On October 21st & 22nd, I attended the First symposium on Biases in Human Computation and Crowdsourcing in Sheffield, UK organized by my colleagues and friends Alessandro Checco (Uni of Sheffield) and Eddy Maddalena (Uni Southampton).
There were 30 participants, majority from UK, Netherlands and US. The program consisted of 2 invited talks – 1 of them mine – 2 keynotes, 9 full talks and 6 short talks.
I spoke about my work on improving the quality of predicted drug indications through data integration, crowdsourcing and machine learning. I discussed our preliminary results of launching microtasks on the Figure Eight platform to employ workers to identify the relation of a given drug-disease pair in the provided drug label. We discussed about the next steps of designing microtasks where workers need to annotate context (e.g. this drug cannot be used by pregnant women) in a drug label. The question is whether we want to show workers what previous workers have annotated or what experts have annotated as context and does this introduce a bias? That’s some thing we need to experiment with.
The first keynote was on “Democratic self-regulation and fairness in future cyber societies” by Prof. Michael Rovatsos (Uni Edinburgh).
He started by discussing the dystopia (enabling ubiquitous surveillance, getting innocent people into prison) and the utopia (measuring citizens needs and desires, rewarding contributions in fair ways) of future cyber societies. He then went on to explain the concept of “Society-in-the-Loop” (SITL) concept (by Iyad Rahwan):
While Human-in-the-Loop AI is about embedding the judgment of individual humans or groups in the optimization of narrowly defined AI systems, SITL is about embedding the judgment of society, as a whole, in the algorithmic governance of societal outcomes.
The focus is on proposing a sustainable model for adapting AI systems design to societal expectations, which requires (i) knowing what types of behaviours people expect from AI; (ii) enabling the public to articulate these expectations to machines and (iii) methods to evaluate AI behaviour against quantifiable human values. These are ways of mitigating bias. Particularly for fairness in Machine Learning, bias is inevitable in order to generalize and detect patterns. There is a need to avoid unintentional discrimination of protected groups and build fair algorithms.
He asked several ethical design questions such as “should all users be given same weight in calculating overall welfare?”, “what parameters of the algorithm can users configure?” and “can we give paying customers preferential treatment?”. In practice today, there are some algorithmic decisions that people consider unfair such as “social media posts not showing on one’s wall”, “recommendations based on information gathered without permission” and “personalized recommendations based on information gathered without user permission”.
The other talks ranged from use cases about real-time crowdsourcing, creating anaphorically annotated resources, description of people images, conversational agents, AI-assisted peer reviews and comparing measurement of perceptions of energy density and carbon footprint via citizens. I found this very interesting: a Cognitive Bias Codex, which breaks down cognitive errors into four quadrants of memory, meaning, information overload, and need for speed:
Each talk focused around exploring, capturing and mitigating bias in human computation and answering questions such as “Do social interactions amplify our social biases?”, “does induced bias help in improving task success?” and “what legal practices do we need to deal with implicit human bias?”. We are on the right path but on a long way to find solutions to these questions. The take away message was that one must not disregard bias affects our results when it comes to human computation and crowdsourcing.