For deep learning to truly take off we must move beyond these cheap parlor tricks towards systems that can do more than write a few simple correlations into code. As companies rush to deploy deep learning, they are building our modern world upon an extraordinarily fragile house of cards. Driverless cars can slam on the brakes or accelerate towards an obstacle with the slightest deviation from their training examples, while image understanding algorithms are rendered helpless by a few spurious pixels. Neural translation algorithms hyped as replacing humans in reality oscillate wildly between human fluency and indecipherable gibberish with the change of a single word. Rather than “learn” about the world, today’s algorithms merely encode databases of simplistic statistical correlations that yield results that are entirely dependent on how similar the inputs are to their training data. The problem is that in the real world this correlative house of cards comes rapidly crashing down. Given ideal circumstances, deep learning models can churn out what appear to be extraordinary feats of near-human or even super-human intelligence. Today’s deep learning shares much in common with the cheap parlor tricks of old.
0 Comments
Leave a Reply. |