[Dprglist] fooling AI

[John Swindle]

 Dave,

The Q&A at the AT&T/IBM Watson hackathon presentation at DPRG a couple months ago showed that the weights in the neural nets have no meaning to people. So, though we think we can grade an AI system, we cannot diagnose it. We can only correct it when it malfunctions. We think that a low rate of false passes and false fails means success. But when we try to diagnose why there were false fails and false passes, we have no clue because the neural net weights have no meaning. We can "fix" the problem by massaging the weights, but we have no freaking clue about what those change have on other results. Indeed, we are not even supposed to know!

Over lunch at a place where they do this sort of thing, my student said, "but people make mistakes, too." It's a very cogent argument, and it stopped me then.

But now I have a story, and it is the story. A purpose-written program (non-AI) has a rationale, a story. If it fails, we understand WHY it fails, because the story was wrong. We understand it. We re-write it.

Does our story give a better result than the opaque weights of a neural net? I don't think so. But I think we can more easily control and CORRECT a story that we understand, as opposed to random weights that just happened to deconvolve a problem that day.

Separately,

I have a completely different proposal from what any of the car companies are doing:

Instead of special-casing hundreds of thousands of traffic conditions, I propose self-driving cars that:

Drive like India, Mexico, and Rome, and dirt farms in Arkansas where I grew up:

1) Stay out of the ditches

and

2) Don't hit anything.

If a lane is available that will not cause the vehicle to become stuck and does not ruin a neighbor's lawn, drive on. Otherwise, stop.

That's what I was taught. The rest is law and courtesy, which I was also taught. But for a vehicle, the rules could be dirt simple.

Stay out of the ditch. Don't hit anything.

Best to y'all.

John Swindle


 

 

-----Original Message-----
From: David Anderson <davida at smu.edu>
To: DPRG <dprglist at dprg.org>
Sent: Sat, Apr 15, 2017 3:02 pm
Subject: [Dprglist] fooling AI

timely:http://www.bbc.com/future/story/20170410-how-to-fool-artificial-intelligence<http://www.bbc.com/future/story/20170410-how-to-fool-artificial-intelligence>cheers,dpa_______________________________________________DPRGlist mailing listDPRGlist at lists.dprg.orghttp://lists.dprg.org/listinfo.cgi/dprglist-dprg.org
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.dprg.org/pipermail/dprglist-dprg.org/attachments/20170419/8095e660/attachment.htm>