Rachel Thomas on Making AI Accessible
How fast.ai is disrupting the learning of AI deep learning
Posted Nov 08, 2018
Artificial intelligence (AI) has emerged from a deep winter of relative dormancy for sixty years into a renaissance in full swing – largely due to advances in deep learning. Based loosely on the human brain, deep learning is a machine learning method that uses layers of artificial neurons (an artificial neural network) that does not require explicit programming in order to “learn” from vast amounts of Big Data input. There is a hiring boom for artificial intelligence (AI) professionals. The share of jobs requiring AI skills has grown 4.5 times in the U.S. during the period 2013-2017, according to Stanford’s The AI Index. To address this growing demand, one company is taking an innovative approach. At Exponential Medicine conference this week, Rachel Thomas, Co-Founder of fast.ai, presented a novel way to make AI accessible to a wide range of people from all backgrounds, not just from the elite institutions – in effect, disrupting the learning of AI deep learning.
Rachel Thomas is one of Forbes “20 Incredible Women in AI,” a TEDx San Francisco speaker, a professor at the University of San Francisco (USF), a faculty member of Exponential Medicine, a popular writer, and an influential keynote speaker. She earned her Ph.D. in math from Duke University, and was one of the early engineers at Uber, among other startup companies. Thomas co-founded fast.ai in 2016 with serial entrepreneur Jeremy Howard with the strategic intent to make learning deep learning accessible.
“When we were creating the course, this was something that I wished had existed five years ago when I was first getting interested in deep learning,” said Thomas.
Traditionally, there are many barriers for coders to acquire enough deep learning skills to produce state-of-the-art algorithms that solve real-world problems. Many existing institutions require a high-level mathematical background or Ph.D. as a requisite, which may take years to obtain. If students do eventually create a working algorithm, it’s usually hypothetical with no real-world application.
“I think many courses are either pretty theoretical, and that make sense … deep learning is growing out of a theoretical field,” she said.
The fast.ai founders had identified a gap in technology education and found a way to modernize it. The traditional approach to teaching deep learning is typically a long, slow process starting at the detailed technical level – a bottom-up approach.
“We wanted something more practical and hand’s on,” said Thomas, “I really wanted to make this accessible to more people, and make it easier for people of all backgrounds and domains to get involved in this field.”
“I am much more interested in what works to solve problems that people are having – whether that’s agriculture, medicine, or manufacturing,” said Thomas.
Her company's methodology is a top-down approach, the exact opposite of traditional teaching methods for deep learning. Students can quickly develop deep learning algorithms with an open-source library of ready-to-use applications and models that Thomas helped to create.
“We want to get people training neural networks right away even though people don’t know the underlying components overall,” said Thomas.
In a very short period of time, fast.ai students are able to rapidly produce high-performance state-of-the-art deep learning algorithms without having to have advanced math prerequisites.
“Over time we’ll get to the details, and if you do the whole course you’ll get a low-level understanding, but it’s just totally flipped in the order,” Thomas said.
“In my background, I have a Ph.D. in math and worked as a software engineer, and data scientist — I could see how powerful this technology is, and that we’re just on the cusp of it,” said Thomas.
As a current research-in-residence at the University of San Francisco (USF) Data Institute, Thomas’ primary focus is on the performance of deep learning algorithms.
Thomas said, “My primary interest is, ‘Do these algorithms work? Are they solving interesting problems and giving accurate results?’
According to Thomas, fast.ai’s long-term vision is to keep making deep learning technology easier to use, to yield even better results. That means increasing the content of fast.ai’s open source library.
In less than two years since fast.ai was founded, over 200,000 students have completed its online course, and several hundred more have taken the in-person classroom curriculum. Fast.ai is blazing a trail to enable coders to become deep learning experts in weeks, versus years — a rapid, practical approach towards solving real-world problems.
Copyright © 2018 Cami Rosso All rights reserved.
Shoham, Yoav; Perrault, Raymond; Brynjolfsson, Erik; Clark, Jack; LeGassick, Calvin. “Artificial Intelligence Index 2017 Annual Report.” The AI Index. Retrieved 11-8-2018 from http://cdn.aiindex.org/2017-report.pdf