W3C home > Mailing lists > Public > public-webmachinelearning-wg@w3.org > June 2021

Re: CfC: adopt WebNN API and publish as a First Public Working Draft; respond by 17 Jun 2021

From: Jonathan Bingham <binghamj@google.com>
Date: Thu, 17 Jun 2021 20:06:29 -0700
Message-ID: <CAEK6eFz0D1QAW2Y3wspvu_ya+DMKCsLhB=F_xoJArAvvw+m0vw@mail.gmail.com>
To: "Kostiainen, Anssi" <anssi.kostiainen@intel.com>
Cc: "public-webmachinelearning-wg@w3.org" <public-webmachinelearning-wg@w3.org>
Several of us at Google  have reviewed the Web NN proposal, from Chrome,
TensorFlow, Chrome OS, and Android. We recommend that the Working Group
move forward with this proposal and publish the First Public Working Draft,
with a few caveats that we have already expressed in the Community Group.

The first caveat is that the field of machine learning is rapidly evolving,
and there is not yet a stable and widely accepted operation set definition
that a neural network API can be based on. The most popular operation sets
-- including those from Google, Facebook, Apple, and Microsoft -- have
grown and changed rapidly, even in the past year, and they are likely to
continue to evolve rapidly. It will require a large effort for web
standards and browser implementations to keep up with the field, even after
the initial launch of a web API. Furthermore, there is a real risk that a
different approach will ultimately be more successful.

The community group has been taking a sensible approach to operation
definition, decomposing operations into the smallest instructions that can
be reused. Backwards compatibility is being considered, and there's reason
to be optimistic that changes can be additive, rather than breaking.

A second caveat is about developer demand. Before adopting Web NN as a
standard and shipping in browsers, it will be important to validate that
existing web standards are inadequate to deliver the performance that web
developers and end users require. As evidence of that demand, we would all
want to see that a critical mass of web developers had tried running ML
models using existing web standards, like WASM, Web GL, and Web GPU, and
found them too slow.

In the community group, we all believe the theoretical case is strong that
better performance is possible and will make a difference. The community
group has been conducting extensive performance benchmarking. It is not yet
representative of the wider developer community. The First Public Working
Draft will help to broaden the conversation, and we hope to see evidence of
developer demand.

In parallel to Web NN, we recommend that the group stay abreast of advances
in ML, and explore multiple options in parallel, including alternative ML
APIs and enhancements to WASM. It remains to be seen whether Web NN will be
the best approach, and there is a real risk that it will launch and quickly
be superseded by alternate approaches.

All of that said, Web NN is well worth exploring, gathering developer
feedback, and moving forward in the standards process. There's strong
potential for meaningful performance gains that can make web experiences
more compelling. We recommend moving forward with the proposal and
publishing a First Public Working Draft.

Cheers,
Jonathan


On Thu, Jun 10, 2021 at 9:26 AM Kostiainen, Anssi <
anssi.kostiainen@intel.com> wrote:

> Hi Web Machine Learning Working Group,
>
> This is a Call for Consensus (CfC) to adopt the Web Neural Network API
> (WebNN API) [1] from the Web Machine Learning Community Group (CG) into
> this Working Group and publish it as a First Public Working Draft (FPWD) as
> per the WG Charter [2].
>
> This specification has been incubated in the Community Group and has
> received initial security, privacy and architecture reviews. The CG has
> also developed and maintains a polyfill and a native implementation that
> evolve together with the specification. Considering these signals, we
> believe the WebNN API meets the expectations of a First Public Working
> Draft and is ready to start its journey on the W3C Recommendation Track.
>
> It should be noted, a specification published as a FPWD is expected to
> change as the work progresses and the group receives further feedback from
> implementers, ML frameworks, web developers, through horizontal reviews and
> other groups working in this space.
>
> As this is the first Call for Consensus we're running in this group, some
> background on what a call for consensus is: the W3C Process requires chairs
> to assess the level of consensus before taking some of the group's formal
> decisions. To enable as broad input from Working Group participants as
> possible, including those that may not be in a position to join synchronous
> discussions during teleconferences, we will in general assess this
> consensus by sending these Call for Consensus messages (sometimes
> abbreviated as CfC) to the Working Group's mailing list, and collect
> feedback for a defined duration (usually between 7 and 10 days).
>
> Please indicate any objections or concerns on this list before EOB 17 June
> 2021. Silence is considered consent.
>
> Thanks,
>
> -Anssi (WebML WG Chair)
>
> [1] https://webmachinelearning.github.io/webnn/
> [2] https://www.w3.org/2021/04/web-machine-learning-charter.html
>
Received on Friday, 18 June 2021 03:17:33 UTC

This archive was generated by hypermail 2.4.0 : Thursday, 24 March 2022 20:38:45 UTC