– DRAFT –
WebML CG Teleconference – 10 January 2019
10 January 2019
<kreeger> + Nick Kreeger
anssik: Welcome back! We've kicked off 2019 with good discussion and new members e.g. from Google, Brain.js community.
-> High level vs low level follow-up https://github.com/webmachinelearning/webnn/issues/3
anssik: Thanks for active discussion and proposals in the high vs low level issue
… It is important we find alignment in the group around this topic before we advance to concrete API spec work
anssik: Goal today to review & solicit feedback on the existing proposals, see if we can agree on the next steps
… Identified 3 proposals to start, open to more suggestions:
… a) Incubate both the operator and load model approaches in parallel
… b) Focus on the operator approach and develop the load model approach as a polyfill initially
… c) New HTML element similar to <video>
… Also please note recently contributed "WebAI" proposal by a Brain.js community member
anssik: Option b) would not require changes to the charter, options a) and c) might require a charter amendment (we have a process for that)
anssik: I'd like to open the floor for comments.
[gregwhitworth summarizing his proposal at https://github.com/webmachinelearning/webnn/issues/3#issuecomment-447080045]
gregwhitworth: would prefer option a)
chrisheilmann: would also support option a)
danielm: How quickly we expect to develop the standard? Can we split the standard development in phases?
kreeger: from Google side prefer operators approach, i.e. option b)
<nsthorat> WebAI proposal: https://github.com/DanielMazurkiewicz/WebAI
[danielm introducing Brain.js]
<kreeger> Google opinion: Prefer ops in this proposal over model loading. Model architecture and format is very fluid and continues to evolve. We think it makes sense to wait until model formats and more flushed out before proposing a format.
anssik: Could you use the operator-based WebNN API as an alternative backend for Brain.js?
danielm: yes, from technical perspective
anssik: hearing we need to continue discussion in the issue toward concensus
<kreeger> I think more debate vs. a and b
-> Web Neural Network API https://webmachinelearning.github.io/webnn/
anssik: Use cases landed, thank you to all contributors!
-> Use cases https://webmachinelearning.github.io/webnn/#usecases
anssik: proposed next step to derive requirements to measure API against
-> [Example] WebRTC requirements derived from use cases https://w3c.github.io/webrtc-nv-use-cases/
anssik: Do people feel a list of concrete requirements would help us make more informed API design decisions?
anssik: Support to derive requirements?
chrisheilmann: I'd prefer to do that, also look into new use cases. Do more outreach.
PROPOSED RESOLUTION: Add requirements derived from use cases to the spec
<daniel> https://github.com/BrainJS/brain.js
Resolved: Add requirements derived from use cases to the spec, and add more use cases
<Ningxin_Hu> semantic segmentation use case: https://webmachinelearning.github.io/webnn/#usecase-segmentation
anssik: Ningxin has an update on semantic segmentation use case investigation
-> WebNN Use Case Investigation: Semantic Segmentation slides https://docs.google.com/presentation/d/1FYiirLCQgAZbGLxm9zHDuRSsSFZpvUmwUVYQqdOpQBM/
[Ningxin discusses material on slides: use case, example, identified requirements of DeepLabV3+ MobileNet V2 model]
<gregwhitworth> custom operator capability
anssik: a good to have a new use case and requirement for custom operators
tomoyuki: We should discuss custom operator proposal more, how to implement it. What is the coverage of operators?
anssik: Call for editors. I'd like to see if there are volunteers to help edit the spec(s) in this group.
… Simply put, the responsibility of an editor is to translate group's consensus position into spec text, resolve and triage issues
anssik: Last month I shared with you a plan for a proposed W3C workshop.
… After discussion with key stakeholders, decided to postpone the workshop from June to October 2019.
[DRAFT] W3C Workshop on Web & Machine Learning
anssik: The workshop draft is open to further feedback.
danielm: If we'd like to include training in scope, how to do that?
<dsmilkov> Thanks everyone and thank you Anssi for leading this meeting!
[Re scope expansion, process for charter amendment https://webmachinelearning.github.io/charter/#charter-change asks for 2/3 votes to pass.]