The Fundamental Flaw of Deep Learning

In the past decade, AI has come to dominate multiple parts of digital media. Search engines such as Google use complex algorithms to translate text into thousands of different languages or to compile millions of search results for a given topic while social media services such as Facebook track people and places across an individual’s photos using enhanced automated software. Much of these advancements have come from a technique known as deep learning: the process whereby computers use layered (hence the term “deep”) networks to statistically classify patterns. Relying on a massive library of defined examples, AI can recognize correlating material to accomplish a given task. Silicon Valley heralds deep learning as the superpower of modern computing. But according to a recent article by Jason Pontin in Wired, deep learning may not be as incredible as the tech industry believes.

Pontin argues that deep learning fails to excel at any activity outside of classification. If placed in the real world, modern AI would flounder amidst the numerous situations for which it cannot classify or identify patterns. Everyday language features infinite variations on wordplay that in themselves contain hundreds of thousands of variations based on context. A computer would take years to understand and decode a day’s worth of human interaction. Since so much of human cognition operates beyond classification, Pontin believes that AI are doomed to fail in the short run if the tech industry continues to rely on deep learning as the answer to all problems. Unless computers can learn to improvise, that is, navigate situations without relying on training data, the odds of successful AI systems mirroring human behavior on a daily basis remains a distant hope.