Are radiologists obsolete?  Machine learning pioneer, Geoffrey Hinton made the startling suggestion that we should stop training radiologists.  Per Hinton, machines are going to surpass humans, in detecting and classifying tumors.   Within the five years that it takes to earn a Radiology degree, machine learning algorithms will be more reliable than humans.  Referring to how blissfully unaware they are to be disrupted, he likens radiologists to cartoon images of coyotes over the edge – unaware that the ground beneath their feet had receded.  Here is a link to the video of his comments, along with other luminaries like Yoshua Bengio, Rick Sutton and Ruslan Salakhutdinov (who recently took over as head of AI at Apple).

Naturally, medical professionals are disdainful.  And in fairness to them, radiologists do far more that merely classifying images.  Embolization’s and other invasive procedures, requiring manual dexterity, are also part of their repertoire.  Robots haven’t arrived at that level of precision…. yet.

Still, Hinton’s assertion is worth considering and got me thinking about other pattern recognition tasks that might be ripe for disruption. What about insurance claims processing, air traffic control and legal e-discovery?  Reaching back to my days as a Gait Lab engineer, pattern matching can even be used to determine the degree of correction required for orthopedic surgeries.  All I did, at the time, was to painstakingly track patient walking patters (using infra-red cameras and markers), and compare them to typical human gait.  We provided this data to orthopedic surgeons, who would rely on their years of experience to make surgical adjustments.  With a large enough training data set, why couldn’t an algorithm do the same job that I did? Not only that, the training data set would be larger than any single surgeon’s data set.  The algorithm could recommend the degree of correction required.  As patients, we may be unwilling to let machines entirely determine our fate.  Still, machine learning systems can definitely augment human judgement.

Deep learning is going to disrupt many multiple domains. Currently, much of the advancements are in the field of games, such as chess, checkers and Go.  Board games are a great petri dish for experimentation.  They help machines learn to make good decisions to maximize benefits in ambiguous environments.  Oh! and no one dies!!!  Watch this space as we continue to explore and experiment with deep learning.  We are super excited to expand our skills in this area and are eagerly anticipating a machine learning enabled future.