A new research poster reads a three-channel classifier as an assembly that decides by vote, and finds its voice channel holding a third of formal power while casting under a fifth of the decisive votes.
A political theory reading of how a multimodal classifier counts its own verdicts, and who it stops hearing
A machine that weighs three kinds of evidence can still end up listening to two of them. That is the finding of The Permanent Minority, a new research poster from the School of Continuous Improvement that brings the theory of representation to the way software combines what it reads, sees and hears.
Every selection panel I scribed for had one member whose view went into the minutes as a dissent and went no further. It was strange to meet the same arrangement inside a piece of software, and good to have the chance to sit with it properly.
— Ms Signy Okpara, Doctoral Candidate, School of Continuous Improvement
The poster, led by doctoral candidate Ms Signy Okpara with Dr Mirela Hanke and Associate Professor Anselm Threlkeld of the School of Settled Matters, treats a classifier screening 1,140 recorded mock interviews as a small deliberative body whose members are its transcript, face-video and voice channels. Bringing the tools political theorists use to weigh voting power into conversation with the logs of everyday automated decisions, the team finds that each channel holds an equal share of formal power, and that the voice channel nonetheless stood alone in 214 of the 331 decisions on which the channels split.
The work positions the School’s long-standing interest in how decisions are recorded at the meeting point of machine learning and constitutional thought, and proposes that an outvoted channel’s verdict be kept in the record as a dissent.
“The Living Dashboard keeps a great many records of decisions,” said Dr Hanke, Postdoctoral Fellow and Deputy Convenor of the Living Dashboard. “This work has given us a good deal to think about the next time we decide what one of them should hold.”
The full poster is available from the University’s research repository under an open licence, doi:10.5555/slop.fkfqft.