Background on review & brainstorm session on Ethical Web Marchine Learning

This short video aims to give you all the information you need to participate effectively in the group sessions.

It’s about 10 minutes long, and I’m going to talk you through background what we’re doing and why, give you a sense of what Web Machine Learning is, give you a very quick primer on ethics and how it applies to ML, talk about how we’ve arrived at the Ethical principles that we’re proposing, run through the principles and then give you a sense of what to expect in the group sessions.

So let’s start with some background.

The W3C has a Machine Learning working group, and its mission is to develop specifications for APIs which enable efficient machine learning inference in web browsers.

The first one is a recommendation for a Web Neural Network API As well as the technical specifications, the group’s charter also has as a deliverable a Working group Note documenting ethical issues associated with using Machine Learning on the Web, to help identify what mitigations its normative specifications should take into account.

And that’s what this video is all about.

So what exactly is Web Machine Learning?

At a very basic level, in web-based ML applications running on a local client, you have a model, which is like a set of instructions, and then input data which needs to be processed.

The model may reside in the cloud, or on the local client.

But the key thing is that the data processing, or inference, can be offloaded to the client, or local hardware, rather than needing to happen in the cloud.

This is already possible, but the lack of access to dedicated client-side capabilities like ML hardware accelerators constrains the scope of experiences and leads to inefficient implementations on modern hardware.

This is what the working group is looking to improve.

So what could this capability be useful for?

Some example use cases include person detection, facial recognition, image captioning, machine translation and noise suppression.

Web machine learning has a number of potential benefits.

It could make large-scale deployment of ML systems feasible without investment in cloud-based infrastructure, opening the door to do-it-yourself web developers.

It aligns this technology with the decentralized web architecture ideal that minimizes single points of failure and single points of control.

Local processing could also enable machine learning use cases that require low latency, such as object detection in immersive web experiences.

With appropriate safeguards, enabling machine learning inference in the browser (as opposed to in the cloud) could also enhance privacy, since input data such as locally sourced images or video streams stay within the browser's sandbox.

So that’s what the working group is trying to do.

Why does it need an ethical approach, and just to back up a bit, what is ethics?

Ethics is about what is right and wrong, good and bad.

It provides a rational framework for thinking about what’s right and wrong and making decisions about how to act accordingly.

So do we need an ethical approach to Web Machine Learning?

Well, technology is never neutral - it will always have social and ethical implications.

The question is whether these are actively considered and addressed, or not.

Given the scale and depth of the impact that AI/ML is anticipated to have, failure to consider the ethical implications could cause great harm.

And indeed we already see examples in the news of the sorts of harms and issues that arise from machine learning.

A couple of the key ones are

• Bias & Fairness: for example where facial recognition technologies work less well for people of colour; or a search for ‘CEO’ returns mostly pictures of white men.
• - Another issue is transparency & explainability: there is widespread concern about ‘black box’ algorithms where it is impossible to know why they arrived at particular decisions, particularly in high stakes areas such as criminal justice. - And another issue is sustainability: increasing awareness that computationally complex ML approaches trained on very large data sets can have a large environmental impact, given the amount of energy required to power the training phase.

So if we have ethical concerns like these, one way to deal with them is to develop an ethical framework, which is really a structured process to get to concrete ethical guidance.

It starts with some principles, which tend to be quite abstract or high level.

They may be accompanied by some guidance to make them more specific.

The principles can then be used as a guide to think about more specific risks and harms associated with any given use case.

And from there, you can think about the ways to mitigate those risks or harms - at which point we’ve arrived at the concrete actions we need to make sure our approach is ethical.

This move from the abstract principles to concrete mitigations is known as operationalisation.

So the first thing to start with is some principles.

So how did we get to our?

We decided to try to base them on an existing set of principles rather than generate our own from scratch - partly because we don’t have the resources to do the sort of wide multi-stakeholder consultation you need, so we’d like to leverage the effort that others have put in to do that.

So to find existing principles, we went looking with some criteria - and that was our first one:

• Universality: we wanted to see evidence of a diverse, global range of stakeholders involved in their development and broad acceptance of the final result.
• Another criteria was coverage: they should be as complete as possible, while not unnecessarily broad, as evidenced by alignment with key principles found in meta-analyses of AI ethical principles.
• And the final criteria was alignment with relevant existing W3C principles and guidance.

So we looked at a number of sets of principles, and the winner was the Recommendation on the ethics of artificial intelligence by UNESCO, which is the United Nations Educational, Scientific and Cultural Organization.

These have been

• developed through a global, multi-stakeholder process, and have been ratified by 193 countries.
• - They cover all the main themes without going too far beyond, although they do lack an explicit principle of ‘Autonomy’.
• - There is also good alignment with relevant W3C concerns, although again, comparing that, you could see the lack of an ‘Autonomy’ principle.

So what are the principles that UNESCO recommends?

They actually have four Values and ten principles, to which we’ve added the addition principle of autonomy for the reasons just discussed.

The values are meant to indicate desirable behavior and represent the foundation of the principles - they’re big picture statements.

They are:

• Respect, protection and promotion of human rights and fundamental freedoms and human dignity
• Environment and ecosystem flourishing
• Ensuring diversity and inclusiveness
• Living in peaceful, just and interconnected societies

The principles are meant to unpack the values more concretely so that the values can be more easily operationalized.

And the principles are what you really need to care about.

They are:

• Proportionality and Do No Harm
• Fairness and non-discrimination
• Autonomy
• Right to Privacy, and Data Protection
• Safety and security
• Transparency and explainability
• Responsibility and accountability
• Sustainability
• Human oversight and determination
• Awareness and literacy
• Multi-stakeholder and adaptive governance and collaboration.

So in terms of developing our ethical framework, we’ve got some principles, and they’ve been supplemented by guidance to elaborate on them and make them clearer and more concrete in the W3C / Web Machine Learning context.

The principles and guidance are part of a draft of the Working Group Note which has been circulated online for review and comment over the past month.

Which brings us to the group sessions.

So what can you expect?

We’ll do some more in-person review and discussion of the principles and guidance.

And then we’ll use the principles to generate some more specific risks and mitigations in a brainstorming exercise.

There are two main benefits, and reasons to do this One is that it’s helping us fill out our ethical framework.

But the framework will never really be finished - this is such a fast moving technology, and the ethical challenges vary so much according to context, that there will always be new risks to consider.

Also, ethics isn’t about ticking off a checklist - it’s about thinking things through.

So the second benefit we hope comes from this is that you’ll be learning by doing - learning a process to take the principles and use them to think through relevant risks and mitigations.

So hopefully that’s all made sense and given you the background you need for the group session - we look forward to seeing you there.

Thank you!