SDS 533: Fusion Energy, Cancer Proteomics, and Massive-Scale Machine Vision — with Dr. Brett Tully

Dr Brett Tully is an accidental Machine Learning Engineer having walked a varied path across 4 distinct scientific areas, held together by a common thread of numerical science at scale. Brett completed a DPhil (PhD) in Biomedical Engineering as a Clarendon Scholar at the University of Oxford studying the interaction of fluid and tissue in the brain. He subsequently joined a fusion power start-up as Head of Simulation, was both Group Leader of Software Engineering and General Manger in a project mapping the proteome of human cancer, and now leads a team of data scientists and machine learning engineers running petabyte-scale geospatial AI.

Jon and Brett had known each other for some time as students and professionals. Brett is not the first Nearmap employee that’s been on the show though this is the first time the podcast has gone in-depth. Brett describes Nearmap as an aerial image company based around IP for cameras that allowed planes to go higher and faster to drop the cost of capturing images. Today, Nearmap takes requests from clients who request specific locations to be captured on images. Nearmap also owns the stack from turning images into a tiled map and they also build massive scale 3D representations of the world to dig into the detail of what the human urban experience looks like.

Brett’s role is in AI output systems which means he works with the deep learning input that comes out of R&D. They have massive amounts of data of images and continue to capture it at a high rate. The AI platform is run across every piece of imagery captured, so Brett’s team paints pixels with meaning. They have a collection of attributes that each pixel can represent. They’re focused on accuracy and aesthetic equally, as many customers are very concerned with the aesthetic as much as the information. They both increase the data and make it richer as they make inferences about the material in a picture, the area within a shape, geospatial information, and so on. One case study Brett went into was the company’s work in multiple vertical works—local government, insurance clients, among others. Brett notes the level of sophistication of teams from local governments looking at locations and land. They model population growth, predict the best school locations, how vegetation changes over time, and other topics. The beauty of the work is it can scale through technology with minimal human effort.

As far as the day-to-day, Brett realized early that he needed to keep technical work as part of his day job. Getting deep into technical problems is how he “stays sane” and achieves a sort of “flow state” in his work. The team has shared ownership over projects that might find Brett researching or simply tackling bugs on the backlog. He utilizes a lot of GeoPandas, a geospatial library in Pandas. The other aspect of his job is the technical strategy and planning 6 months to 2 years out in addition to ensuring his team is happy and functioning well. In terms of hires, Brett looks for people who know how to code in a team environment who can help the team run prototypes and embed work into the massively scalable system—which usually results in something breaking. So, adaptability in coding work and understanding potential pitfalls, and responding are key to succeeding on Brett’s team.

In addition to his work now at Nearmap, Brett has a rich history as a researcher and academic in biomedical work. He studied hydrocephalus which is an overabundance of spinal fluid in the brain, which shows up in children with congenital conditions. What Bret specifically studied was normal pressure hydrocephalus which presents in older folks and often looks like dementia. The good news is, unlike dementia, this can be reversed and corrected. So determining the diagnosis is important. He utilizes models to look at how fluid should flow through the brain by taking formulas originally designed for fluid movement in the gas and oil industry and applying them to this work. This is an example where data and formulas become important since actually experimenting would obviously be unethical. But Brett’s research helped to understand the physics of the problem as well as the physics of the treatment. Another very interesting topic he worked on was nuclear fusion at First Light. Brett got involved when one of his lab mates got involved in fusion through their work in fluids simulation and Brett joined the fledgling fusion company as the very first employee when the founders had the idea to look at collapsing bubbles that can result in fusion reactions. First Light Fusion since then has raised $62 million for their research. As for the future of fusion research, Brett thinks the achievement of fusion is inevitable at this point though he won’t make a prediction on exactly when we’ll achieve it. He thinks fusion is the key to maintaining society’s function, as we understand it now, into the future.

Brett also discussed the interesting application of data science he’s done with the ProCan project which seeks to tackle and understand the cause of cancer and how a collection of data around genomes and proteins can become a data problem. He discussed examples of creating cohorts from the data sets and looking at their cancer types and potential treatments. One fundamental interest of the program was childhood cancer and understanding childhood cancer by studying adult cancer. Childhood cancer is rare and often very fatal, which makes gathering data difficult. So ProCan studies protein structures in more general forms of cancer to understand the causes of and treatments for childhood cancer.

Throughout his career, Brett has been utilizing large-scale data that has been made practical and then applied the ideas to study numerical problems of immense difficulty across multiple leading-edge topics.