Dr. Eng Lim Goh on Swarm Learning and Steering HPC Simulation with AI

Print Friendly, PDF & Email

Dr. Eng Lim Goh from HPE

In this fireside chat from SC19, Dr. Eng Lim Goh from HPE describes how the convergence of HPC and AI is changing the way scientists and engineers do their simulations. He also cites a case study of “Swarm Learning” where hospitals were able to train AI diagnostic models without sharing private patient data.

Transcript:

insideHPC: Hi, I’m Rich with insight HPC. We’re here at SC 19 in Denver, Colorado and today I’m here with Dr. Eng Goh from HPE. Dr Goh, we’ve had these fireside chats for years now, but today I wanted to ask you about HPC and AI. It seems like they’re converging. Can you tell me some examples that might show how this works?

Eng Lim Goh: Yes, there are many levels of integration. I think one of the earliest kinds that we’ve seen applied in operation is in the HPC modeling simulation world where you are doing, for example, fluid dynamics simulation, right? A lot of times, you run your simulation and then eventually you realize that that’s the wrong way to go after quite a lot of time and sometimes all the way to completion? Then you come back, change your initial parameters, and run your simulation again. So we’ve seen an example where a customer of ours took all the old failures scenarios and all the successful scenarios to train a neural network. The idea is to recognize what looks like a successful outcome as the computation is progressing and what looks like it might not be a successful one. And then they can use the AI system to tell them when to stop the simulation early because it is likely not going to be successful. So this is one great example I believe in increasing the productivity of modern simulation HPC site using AI.

insideHPC: Because that’s an expensive resource to run on the supercomputer for hours if it was going down the wrong track, right?

Eng Lim Goh: So Imagine operations where they are restricted in the amount of modeling simulation time, like they are in the case of formula one racing cars. Here is a great example where there is a limited budget for modeling simulation. So they had to be very careful about what they do there. A lot of unsuccessful runs is very costly to them. So yes, by applying this AI system to stop simulation early, they can do more successful runs within their computational budget.

insideHPC: That’s a great example. I also wanted to ask you about things like healthcare and AI because it seems to me the data there is heavily regulated. How can you create a model based on this private information without exposing it to the world?

Eng Lim Goh: You get biased results. Yes. Let’s use a hospital as an example. I’m seeing lots of cases of where they are diagnosing chest condition with lots of chest x-, but they seldom see one particular kind, say for example, pneumonia toric. Okay. If they train a neural network that way, that hospital will be biased against pneumonia. torics because it’s not seen in many of those cases. In order for it to improve, it has to collect data from other hospitals that sees lots of those conditions and train with them. But then there is a problem, right? Privacy. Not just privacy because of HIPAA, but also privacy across a country sovereignty, right? Lots of privacy requirements.

How do you share your learnings across hospitals without sharing data? That was a reason we developed a technique called swarm learning. Basically what it does is every hospital does their own learning of their own data bias. And other hospitals pass their own learning bias a different way. But after each cycle of learning, we use the system to collect all of our learnings, the weights of the neural network, average them, and send it back down. Then they continue learning, collect your weights. We’ve done a test of three hospitals’ data and all three had the biases removed here. All three could detect conditions that they didn’t have individually. We ended up allowing sharing of the learnings without sharing data.

insideHPC: Well, that’s really compelling. I wish we had more time to talk, but thanks again for having me and I look forward to next time.

Dr. Eng Lim Goh is Senior Vice President and Chief Technology Officer for AI at HPE. As such, leads the development of the next generation computer architecture. His current research interest is in the progression from data intensive computing to analytics, inductive machine learning, deductive reasoning and artificial specific to general intelligence, human perception for user interfaces and virtual and augmented reality. In collaboration with NASA he is currently principal investigator of a year-long experiment, aboard the ISS to operate high performance computers for long duration space travel.

In 2005, InfoWorld named Dr. Goh one of the World’s 25 Most Influential CTOs. He was included twice in the HPCwire list of “People to Watch”; 2005 and 2015. He co-presented with NASA at the inaugural 1st plenary of the Supercomputing 2014 Conference to an audience of 2,500. Before joining SGI, Dr. Goh worked for Intergraph Systems, Schlumberger Wireline and Shell Research. A Shell Cambridge University Scholar, Dr. Goh completed his PhD research and dissertation on parallel architectures and computer graphics. Dr. Goh has been granted six U.S. patents, with three others pending.

See our complete coverage of SC19

Check out our insideHPC Events Calendar