Douglas Fisher

This fall I’ll be “teaching” Vanderbilt’s undergraduate course in artificial intelligence (AI) at a physical distance — from Arlington, VA. It’s an opportunity to shatter the way I’ve taught AI in the past. As uncomfortable as giving up the ol’ PowerPoint presentations are, and as creative an outlet as it was to prepare and still is to refine/adapt those slides (I’m absolutely serious — those slides are sometimes art — I’m seriouuuuuuus!), I’ve fallen back on them too much — they’ve encouraged an automaticity of teaching and a psychological distance from students that is not healthy. Gutsy was/is the professor that attempts a proof or writes a program for the first time (perhaps, in a long time) on a board in front of a class. And so long as the class recognizes that they are witnessing a search for an answer rather than an answer, its a-OK: “I’m about to demonstrate what it is to search for a proof, with some commentary, I hope, about how this search is systematic in advancing towards an answer, and if I/we don’t find the proof today, you and I can race to see who can post a valid proof by next week.” It’s scary, quite frankly. I think its possible that a fair number of professors got into academia, in part, because they saw a safe psychological distance modeled in the classrooms in which they were students.

Prefabricated slides also add friction that opposes mobility — if my slides are synced to a textbook, for example, it’s harder to switch textbooks if I feel obligated to prepare yet another set of slides, or even to adapt my current slides to the new text.

When I started this post, I really wasn’t intending to get into the above, but rather launch directly into some ideas that are arising as I think about the AI course as critical thinking “lab” — as we study ways of designing ‘intelligent’ computers, we reflect on our own thinking — for inspiration, comparison, and validation. My study of AI, and in fact computer science generally, has certainly shaped the way I think about my thinking and the thinking of others, but frankly its shaped my thinking itself — I’m quite certain that I think differently than I would have if I had gone to Humboldt State to study forestry, though probably not radically different.

As critical thinking lab I want to reflect on everyday scenarios (and not so everyday scenarios too). Not long ago my sister posted a status report on Facebook on how eerie it was that Facebook was making friend suggestions that were spot on — that is, she knew them, but she had no idea how Facebook would know that! This is certainly the case with me. There are, of course, friend suggestions that seem no brainers — e.g., I and suggested person X have 2 friends in common. Sometimes these no brainer suggestions have a plausibility, but in other cases they are not terribly suggestions, and Facebook has enough information to infer this — if the two common friends are married and there is no other substantive link between me and the suggested person that I can see, this suggestion loses credibility as someone I might reflect on, and if Facebook knows of the marriage link between the two common friends, you would think that their algorithms could be written to take this into account (and perhaps the algorithms do!?).

The mystery of course is in the suggestions on which I am clueless on how Facebook made a connection — do they have some kind of crystal ball? Two clues as to what Facebook could do if they chose to do so appear in the Facebook margins itself. One is this app that keeps popping up — “Find out who is searching for you!”, with a very good-looking young woman smiling out at me ). Who knows, maybe the app is a scam, but is does suggest that who I search for and who’s profile I am clicking on (including perhaps through a Google search result) is recorded, and why wouldn’t it be!? It also seems clear that the ads that pop up in the margins are customized to me in some fashion (e.g., “Eating out in DC?” “Class of 1975 memories”). So, why couldn’t friend suggestions take into account who/what I’ve searched for and who’s (partal) profile I’ve gone too? — I just went to the “Glendora High School Class of 1975″ site and clicked on a couple of non-friend people (”Is this Betty Friedland the Betty Moorehead that I had a crush on in first grade??!!”) Now, it seems unlikely that having looked at Betty’s partial profile and having not friended her that she would be suggested as a friend for me (but I’ll let you know). But is Betty Friedland going to see my name as a suggested friend? Or more generally will that I looked at Betty Friedland be a factor that’s taken into account in friend suggestions to Betty? I don’t know, but I’d like my students to be asking themselves these kinds of questions on everyday kinds of scenarios like this. Exercises like this (a) cast a light on my own thinking, (b) suggest what others can infer and what can be automated (because whatever I can infer from Facebook data can be inferred by others and by properly designed automated methods), and (c) beg some interesting ethical discussions concerning AI (and in this case, online social networks). In the context of AI there are a lot of mundane things that one can infer about me — e.g., I’m not part of the Washington DC network, but given my photo gallery and many of my friends, it’s not hard (for a human) to realize where I am living right now, though to program a computer to take these modest inferential steps would be quite interesting. And there are even more ambitious inferential leaps that can be made, by human and (thus) ultimately by computer — there are all kinds of research going on into what can be inferred from social networks, and exploring designs that guard against those inferences.

At age 52, I am a lot less concerned with my own privacy than (I think) I used to be — life is short as I’ve been aware of lately, and while I am much more self-revealing than I once was, I keep a low profile on politics and religion, at least so long as I’m in this government gig, though you could dig, but I don’t think a simple story would be revealed (though the simplicity of the story often has more to do with the “reader” than the “author”). I think that I might prefer automated inference methods that include explicit representation of uncertainty to human inference methods that sometimes (incorrectly) dismiss the uncertainty — of late uncertainty has been much more a friend than an enemy. To return to my original thread, however, I see online social networks as a treasure trove for AI class discussions. Music and environment are others. More to come, I expect.