Building experts into AI
One widespread concern over AI’s place in the future is that many, or even most, meaningful areas of human employment will be made redundant. Already, we’re seeing jobs lost to automation in sectors we couldn’t have foreseen a decade or two ago.
But what about the jobs that AI is creating? While the emergence of AI has certainly meant lots of meaningful work for the programmers who actually write the learning algorithms, more numerous by far are jobs in data labelling: the creation of the vast quantities of training data that AI algorithms learn from. In our last blog we underlined the importance of ethical labelling practices as a component of any ethical AI system. Is it possible to reconcile ethical labelling work with competitive AI practices? When it comes to labelling, won’t the lowest bidder always win?
One area of promise here is the so-called “human-in-the-loop” (HITL) approach to AI. In fields where the abilities of AI are stretched to their frontiers, like aerial whale detection, HITL systems get the algorithm to do the bulk of the work, then defer to human experts’ judgements on “edge cases”, where the detection algorithm is less confident (Yes, computers can be uncertain too!) This way computers do the parts of the work where they can add the most value (in speed/efficiency), while humans do the parts of the work where they add most value (via their discernment in edge cases).
As an added bonus, labelling edge cases are precisely the kind of work that’s most interesting to humans. Think: “huh, that’s a wave, but I can really see how it looks like a whale”, as opposed to: “wave, wave, wave, wave, whale, wave, wave…”
Our Lead Biologist, Bertrand Charry, on the job.
And it gets better: besides leveraging human talent more effectively, HITL also allows the model to re-train and improve, since the expert’s judgements on edge cases target precisely the gaps in “knowledge” of the original algorithm. What’s good for humans, in this case, is also good for AI — and the interchange between them allows for higher accuracy and speed than either could achieve on its own. This is especially true for tasks that push the boundaries of machine learning capabilities. For instance, Stitch Fix is an online personal styling service that uses a HITL approach to provide clothing recommendations to customers that improve over time, using human acumen to fine-tune AI-generated recommendations.
Whale detection is another task that’s ripe for the “HITL-ing”. Manual detection is slow and often tedious work, but fully automated detection often loses in accuracy what it gains in efficiency, or else is tied to the specific context it was trained for (e.g. calm water, no ice, different target species). That’s why Whale Seeker is excited to introduce its new labelling tool, Möbius, which applies the HITL approach to marine mammal detection. Rather than replacing expert biologists, we’ve built them into our product, empowering them to cover more imagery, faster. Whale Seeker is also developing detection approaches that go beyond the spectrum of visual light (using infrared, for example), further expanding human capabilities. These added efficiencies don’t replace biologists — they give them super powers!
