About a year ago I was at a tech conference and there was one topic that threatened to overwhelm all others: no matter how a conversation started, it always ended up being about the fear and uncertainty of what would happen when the robots take over our jobs. Last month I attended O’Reilly’s Artificial Intelligence conference in San Francisco and, perhaps not unexpectedly, the dominant topics were completely different.
Data: more is more
The biggest theme throughout the conference was data. Machine learning and deep learning in AI presents infinite potential and could impact multiple industries, yet, nobody seems to have enough data.
To be more specific: not enough labelled data. This is a problem across all domains and although some fields benefit from the existence of large open data sets (e.g. ImageNet), even these areas would appreciate fresh data to explore. In recognition of this need, companies are popping up that offer a data labelling service. You send them your dataset (unlabeled), they crowdsource a team to label it for you and send it back.
Clearly, in today’s gig economy there are endless opportunities for exchanging value. And anyone working in this space will agree that this truly is a valuable service.
The challenge for businesses is that it is often impractical to outsource the labelling of data either due to the need to protect competitive advantage or the need to protect the private information of customers. In fact, even without these concerns, in many cases the ‘crowd’ would not have sufficient domain knowledge and expertise to label the data accurately and so the availability of data will continue to limit how much can be achieved with these techniques inside businesses.
Context is king
The framing of the problem was another interesting theme that emerged. There was a lot of discussion about design and, while this included the concept of understanding the end user of a system and designing appropriately for them, it also extended to having a deep awareness of the context of the problems we are solving.
Although this might sound trivial, it often isn’t. We might know the problem that we are trying to solve, but the true end goal or outcome is sometimes obscured by complications along the way, or perhaps more commonly, we aren’t always aware of the peripheral effects of our solution.
There is a ‘last mile problem’, which is a concept from behavioural economics, where there can be a gap between the immediate outcomes of a system and the overarching goal for which it was built in the first place. How do we ensure that the capability we are creating through AI is integrated into society in a way that can bring about the intended benefits?
Anyway, it’s not the robots we have to worry about
Along with the intended benefits, there are always some unintended consequences too. This is the concept of a runaway objective function where a system diligently pursues its stated goal but can have disastrous consequences if the peripheral effects are not explicitly accounted for as part of its design.
This is the modern version of cautionary tales of genies granting wishes accurately but not to the satisfaction of the requester. In reinforcement learning, this concept is illustrated quite literally when reward functions fail to take all factors into account (e.g. planning a route with no traffic but failing to penalize for excessive distance travelled and time taken could result in a very indirect and impractical solution).
The extreme of this runaway objective function is central to many science fiction plots of robots ‘taking over’ (think Isaac Asimov’s iRobot). Although this is something designers of AI systems should keep in mind, I agree with Zeynep Tufekci who points out that “…too many worry about what AI – as if some independent entity – will do to us. Too few people worry what power will do with AI”.
by Bryony Martin