- This report is a part of the following series: SCC Distinguished Seminar Series.
- This piece was written over a year ago. It may no longer accurately reflect my views now, or may be factually outdated.
- This piece was originally written for my old site, Oh What? Oh Jeez! As such, it may not have transferred over properly and some images and links might be broken (and, to a lesser extent, my writing from years ago is about 80% run-on sentences).
Rather than aiming to build systems which replace human capabilities, we need to design systems that work in partnership with users.
Abigail Sellen, of The Myth of the Paperless Office fame, came to deliver an instalment of the SCC Distinguished Seminar Series on the
process of . That final line cropped up a few times, and forms part of the most interesting takeaway of the seminar.
symbiotic design whereby user behaviour is made intelligible to machines and machines are made intelligible to users
After a brief introduction (including a summary of the history of AI development and a comment about the Google Autocomplete results for
multiple sclerosis, not
Microsoft). There’s a detailed writeup here, but in short she explained that the current MS diagnosis tool—the EDSS scale—is
coarse-grained and subjective. A solution was desired that would be more consistent in its judgement of disease severity, as well as allow finer-grained ratings.
The first step, and tying in to the first idea of needing
user behaviour [to be] made intelligible to machines, was ethnographic research into how neurologists currently conduct the tests. It became apparent that despite the standardisation of the test, there was still a lot of workflow variation between practitioners which produced a range of effects, e.g. getting the patient to perform a Romberg’s test before the finger-to-nose test can make them perform the latter ventrally, rather than laterally. This is no issue for a human neurologist, but throws up issues for a machine learning algorithm. She also mentioned the rather clever idea of putting the neurologist*rsquo;s screen on the back of the unit, encouraging them to stand behind it rather than blocking the camera.
Tying into the requirement that
machines are made intelligible to users was the second half of the case study, focusing on how best to represent the data to the neurologists. After trying heatmaps and a number of other methods of data display, the consensus of preference was for the original videos as the most important tool for the neurologists to analyse, with the slew of other visualisations viewed as helpful tools to assist the expert diagnosis.
The second case study was less interesting — a consumer app that allows the user to take a photograph of objects with their phone camera and have the focus image segmented out and stored in a kind of
object scrapbook. Users could then tag images and organise them into collections, with the standard Microsoft app demo video (with soft folksy backing music, natch) showing collections for things like
action figures and
Mary’s art. Seconds after I wrote in my notes that it was basically Pinterest, Sellen brought up that many people erroneously think the app is basically Pinterest. However, her argument that Pinterest is for collating existing web content whilst the app in question was for collating physical objects didn’t hold water for me — surely the web content on Pinterest was originally physical objects too? I don’t think the addition of image segmentation is a big enough difference, nor do I particularly like the thought of physical objects—that last bastion of sentiment and meaning—being reduced to content as ephemeral as a Facebook photo.
However, as someone who quite likes using social cataloguing sites (e.g. Goodreads) to keep track of different forms of media, I can see a possible appeal for collectors of things like action figures, especially with the addition of backdrops for objects, including podiums and a bookshelf. Sellen also suggested the real reason Microsoft is pursuing the project when she said that it generated a large, semantically meaningful dataset that can be used to train other machine learning algorithms within the company. There was also a great point by an anonymous user about the fallibility of the algorithms:
For example, I added my Princess Peach figurine today and the suggested tags werepaper towel,diaperandteddy.
The third, shortest case study was even less convincing — a concept video for a The Lord of the Rings Top Trumps app where you take a photo of a friend and the app determines which character they most resemble and then creates that card in your deck with their face superimposed over the characters. After that, it’s just standard Top Trumps. It looked pointless.
To round off, Sellen elaborated on the four steps that users must understand in order to fulfil the goal of
design[ing] systems that work in partnership with users: sensing, or users understanding what a system is capable of seeing (e.g. the screen on the back of the Assess MS unit); perceiving, or users understanding how a system perceives forms and patterns in data (e.g. a shiny chair messing with a depth view whilst other clutter was safely ignored, whereas a human would ignore all the clutter); recognition, or users understanding how the system labels forms and patterns (e.g. the consumer app offering a suggested tag of ‘hermit crab’ for a scrunched-up bit of paper that vaguely resembled one); and acting, or users understanding the rules by which the system responds to data, which engenders trust. The recognition part came with the aphorism that it doesn’t matter if the system isn’t always correct, as long as users can understand why it was wrong.
The followup questions got me thinking though. Dr Mark Rouncefield talked about how all the old critiques of AI have not been answered, they’ve merely been ignored and the focus of the field moved to that of human intelligence in said. Whilst by the end of it he had forgotten to ask a question,
AI is dead, it died a long time ago was certainly a characteristically polemical point. Sellen riposted that it wasn’t dead, it had merely developed to using new techniques, such as neural nets. Another academic suggested the AI didn’t mimic human intelligence perfectly, and that was its strength —
the machine finds connections that we don’t spot.
A final point was made from someone in the psychology department that evolutionary psychology models the human mind as a computer, and Sellen added that evolutionary biology and genetic algorithms do much the same blurring of disciplines. She concluded with the driving thesis of symbiotic computing — using machine learning to produce consistent judgement over similar cases, whilst keeping a human around for a creative element, e.g. an MS patient with one arm.
On the downside, she said paradigm shift a lot.