How do we prevent pre-existing biases from bleeding into the artificial intelligence systems we’re increasingly relying on for day-to-day tasks? By paying closer attention to the data and machine learning used to build those AIs.
Google’s new initiative, PAIR, or People + AI Research, unveiled on Monday, seeks to put human considerations like this at the front of AI development.
In a blog post on PAIR, senior staff research scientists from the Google Brain team, Fernanda Viégas and Martin Wattenberg, write, “By building AI systems with users in mind from the ground up, we open up entire new areas of design and interaction.”
The team believes part of the solution for how to build AI systems that consider actual human needs is to treat AI less as a collection of algorithms that you can apply to various problems and more as a design material or medium like paint. They also want to pair that perspective with tools and training that help those building AI and machine learning systems view it and develop for it in a more people-centric way.
To help guide AI development, the team will focus on three areas: Tools for AI engineers and researchers, how they can build tools to help people working in verticals like healthcare, farming and music effectively apply AI, and the inclusivity of AI systems.
It’s that last part that may resonate with people inside and outside of Google.
PAIR is not Google’s first “let’s get a handle on AI” effort. The tech giant is a founding member of the Partnership on Artificial Intelligence to Benefit People and Society (“Partnership on AI” for short). Both groups recognize that the key to unbiased AI is better training data (Machine Learning systems are typically trained with mountains of example data to build artificial intelligence around things like object and face recognition, speech translation, and even cancer research).
Last year, Partnership for AI founding member and Microsoft Research Technical Fellow and Managing Director Eric Horvitz told me, “Bias in data can get propagated to machine learning that can lead to biased systems.” He also worried that biased AI could impact, “how racial minorities are handled when it comes to visual perception and facial recognition.”
While Google acknowledges that Partnership on AI and PAIR are complementary efforts, the two are not directly tied together. Even so, a key component of PAIR is the open-sourcing of two Google visual tools, Facets Overview and Facets Drive. Both deal with training data.
“We think this is important because training data is a key ingredient in modern AI systems, but it can often be a source of opacity and confusion. Indeed, one of the ways that ML engineering seems different than traditional software engineering is a stronger need to debug not just code, but data too,” wrote the Google Brain Team.
It’s not just the AI world at large that has work to do on AI data and output. In April, researchers at University of Bath in the United Kingdom and Princeton University found gender-based bias in numerous AI systems, including Google Translate where it automatically “converts gender-neutral pronouns from several languages into ‘he’ when talking about a doctor, and ‘she’ when talking about a nurse.”
Google never explicitly mentions “bias” in its PAIR announcement, but it’s clear that the company wants to use early human input to make AI output a little less artificial.