News Feature: What are the limits of deep learning? | PNAS
In addition to its vulnerability to spoofing, for example, there is its gross inefficiency. “For a child to learn to recognize a cow,” says Hinton, “it’s not like their mother needs to say ‘cow’ 10,000 times”—a number that’s often required for deep-learning systems. Humans generally learn new concepts from just one or two examples.
There is a nice review on Deep Learning in PNAS. The spoofing referred to, is an ‘adversarial patch’ — a patch comprising an image of something else. In the example here, a mini-image of a toaster confuses the AI such that a very large banana is seen as a toaster (the paper is here on arXiv — an image is worth more than a thousand of my words).
Hinton, one of the giants of this field, is of course referring to Plato’s problem: how can we know so much given so little (input). From the dermatology perspective, the humans may still be smarter than the current machines in the real world, but pace Hinton our training sets need not be so large. But they do need to be a lot larger than n=2. The great achievement of the 19th century clinician masters was to be able to create concepts that gathered together disparate appearances, under one ‘concept’. Remember the mantra: there is no one-to-one correspondence between diagnosis and appearance. The second problem with humans is that they need continued (and structured) practice: the natural state of clinical skills is to get worse in the absence of continued reinforcement. Entropy rules.
Will things change? Yes, but radiology will fall first, then ‘lesions’ (tumours), and then rashes — the latter I suspect after entropy has had its way with me.