These primitive correlations mean our models are extraordinarily brittle and offer little more than the cheap parlor tricks for which they are known. An image of a dog in a park is not “seen” as a four-legged fur-covered mammal standing in grass, but rather as an assortment of pixels of certain colors and textures in certain arrangements that are found more often in images labeled as “dog” than in other images. Worse, these correlations lack the ability to reason about their training data in terms of semantic relationships rather than merely syntactic co-occurrence. The reality is that these simplistic piles of correlations cannot reach beyond the confines of their training data. Much as a Pearson correlation might seem to the uninitiated a magical feat of “machine learning,” so too can collections of these correlations working together seem to represent something far greater, a silicon “intelligence” that is thinking and reasoning about the world. Instead, they merely encode the patterns of their training data as precisely as possible and attempt to match incoming data with the most similar pattern extracted from their training examples. 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.

Parlor Tricks by N. Kuhn