What Tesla and Medical Innovation have in Common
Artificial Intelligence (AI) is everywhere!
Robots are either stealing jobs and launching SkyNet 🙀 or solving every looming catastrophe and ushering in the Singularity Nirvana.
Which one is it internet?!?
Deep Learning
For all the hype, one AI approach in particular (Deep Learning) is driving the excitement behind many of the world’s most innovative products, perhaps the most famous of which is Tesla and their ardent pursuit of self-driving cars.
Deep learning generates a model that processes streams of almost any type of data at a fast pace. In Tesla’s case, deep learning processes images of the road and laser radar data, then outputs driving instructions.
Deep Learning is a contemporary term for neural networks that typically run on commodity hardware like GPUs.
The technique allows for arbitrary combinations of convolutional¹ and recurrent² networks that consume data, typically labeled by humans. Tesla’s head of AI recently explained that the Autopilot engineers are stepping back from designing software and instead utilizing machine learning to create the programs.
“Now that the neural net is slowly taking over the code of Tesla’s Autopilot, Karpathy says that the team is focusing on labeling and creating the dataset infrastructure.” — Electrek
At their core, vehicle vision and X-ray reading algorithms run on the same technology, however their “intelligence” will only be as robust as the data set used to train the model. When daily use presents a neural network with a scenario it was not trained on, it may generate commands that are unintuitive and hard to categorize; that makes applying frameworks for risk classification a challenge.
Deep learning is also driving one of the hottest segments of medical technology; the automation of radiological diagnosis (X-rays, MRI, PET, etc.). Almost weekly new headlines report academic advances in detection of cancer or pneumonia. Affordable technology and the availability of high-quality data sets from agencies like the NIH have laid the way for a creative explosion in this field.
The researchers claim the resulting models rival or beat trained human radiologists. The models perform so well that it encourages both large companies and cash-strapped startups to bet heavily on using it in their products (again, like self-driving cars).
The Role of the FDA
However, in the United States between medical innovation and the marketplace stands the Food and Drug Administration (FDA). To many of us, the idea of software that is non-deterministic, meaning that we can’t explain precisely how it works, is new and a bit disconcerting. The FDA shares this apprehension. In doing their duty, the FDA has questions about the risks associated with this new type of software.
Regulators are concerned not with just knowing that the software works, but also how it works.
Moreover, some of these neural networks-based products intend to continuously-learn and update while in use. This idea of a product that changes constantly, without additional regulatory review, is entirely foreign to the FDA. It’s all well and good if AI models are always better, what happens if they aren’t? How do we predict that?
For Tesla and other autonomous vehicles, regulation is emerging at the state level where most on-road testing occurs (California, Arizona, etc.), and in the case of a crash, sometimes prompts a federal level investigation by the NTSB. The FDA, however, has been long established and is evolving from legacy perspectives on technology to the world of AI.
The Way Forward
Self-driving cars and clinical AI will find a way to make use of deep learning’s magic. For companies of both types to succeed in the long run, they’ll also have to find a way to work well with regulators and account for hard-to-spot risks to ensure the public gets the best product possible while keeping them predictably safe.
Footnotes:
Âą recursive calls that model animal neurology, ideal for image processing
² chronologically spaced passes through a node ideal for text processing
