Deep Learning Accessibility, Programming Languages, and AI in Medicine with Jeremy Howard

Jeremy Howard, founder of fast.ai, discusses his journey through programming languages and the mission of fast.ai to democratize deep learning. He highlights the critical need for accessible, practical deep learning education, contrasting fast.ai's approach with the "BS" often found in popular educational content. Howard's early programming experiences, including a Commodore 64 BASIC program to analyze musical scales, reveal a lifelong interest in applying computational tools to solve problems and understand complex systems. This foundational drive for utility and problem-solving has guided his career, from building an email company with Perl to his current work in AI.

A significant portion of the discussion delves into the evolution and future of programming languages, particularly in the context of data science and machine learning. Howard expresses a deep appreciation for languages and environments that prioritize productivity, expressiveness, and hackability, such as Microsoft Access (VBA), Delphi, and the array-oriented languages APL and J. He critiques Python's limitations, particularly its slowness and "unhackable" C-level components, which hinder innovation in areas like recurrent neural networks and sparse convolutional neural networks. Howard advocates for Swift, led by Chris Lattner, as a potential future language that could offer the desired "infinitely hackable" environment, allowing researchers to deeply experiment and optimize algorithms from top to bottom.

The conversation also explores the challenges and opportunities in optimizing GPU programming for deep learning. Howard explains that writing fast GPU code is inherently complex due to low-level details like thread synchronization and memory management. He points to emerging compiler technologies like MLIR, Tensor Comprehensions, and Halide as crucial advancements that enable the creation of domain-specific languages for tensor computations, drastically reducing code complexity while maintaining performance. This approach, he argues, could abstract away the hardware-specific boilerplate, making GPU programming more accessible and fostering innovation beyond NVIDIA's current dominance, potentially supporting other hardware like AMD, Graphcore, and Vertex AI.

Finally, Howard shares the origin story of his previous startup, Inlytic, which focused on deep learning for medicine. Driven by the global shortage of doctors, particularly in developing nations, Inlytic aimed to leverage AI for diagnosis, treatment planning, and triage, thereby magnifying the productivity of existing healthcare workers rather than replacing them. He emphasizes that AI's role is to identify high-risk cases, allowing human experts to focus on complex problems. Howard discusses the significant hurdles to AI adoption in medicine, including the medical community's initial lack of awareness, regulatory complexities, and hospital lawyers' risk-averse interpretations of data privacy laws like HIPAA. He advocates for "doing more with less data" through techniques like transfer learning, challenging the industry's tendency to demand ever-larger datasets and computational resources.

"when someone asked me how do I get started with deep learning fast AI is one of the top places that point them to it's free it's easy to get started it's insightful and accessible and if I may say so it has very little BS"

"I've always been interested in doing useful things for myself and for others which generally means getting some data and doing something with it and putting it out there again so that's been my interest throughout"

"J is the most expressive composable language of you know beautifully designed language I've ever seen"

"this this this this path of programming languages it's just so much that are not so much more powerful in every way than the ones that almost anybody uses every day"

"I hope Swift is successful because the goal is Swift the way Chris Lattner describes it is to be infinitely hackable and that's what I want"

"at the moment our understanding of deep learning is incredibly primitive there's very little we understand most things don't work very well even though it works better than anything else out there"

"there's a about a 10x shortage of the number of doctors in the world and the developing world that we need expected it would take about three hundred years to train enough doctors to meet that gap"

"one of my areas of focus is on doing more with less data which so most vendors unfortunately are strongly incented to find ways to require more data and more computation"