• If you are citizen of an European Union member nation, you may not use this service unless you are at least 16 years old.

  • Finally, you can manage your Google Docs, uploads, and email attachments (plus Dropbox and Slack files) in one convenient place. Claim a free account, and in less than 2 minutes, Dokkio (from the makers of PBworks) can automatically organize your content for you.


Is AI Riding a One Trick Poney

Page history last edited by swanson@... 2 years, 11 months ago








Somers, J. (2017). Is AI Riding a One-Trick Pony?. MIT Technology Review, 120(6), 28-36.




"It’s only when you leave the room that you remember: these “deep learning” systems are still pretty dumb, in spite of how smart they sometimes seem. A computer that sees a picture of a pile of doughnuts piled up on a table and captions it, automatically, as “a pile of doughnuts piled on a table”

seems to understand the world; but when that same program sees a picture of a girl brushing her teeth and says “The boy is holding a baseball bat,” you realize how thin that understanding really is, if ever it was there at all. Neural nets are just thoughtless fuzzy pattern recognizers, and as useful as fuzzy

pattern recognizers can be—hence the rush to integrate them into just about every kind of software—they represent, at best, a limited brand of intelligence, one that is easily fooled. A deep neural net that recognizes images can be totally stymied when you change a single pixel, or add visual noise that’s

imperceptible to a human. Indeed, almost as often as we’re finding new ways to apply deep learning, we’re finding more

of its limits. Self-driving cars can fail to navigate conditions they’ve never seen before. Machines have trouble parsing sentences that demand common-sense understanding of how the world works.


"Deep learning in some ways mimics what goes on in the human brain, but only in a shallow way—which perhaps explains why its intelligence can sometimes seem so shallow. Indeed, backprop wasn’t discovered by probing deep into the brain, decoding thought itself; it grew out of models of how animals learn by trial and error in old classical-conditioning experiments. And most of the big leaps that came about as it developed didn’t involve some new insight about neuroscience;

they were technical improvements, reached by years of mathematics and engineering. What we know about intelligence is nothing against the vastness of what we still don’t know" (p.34) 


" Suppose you take your first training image, and it’s a picture of a piano. You convert the pixel intensities of the 100x100 picture into 10,000 numbers, one for each neuron in the bottom layer of the network. As the excitement spreads up the network according to the connection strengths between   neurons in adjacent layers, it’ll eventually end up in that last layer, the one with the two neurons that say whether there’s a hot dog in the picture. Since the picture is of a piano, ideally

the “hot dog” neuron should have a zero on it, while the “not hot dog” neuron should have a high number. But let’s say it doesn’t work out that way. Let’s say the network is wrong about this picture. Backprop is a procedure for rejiggering the strength of every connection in the network so as to fix the error for a given training example. The way it works is that you start with the last two neurons, and figure out just how wrong they were: how much of a difference is there between what the excitement numbers should have been and what they actually were? When that’s done, you take a look at each of the connections leading into those neurons—the ones in the next lower layer—and figure out their contribution to the error. You keep doing this until you’ve gone all the way to the first set of connections, at the very bottom of the network. At that point you know how much each individual connection contributed to the overall error, and in a final step, you change each of the weights in the direction that best reduces the error overall. The technique is called “backpropagation” because you are “propagating” errors back (or down) through the network, starting from the output. The incredible thing is that when you do this with millions or billions of images, the network starts to get pretty good at saying whether an image has a hot dog in it. And what’s even more remarkable is that the individual layers of these image recognition nets start being able to “see” images in sort of the same way our own visual system does" (p. 31-33)



Backprop = backpropagation


From Rumelheart, Hinton & Williams 1986


Paper at heart of AI 


A 2007 paper by Hinton: Learning multiple layers of representation by Geoffrey E. Hinton



Why we should be Deeply Suspicious of BackPropagation






Comments (0)

You don't have permission to comment on this page.