[Haskell-cafe] Accelerating Automatic Differentiation

Trevor McDonell trevor.mcdonell at gmail.com
Sun Mar 25 23:57:53 UTC 2018


Regarding TensorFlow (and PyTorch, etc), it is important to note that TF is
designed specifically for the deep-learning style of machine learning, but
automatic differentiation itself is a much more generally useful and finds
applications in physics, finance, optimisation... not just machine
learning. So, I think there is value in implementing AD from a more general
parallel array library like Accelerate.

I'm generally happy to add mentors. If you don't want to sign up
officially, I'm sure we can just add you to the discussion board / email
thread / whatever, and you can still contribute to the discussion.

Cheers,
-Trev

P.S. sorry for not replying sooner, I managed to lock myself out of my
office on Friday evening, so didn't have my laptop over the weekend.


On Sun, 25 Mar 2018 at 23:42 Charles Blake <cb307 at st-andrews.ac.uk> wrote:

> Thanks so much for the feedback - I really appreciate the attention you
> guys have given me. My draft proposal can be found here:
> https://docs.google.com/document/d/1bJqoqVsW5pvXJmdkfn3EXY1QKEiTXxQkl2BmREAtBsQ/edit?usp=sharing
>
> If you have the chance to look at it and are interested, any feedback
> would be valuable!
>
>
> In terms of mentors for the project, I'm not entirely familiar with the
> process involved. I've contacted the mentors originally listed for the
> project with a copy of my draft proposal, but of course I appreciate your
> input as well and if there is the option of others joining to mentor the
> project then that sounds fantastic.
>
>
> Charlie
> ------------------------------
> *From:* Michal J Gajda <mgajda at mimuw.edu.pl>
> *Sent:* 25 March 2018 09:08:54
> *To:* dominic at steinitz.org
> *Cc:* Charles Blake; Mikhail Baykov; haskell-cafe at haskell.org;
> accelerate-haskell at googlegroups.com
>
> *Subject:* Re: Accelerating Automatic Differentiation
> Great! If they all agreed, we can just add Mikhail, and done?
> On Sun, 25 Mar 2018 at 15:17, <dominic at steinitz.org> wrote:
>
> Hi Michal,
>
> I didn’t volunteer to be a mentor for this. The project already lists:
>
> *Mentor*: Fritz Henglein, Gabriele Keller, Trevor McDonell, Edward Kmett,
> Sacha Sokoloski
>
>
> I doubt there is much I could add to such an illustrious list.
>
> Dominic Steinitz
> dominic at steinitz.org
> http://idontgetoutmuch.wordpress.com
> Twitter: @idontgetoutmuch
>
>
>
> On 25 Mar 2018, at 06:11, Michal J Gajda <mgajda at mimuw.edu.pl> wrote:
>
> Given that Marco did not confirm, I am just now confirming that we can get
> you mentored by  Mikhail Baikov as the third mentor (beside Dominic and me).
> Both me (Michal) and Mikhail are performance optimization experts (I am in
> parsers, and data analytics, Mikhail is in real-time systems, he has his
> own top-notch serialization library - Beamable that outperforms Cereal in
> both data size and speed). Dominic is expert in numerical computing (ODEs
> and Julia, among other things).
>
> I believe that with these three excellent mentors you have very good
> chance to make outstanding contribution.
> We just make sure that you prep application by 27th deadline.
> --
>   Cheers
>     Michal
>
> On Sun, Mar 25, 2018 at 6:32 AM Michal J Gajda <mgajda at mimuw.edu.pl>
> wrote:
>
> As a mentor I would say it is certainly possible to outperform exsting
> mega-solutions in some narrow domain.
> Just as I did with hPDB https://hackage.haskell.org/package/hPDB
> But it requires a lot of skill and patiece.
>
> Please proceed with this project with the current list of mentors.
> I think me and Dominic have already declared committment.
>
> You might also start by making a table of best competing solutions in
> other. languages, their respective strengths, and ways that we can possibly
> improve on them!
>
> Where do You keep Your application draft? Ideally it should be a shared
> space where you can add mentors as co-editors.
>> Cheers
> Michal
> On Sun, 25 Mar 2018 at 02:32, <dominic at steinitz.org> wrote:
>
> The list of mentors for this project looks great to me. I am not sure if I
> can add much other than I think this is a nice project. Perhaps it would be
> best to get the advice of some of the mentors?
>
> For some very simple tests with an ODE solver, I concluded that accelerate
> can perform at least as well as Julia. It would certainly be very helpful
> to be able to get Jacobians for ODE solving and for other applications.
>
> Dominic Steinitz
> dominic at steinitz.org
> http://idontgetoutmuch.wordpress.com
> Twitter: @idontgetoutmuch
>
>
>
> On 24 Mar 2018, at 17:20, Charles Blake <cb307 at st-andrews.ac.uk> wrote:
>
> Thanks for the response Michal,
>
> Yes, this did cross my mind - and I wouldn't be expecting to outperform
> those frameworks in the timeframe available! I assumed that the reason that
> this project was suggested was perhaps:
>
> a) there is some intrinsic value in implementing these algorithms natively
> in haskell (hence why the 'ad' library was developed in the first place),
> so that those who want to use parallel automatic differentiation / the
> machine learning algorithms built on top of it can do so without leaving
> the haskell ecosystem,
>
> and b) because the challenges involved in implementing parallel ad in a
> purely functional language are a little different to those involved in
> doing so in OO/imperative languages - so it might be interesting from that
> angle as well?
>
> So perhaps my aim would no be to do something unique, but rather to do
> something that has already done well in other languages, but has not yet
> been provided as a haskell library. Does this sound like a reasonable
> approach or do I need to find a slightly more unique angle?
>
> Thanks,
> Charlie
>
> ------------------------------
> *From:* Michal J Gajda <mgajda at mimuw.edu.pl>
> *Sent:* 24 March 2018 16:56:35
> *To:* Dominic Steinitz; Marco Zocca; accelerate-haskell at googlegroups.com;
> Charles Blake; haskell-cafe at haskell.org
> *Subject:* Re: Accelerating Automatic Differentiation
>
>
> Hi Charlie
>
> It certainly looks like exciting project, but the bar is currently placed
> very high.
> TensorFlow package not only provides automatic differentiation for whole
> programs, but also optimizes data processing both on GPU, and reading to
> achieve large batches.
> This field has a lot of hot developments, so You would either need to
> propose something unique to Haskell, or You risk being outclassed by
> PyTorch and TensorFlow bindings
>   Maybe Dominic suggests something too.
>  Cheers
>       Michal
>
>
> _______________________________________________
> Haskell-Cafe mailing list
> To (un)subscribe, modify options or view archives go to:
> http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe
> Only members subscribed via the mailman list are allowed to post.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/haskell-cafe/attachments/20180325/b81c9e51/attachment.html>


More information about the Haskell-Cafe mailing list