SAVE THE DATE! KMWORLD 2019 in Washington DC NOVEMBER 5 - 7, 2019


Perspective on knowledge: Tools, senses, and machine learning

<< back Page 2 of 2

The alien nature of machines

But sometimes their delight is in the impossible things machine learning gets right. For example, without any instruction in what words mean or how they go together, after analyzing a large body of text represented as sequences of randomly assigned numbers, a machine learning system can infer that a prince is to a king as a princess is to a queen. Or, by analyzing scans of retinas, a system is able to gauge the eyeball owner’s gender, age, and whether or not they smoke. In fact, it’s able to predict the owner’s risk of a heart attack with about the same accuracy as our prior methods. But, at least so far, no one—computer scientists or physicians—can figure out what about those scans led the machine learning system to its conclusions. When one of the developers I talked with described that system—not one that he was working on—he chuckled in delight.

That delight comes from the alien nature of these machines. We can argue about whether machine learning’s neural networks work the same way as our brains’ neural networks do, but if you’re trying to train a machine to give accurate results, you’re not thinking about how you would train a brain. You’re deep into analyzing the quirks (and far worse) of datasets, and about the math that will get a system to produce accurate results when tested and deployed. You’re trying to figure out why your image identifier thinks a 3D-printed turtle is a rifle, and, in a far worse case, why your job resume sorter is accepting way too few women.

Beyond comprehension

The answers are in the data and the math. Sometimes, the answer can’t be pinned down because the networks that the data and the math construct are just too complex. It’s frustrating when the system is giving wrong results, but when it’s giving correct results by means that we just can’t fathom, it’s a lesson in humility. 

This makes machine learning a different type of tool. Unlike a knife, wrench, or computer, it is not extending our senses and minds by magnifying our existing powers. Machine learning sometimes extends our cognitive abilities in ways that are alien to our minds. If a knife shows us a tomato as something that can be cleanly cut, machine learning shows us the world as too complicated and complex for us humans to grasp. It extends our senses, but in a way that is beyond our senses and sometimes beyond our comprehension.

<< back Page 2 of 2

Search KMWorld