running out of time we haven't managed to work in subgroups, and haven't managed to see parallelism yet. Discussed further plans. Thinking that it is on individuals and subgroups to work out things between meetings; list of things to work on for such individuals or subgroups:

    • using matrices in Julia (using small interesting example(s)?) (perhaps Nick)

    • path search in Julia (Sam)

    • what is coming up in the course? Is there an alternative we should take? (Christian)

    Mentioned but not written down during meeting:

    • reading up about neural networks (and AI in general?) (Paul)

      (Note from Christian: check "Wikipedia" and "Online books" sections in Teaching.)

    Not listed during meeting:

    • parallel programming in Julia (Christian)

  • Meeting on Sun May 29 (7pm intended, 7:30pm start since people using the room relied on public calendar that wasn't updated):

    Sam (and Christian) showed Julia version of the path search. The implementation shown is purely functional. There's still work that could (but not necessarily will) be done:

    • add tests
    • 'check visited cities' optimization
    • set / queue instead of arrays (optimization) -> 3rd party library seen on Github
    • newer algorithms
    • parallelize

    Christian showed what he found with regards to parallelization in Julia (using its core infrastructure only) as well as in C for comparison. (Topic partially unfinished, 3rd party libraries should offer better performance.)

    Suggestion to perhaps do some sessions Code dojo style (decide on a problem to solve/implement, someone (maybe changing) sits at the computer projected to the screen, the others tell how).

  • Meeting on Sun May 5 (7pm intended, ~7:15 actual start)

    • two new members (Peter and James), Bob was there, too. Those three discuss a lot amongst them, we ponder various ways going forward.
    • Peter demonstrated Jupyter
    • Christian still didn't manage to watch videos. Showed Git slides (seemed a bit pointless though)
    • Mentioned more path search work but holding showing it off until perhaps implementing tests, (newer algorithms?, optimizations, and) parallelism

Next steps

  • Jupyter on Borg: make accessible to all. Figure out security concept. How to Git commit files. Whether working on the same sheet by multiple people at the same time works, whether working on different sheets by multiple people works.

  • should we connect Peter's machine to the net and leave accessible? How is power usage of that?

  • install LDAP user accounts in the chroot (and separate home from ubuntu)?

  • install Torch on the Debian on the Borg? (Bob has been using it.) Figure out why it doesn't use the available cores much. Install TensorFlow, Leaf (see More links).

  • Tony reads Paradigms of Artificial Intelligence Programming: Case Studies in Common Lisp

  • continue watching the "Intro to AI" videos and implement them in Julia?

  • decide whether and when to jump forward to neural network algorithms?

  • play a bit with TensorFlow (and/or Leaf and/or Torch) anyway already?

Suggestions