This recent NY Times article about crowd estimates for Hong Kong's annual pro-democracy protest is a good follow-up to my post about the number of people who, allegedly, showed up to last month's NBA Championship parade here in Toronto.

For years, Hong Kong has been seeing divergent estimates for its annual protest. Organizers typically overstate. And the police typically understate. This year, organizers claimed 550,000 people in attendance, whereas the police claimed only 190,000.

The difference this year is that a local tech company has started using AI software (loaded up onto iPads) to help supplement the standard practice manual counts. This year they concluded -- perhaps more definitively -- that 265,000 people protested in the streets of Hong Kong.

Image: NY Times

This is an interesting study from a team of AI researchers at Stanford. What they did was use car images taken directly from Google Street View (so images of cars parked on-street) to predict income levels, racial makeup, educational attainment, and voting patterns at the zip code and precinct level.

Admittedly, it’s not a perfect survey, but when they compared their findings to actual or previously collected data (such as from the American Community Survey), it turns out that their study was actually remarkably accurate. Google Street View allowed them to survey 22 million cars, or about 8% of all cars in the US.

Here are some of the things they found:

-  Toyota and Honda vehicles are strongly associated with Asian neighborhoods.

- Buick, Oldsmobile, and Chrysler vehicles are strongly associated with black neighborhoods.

- Pickup trucks, Volkswagens and Aston Martins are strongly associated with white neighborhoods.

Interestingly enough, the ratio of pickup trucks to sedans, alone, is a pretty reliable indicator of voting patterns. If a neighborhood has more pickup trucks than sedans, there’s an 82% chance it voted Republican in the last election.

Perhaps this isn’t all that surprising given that car purchases are highly symbolic. But given that the American Community Survey costs $250 million a year to administer, this study is a good preview of what cheaper and more realtime data collection might look like.

I should read The New Yorker more often. I’m going to subscribe right now (done). The articles are fantastic. In the April 3, 2017 issue, Siddartha Mukherjee wrote a piece talking about automated medicine. It’s called A.I. Versus M.D.

Here is an excerpt:

His prognosis for the future of automated medicine is based on a simple principle: “Take any old classification problem where you have a lot of data, and it’s going to be solved by deep learning. There’s going to be thousands of applications of deep learning.” He wants to use learning algorithms to read X-rays, CT scans, and MRIs of every variety—and that’s just what he considers the near-term prospects. In the future, he said, “learning algorithms will make pathological diagnoses.” They might read Pap smears, listen to heart sounds, or predict relapses in psychiatric patients.

This is a world where you could snap a photo of your unsightly rash, submit it, and then have a machine provide a diagnosis. This is a world where early detection and prevention would become common place. Already, today, there are instances where learning-based algorithms (distinct from rule-based) have proven to be more accurate than raw humans.

I don’t know how you all feel about this, but I actually feel a sense of comfort thinking about this as the future of medicine. There’s piece of mind that comes with large amounts of data and the discipline of algorithms. At the same time, this is not necessarily about replacing humans. It is also about augmenting human ability.

If it leads to better health outcomes, then why wouldn’t we?