Ken Washington, Senior Vice President and Chief Technology and Innovation Officer, Medtronic [Photo courtesy of Medtronic]
Ken Washington, Medtronic's senior vice president and chief innovation officer, recently briefed the 15 general managers who run the sales divisions of the world's largest medical device maker.
Midway through the first diagram of his presentation on artificial intelligence, one of the leaders stopped him.
They reportedly said: “I just don't understand all the buzzwords about AI. Can you tell me what types of AI there are? How does it all work? And the difference between Generative AI and Deep Learning What is it?”
Mr. Washington joined Medtronic in June 2023 after serving as vice president and general manager of consumer robotics at Amazon and CTO at Ford Motor Company. He pulled out his easel, grabbed a marker, and explained the basics to the group.
In an interview with Medical design and outsourcing, Washington provided an abbreviated version of that tutorial. The following quotes have been lightly edited for space and clarity.
“What we have to do as leaders is to help business leaders, people who don't practice this every day, people like me who aren't in a position to run robot programs, or people who haven't done something before. What Ha Hong, Medtronic Endoscopy's chief AI officer, has done is explain what these terms mean, why they're important, and how they've evolved. what's good and what's bad, what's hype and what's reality, and what makes us more fluent. And you can apply this in important strategic ways.
“The term AI was first coined in 1956. It is the mathematics and science of emulating human decision-making on a computer. And its first iteration was in a rule-based system. There are some logic machines specifically designed to do that, but none of them are very good; they don't solve any appreciably large problems, and they're not very general-purpose either. There wasn't, and the field stagnated for a long time.
“And then machine learning came along rapidly. That's the first specialization that happened. It's the science of teaching machines how to make decisions by creating digital neurons and connecting them with weights. The weights are actually established by teaching the neurons how to interact with each other by feeding them data; the more data you feed them, the better the weights become. You'll be able to interact with it, take input, and actually provide meaningful output. And it worked pretty well.
“But until Jeffrey Hinton came along, the number of layers was limited because it was too computationally expensive to actually assign weights and run the algorithm to get the output.” Followers have figured out how to add various additional layers to these networks, creating something called deep neural networks or deep learning, which allows these weights to be trained on large amounts of data. We were able to build models with many layers and many digital neurons connected by . They are equipped with all kinds of algorithms and techniques to perform backpropagation and forwardpropagation. At the same time, the amount of data available to train these networks has exploded exponentially. has gotten much faster, and graphics processing units (GPUs) have gotten a lot better, so NVIDIA has thrown in a lot of big, dedicated GPUs to do these calculations. , and these new algorithms for performing deep learning forward and backward propagation, have made deep learning actually possible for the first time. They have also become so good that in many cases they can now perform almost as well as humans at image recognition, speech recognition, and other problem-solving.
“Then the smart people at Google came up with these Transformer models, which allowed us to build not just deep models, but large multi-billion parameter models trained on large datasets. And that's what created the underlying model and the generative AI that not only does image recognition and speech recognition very well, but also creates new content that looks, sounds, and looks real based on what it's learned in the past. Generative AI is a very special branch of deep learning that uses large underlying models that need to be trained on massively parallel supercomputers that major technology companies have access to. It came onto the scene around 2017, and that's why it exploded in scale and hype and excitement last November. Because they packaged it into a consumer tool and put it on the web: ChatGPT and the rest is history.
“The sweet spot for medical technology is deep learning. The information has to be enough to help clinicians make decisions and get information and advice. Deep learning is very good at this. And we don't need generative AI. In fact, generative AI has all sorts of side effects that aren't very suitable for medical technology, such as hallucinations.
“Although AI technology has been around for a long time, the modern version of AI is that it is even AI solutions built on top of these large underlying models and deep neural networks with very sophisticated algorithms. We have to remember that the world's most complex cloud supercomputers and some on-premises supercomputers have only been used to train networks for a year or two. Therefore, we should spend time and effort to educate ourselves and teach others about the importance of these technologies.
“The future is bright, and we are just getting started.”
Read more about Washington: How Medtronic uses AI: Artificial Intelligence Insights and Advice
