Meeting minutes
<paulmcd> Good morning !
[We're shuffle the agenda to match participation]
Inference API to load and run a model
anssik: Discuss current status and next steps (move repo to CG?), the group resolved to add the API to the group's scope.
anssik: Jonathan created an issue to track the next steps, Jonathan to give a brief update?
jonathan: I filed an issue for the defining the plan, at this point no details yet, but have talked with folks internally familiar with the standards work and process
…
gregwhitworth: WICG is equivalent forum to dedicated CGs like this one for incubation
anssik: WICG is catch-all CG, for this spec WebML CG is the right topic-specific CG
gregwhitworth: Jonathan is doing the right things with the incubation, have explainer available in public and accepting feedback
… when it graduates, the incubation moved to a Working Group, in this case to-be-created WebML WG
jonathan: that sounds great
gregwhitworth: if you like, you can create a thread in WICG Discourse and add a pointer to the WebML CG where the explainer and incubation happens
<gregwhitworth> https://discourse.wicg.io/c/html
anssik: we can migrate https://github.com/jbingham/web-ml-inference to https://github.com/webmachinelearning/web-ml-inference
jonathan: sounds good
Rafael: is this API ready for prototyping at this point or for discussion?
jonathan: a group of Google folks are interested in looking at this, figure out who the engineers working on prototype would be
… we have some ideas how to prototype this in custom Chromium build
… a bit similar to what Ningxin is already doing
… we'd also like to get to agreement with Microsoft on details to get backends aligned, much of the code would like in Chromium and everyone working off that codebase would share code
RafaelCintron: agree, that sounds like a good model, so everyone could integrate with their own backends
jonathan: allowing polyfilling, ONNX.js and TF Lite we may be able to do that
<paulmcd> Love those . Poly fill, onnx, tflite
RafaelCintron: there's no active development on ONNX.js but could use it for experimenting with this API
Virtual workshop
anssik: As you're aware, our in-person Web & ML workshop was postponed due to the coronavirus outbreak
[POSTPONED] W3C Workshop on Web & Machine Learning
anssik: to overcome this situation without blocking, I've been planning with W3C's Dom a virtual workshop that'd replace the in-person workshop. Our estimate is it'd be at least 6 weeks to pull this off, so earliest it'd happen mid-May onwards.
anssik: First, the Program Committee would build an agenda based on its assessments of what topics would benefit to be covered. We'd use the existing in-person workshop agenda as a starting point (https://bit.ly/webml-workshop-talks).
anssik: The agenda would be built around 2 main modes: an offline mode and an on-line synchronous mode.
anssik: - an offline mode, built from pre-recorded videos of identified speakers on identified topics - ideally, reasonably short, focused talks that give a high-level of the current status on the said topic, and highlights the unsolved aspects, and the role of standardization on this.
anssik: - an on-line synchronous mode organized as a series of four 90 minutes video conferences spread over 4 days, at a single timeslot (optimized to allow as many of the registered participants to join), where all the (selected) participants to the virtual workshop are invited to join and participate. The said live sessions would be built over a well-defined agenda allowing to exchange on questions and topics exposed during the talks,
with speakers expected to participate at least in the sessions where their topics are covered.
anssik: Any comments on the modes: offline mode and on-line synchronous mode? Does this sound good? We need to consider timezone differences, which was a key consideration in this proposal.
gregwhitworth: I like the offline aspect
… when reading through the in-person agenda, think we can subset the existing agenda and keep relevant topics
… Web Components F2F did 2-3 hours over two days and it was very productive
<gregwhitworth> that sounds good
anssik: To make this work, the following infrastructure would be provided by W3C:
anssik: - a platform to serve the video of the talks to a global audience
anssik: - a support team to inform and remind speakers of the tools, logistics and timelines to capture and upload the recorded talks
anssik: - a platform to host and manage the live conversation with up to 100 (or more?) participants (most likely, Zoom); optionally, the said conversation could be augmented with live audio translation of the discussions
anssik: - a slack instance (?) reserved to registered participants to meet & greet with other participants before, during and after the live sessions
anssik: To make this work, the expectation would remain that the live and interactive (slack) parts of the virtual workshop would be reserved to selected participants. To that end, interested people would be requested to apply indicating their relevant background on the topic
anssik: So that's the general plan, open to comments and suggestions.
RafaelCintron: I've found other groups using Google Meets, e.g. WebGL/GPU
… could also use Microsoft Teams, makes sharing slides easy
RafaelCintron: WebGPU CG meetings are 7-8 people across the world, using Meets
anssik: we'd need a platform that can scale up to 100 participants
https://bit.ly/webml-workshop-talks
RafaelCintron: I like the concept of offline
… also like the effective use of sync time where everyone has seen the offline material
<paulmcd> I like the offline up front as well
anssik: let me start look into a subset of the agenda that would be interesting for the audience, I'll ask your feedback on that