Deep Learning: Modular in Theory, Inflexible in Practice
The high-level view of deep learning is elegant: composing differentiable components together trained in an end-to-end fashion. In reality, it isn't that simple, and the commonly used tools greatly limit the kinds of architectures one can create. In this talk, we will discuss some of these limitations, what we can do about it, and how those things can result in much better models in practice. We will show how we've used these approaches to deal with some of the technical challenges of analyzing medical images, such as their large size (150MB+ per scan) and small features (medical anomalies can be 3 pixels or less).
Jeremy Howard is a serial entrepreneur, business strategist, developer, and educator. He is the CEO of Enlitic, a startup he founded to use recent advances in machine learning to transform the practice of medicine, and bring modern medical diagnostics to billions of people in the developing world for the first time. He is the youngest faculty member at Singularity University, where he teaches data science, and is also a Young Global Leader with the World Economic Forum. Previously he was the President and Chief Scientist of Kaggle, a community and competition platform for over 150,000 data scientists. Before working at Kaggle, he was the top ranked participant in data science competitions globally, in 2010 and 2011. He founded two successful Australian startups (the email provider FastMail, and the insurance pricing algorithm company Optimal Decisions Group), both of which grew internationally and were sold to large international companies. He started his career in management consulting, working at the world's most exclusive firms, including McKinsey & Co, and AT Kearney (becoming the youngest engagement manager world-wide, and building a new global practice in what is now called "Big Data"). He is also a keen student, for example developing a new system for learning Chinese, which he used to develop usable Chinese language skills in just one year. Jeremy has mentored and advised many startups, and is also an angel investor. He has contributed to a range of open source projects as a developer, and was a regular expert guest on Australia's most popular TV morning news program.