Learn neural nets in 7 weeks? That’s fast.ai

This picture shows Fast.ai co-founders Jeremy Howard (left) and Rachel Thomas. (Image credit: fast.ai)

Fast.ai co-founders Jeremy Howard (left) and Rachel Thomas. (Image credit: fast.ai)

What does it take to work in state-of-the-art deep learning? Conventional wisdom would have it that artificial intelligence is the most elite corner in the already very elite world of tech, and that AI practitioners must have an advanced degree from the likes of Stanford or Carnegie Mellon, as well as access to the world’s most advanced, AI-optimized computer systems.

On the other hand, you could launch your deep learning career by taking Practical Deep Learning for Coders, a free seven-week online course offered by fast.ai, the San Francisco-based nonprofit that wants to bring artificial intelligence to the people.

Founded in 2015 by Jeremy Howard and Rachel Thomas, fast.ai’s motto is “making neural nets uncool again” — because “cool equals exclusive,” as Thomas quipped in a recent TEDx Talk (above). By offering deep learning courses and tools for people with only very basic coding skills, Howard and Thomas are debunking the myth that AI is only for elites.  

“It’s extremely important to get a more diverse and representative group of people involved in building the technology that we’re all using,” Thomas told All Turtles in an interview.

Howard, a philosophy major with no graduate degree or formal ML training, is a self-taught data scientist and the founder of Enlitic, a deep learning medical diagnostics company. Thomas has a math Ph.D. from Duke and was formerly an engineer at Uber. Neither come from the academic world of AI and both are extremely passionate about its potential to address problems in medicine, agriculture, education and beyond. But for AI to reach its potential to help people, the pair believes that experts from all walks of life need to be learning the technology and pushing it forward. “We believe that this requires allowing domain experts to be able to leverage the technology themselves, rather than leaving it in the hands of a small and exclusive group of mathematicians,” Howard wrote in a company blog post about their company’s launch.  

More than 200,000 students have taken fast.ai's 7 week course, Practical Deep Learning For Coders, Part 1. (Image credit: Screenshot/fast.ai)

More than 200,000 students have taken fast.ai’s 7 week course, Practical Deep Learning For Coders, Part 1. (Image credit: Screenshot/fast.ai)

Object recognition: seeing is believing

When they formed fast.ai three years ago, Thomas says it was an open question whether “we could really teach deep learning to people without an advanced math degree.” But the project has been a huge success with thousands of people from over 50 countries taking fast.ai courses, forming an online community, and using the fast.ai library of applications and models. Earlier this year, a fast.ai team of programmers beat Google and other tech giants in an object recognition challenge at Stanford, a victory Howard said shows that anyone can “get world-class results using basic resources.”

Thomas, who is sometimes cited as a leading female figure in the male-dominated field of AI, is open about her fears regarding the dangerous potential of AI if it isn’t checked. “Bias is getting encoded and magnified into AI and machine learning,” she said, citing the example of the poor performance of facial recognition software — software that is nevertheless already being used by some law enforcement agencies — on women and people with dark skin. Other bias issues emerge through the absence of data sets on particular topics. Researcher Mimi Onuoha has identified many such gaps, including “LGBT older adults discriminated against in housing” and “poverty and employment statistics that include people who are behind bars.” Onuoha points out that data gaps in a given field “typically correlate with issues affecting those who are most vulnerable in that context.”

This is a photo of Researcher Mimi Onuoha says "I aim to trouble the assumptions baked into the technologies that mediate our experiences." (Image credit: Brandon Schulman/http://mimionuoha.com)

Researcher Mimi Onuoha says “I aim to trouble the assumptions baked into the technologies that mediate our experiences.” (Image credit: Brandon Schulman/http://mimionuoha.com)

Thomas is also concerned about how ML algorithms used by YouTube and other social media networks have ended up promoting racist and anti-science conspiracy theories in order to optimize time on site. “We’re seeing the social and political impact from not having a more communal sense of what reality is or what science is saying,” Thomas said.

One brutal outcome occurred in Burma, where the circulation of racist propaganda on Facebook, which uses ML algorithms to populate its news feed, contributed to genocide against the minority Rohingya population.  Many people warned Facebook that “the platform was being used to incite violence” but it failed to take effective action, according to Thomas. “As of 2015 they only had four contractors who spoke Burmese.”

So when skeptics ask Thomas whether it’s safe to put the power of AI into the hands of people outside academic circles, she points to this and other examples and says it’s just the opposite, that new people with new ideas are essential.  “Why would anyone think that the people who created all these problems would be the best people to solve them?”

Thomas is tremendously optimistic about the potential for practical AI, especially as the field diversifies and the technology gets into the hands of domain experts in different fields from different regions globally. A fast.ai student from India used ML to improve agricultural lending models so that small farmers can get access to loans on better terms. Another student has worked on using deep learning to identify the exact location of where metastasized cancer originated, thus improving treatment outcomes; others have designed wearable devices for Parkinson’s patients and an app to help the visually impaired.

MOOC, meet AI

When Howard and Thomas announced their first fast.ai deep learning course three years ago, it was to be held in person at the University of San Francisco Data Institute.  But then a researcher from Pakistan who said he had exhausted every massive open online course, or MOOC, on artificial learning topics, contacted them and asked if he could somehow participate. That man from Pakistan inspired them to create fast.ai.live, an online version of the course.  This year, over 2,700 students have enrolled. The in-person class, which has diversity scholarships, has grown to 350 students. The Pakistani student went on to create “the largest corpus of Urdu text ever assembled,” said Thomas, which is significant because “Urdu is a really under-resourced language.”

Thomas is especially optimistic about the growing deep learning community in Africa. Based in South Africa, Deep Learning INDABA runs educational programs and hosts conferences around the continent; Google will soon open an AI research center in Accra, Ghana. “People outside of the stereotypical Silicon Valley employee are making it clear that they’re here and they’re doing research,” Thomas said.

The fast.ai approach to artificial intelligence is practical and down to earth, the opposite of the media hype about computers achieving human consciousness. “I feel some companies are trying to sell things through the idea that AI is too hard and fancy for you to understand,” Thomas said. “I want people to be able to recognize how useful the technology can be in their lives, to know what the risks are, and to evaluate the information about the products. It would be helpful for everyone to have a more general understanding of AI.”