Meeting minutes
Announcements
anssik: welcome everyone to the last WebML WG teleconference of 2021
… we have a special guest today and a packed agenda as usual
anssik: we're welcoming James to talk to us about ethics in a minute
anssik: we've created a "how to" resource to folks who want to help with scribing, linked from the agendas from now on
Intent to Prototype in Chromium: Model Loader API
WebML CG telcon for Model Loader API poll results
anssik: we're planning to run Sydney-friendly meetings on the model loader API
… plan to kickoff on Jan 12 at 5am UTC
Geun-Hyung: will check my availabilities
TPAC meeting follow-up (cont'd)
Ethical issues in using Machine Learning on the Web
anssik: one of the topics we discussed briefly at TPAC was ethics
… it got also highlighted in the intent to prototype discussion in Chromium
… work has started on this
… I'm happy to introduce James Fletcher today who is interested in helping us
… Dom and I met a couple of times with him
… and James is bringing his ethics experience in this space
Slideset: https://
James: happy to be getting involved in this project
… want to give a quick intro on myself and the BBC work in this space
… and ideas on how to move this ethics work forward
James: the challenge I'm getting involved in is the development of the ethical considerations around ML and potential mitigations
James: The challenge "What are the ethical considerations that should inform the development of the Web Neural Network API, and what mitigations should be taken into account?"
James: I lead the work in the BBC in how we use ML as a user-facing organization
… I'm not a technical person, but hopefully I can bring a different perspective
James: BBC plays a role in standard as part of its R&D of work, as our charter recognizes
… we work in ~40 standards groups, including W3C, esp in the audio/video/accessibility areas
… It's exciting that we're getting involved in the ethical side of things
James: wrt AI/ML, our approach is anchored to the response we gave the house of lords committee in 2017
… my role is anchored in our 3rd commitment to responsible technical development
… #2 is also relevant to getting involved here
James: our work in this space is anchored in MLEP, Machine Learning Engine Principles
James: trust is a key value for BBC, which is reflected in our approach in this space
James: The Principles
… Reflecting the BBC’s Values
… Improving Audience Experience
… Clear Explanations
… Editorial Values & Broadening Horizons
… Taking Responsibility: Review, Security & Fairness
… Human in the Loop
James: MLEP Checklist 2.0 Overview
[A diagram depicting the checklist]
James: Scoping your ML project
… Intended ML applications
… Impact
… Risks, opportunities and consequences
James: Consequence Scanning Toolkit
… to think through ethical considerations
James: AI & ML at the BBC
… two dimensions:
… internal-facing <-> audience-facing
… content data <-> audience data
James: How might we apply an ethical approach to the web neural network API?
James: Key considerations
… Build on existing W3C ethics work
James: Involve a wide range of stakeholders
… Use internal + external expertise
… Build towards broader web machine learning principles
James: Discover > Define > Develop > Deliver
James: How might we get there?
… Discover: Internal + external literature review, Consult experts
… Define: Guiding principles
… Develop: Workshops: Issues & Mitigations
… Deliver: Ethics Recommendations
James: Questions from me ...
… Thoughts on the proposed process
… Timeline considerations
… Recommendations for key stakeholders, internal and external experts
anssik: thanks for the great presentation!
anssik: As for timeline
… we're aiming for a spec milestone "Candidate Recommendation" in 1H 2022, an ambitious goal but an important one
… would love to have a fleshed out document on Ethical Web Machine Learning to link to by that time.
James: would the ethics recommendations include mitigations that are prototyped?
Rachel: thanks for the great presentation and Anssi for connecting all the people
… about the framework, have thoughts of organizing an event where we can conduct policy research across different companies
… does policy research across companies make sense to you?
… companies outside of W3C also
… what is achievable from my end, within W3C community and outside at the corporation level
… including government, community and corporations
… policy research initiative to see how this is reflected at each corporate level
James: AI ethics policies companies, govs have?
… that'd be part of literature review
… gov bodies and pan-gov entities have, this is meta research that looks into similarities
… similar themes are surfacing from many places
Rachel: I'm proposing us to facilitate research across companies, within W3C community and outside, looking at existing policies
James: literature review is meant to provide view into that
Guen-Hyung: what is the approach in context of web technologies?
… the scope of ethics, includes bias
Rachel: excellent question between policy and implementation
… policy driven implementation and the other way around
… we hear news people are surprised how companies use AI today
… everyone see ethics differently, but there are some key themes that surface
… there are gray areas
James: two thoughts, in terms interpretations of ethics, absolutely!
… getting people together with diverse perspectives is the way to go
… we have UNESCO working on global level
… principles such as autonomy, allows the users to make their own decisions
… a lot of ethical principles are like that, privacy, security safeguarding individuals
Path to Candidate Recommendation
anssik: we are chartered to push toward Candidate Recommendation specification maturity in 2022.
… this is an important spec milestone, essentially says "the spec is feature complete and can be implemented as is"
… initially, this was a spec publication that triggered implementation work, in practice today implementation work start before CR
… to increase transparency and allow us to move toward this CR milestone in a coordinated fashion, I created two meta issues that act as trackers.
… the first one, Candidate Recommendation readiness tracker, turn process requirements into checkboxes we want to tick:
Candidate Recommendation readiness tracker #240
anssik: key requirements demystified:
… 1) "must show that the specification has met all Working Group requirements,"
… spec addresses the requirements defined at chartering time and satisfies its use cases
… the W3C process does not micromanage the requirements or use cases, up to the WG to be responsible
… 2) "must document how adequate implementation experience will be demonstrated"
… essentially tests and early implementations, with a caveat implementations do not need to exist to get to CR
… also no requirement to get explicit support signals from all implementers
… 3) "must show that the specification has received wide review"
… this is an important W3C process step we must complete, we're tracking this separately:
Jonathan: I see the Model Loader API incubation and WebNN API on the Rec Track as complementary approaches, good to get feedback from web developers for both
… WebNN is more mature at this point, fine to be further along and it makes sense
… we need to allow time for wider feedback and OTs for both of them
Happy Holidays!
anssik: We plan to resume WG telcons 13 Jan 2022, check the invite for details
… and kick off CG telcons on 12 January
… please enjoy the upcoming Holiday season!
… 2022 looks exciting for this WG with many important milestones.
Rachel: thanks for your leadership Anssi and James for an amazing presentation!