On “Deep Learning” AI and Not Sitting At the Back of the Bus
In my post on the emerging tech of augmented reality (AR), I argued that the AR offers community groups — especially those in communities that our society has written off — a unique opportunity to change the course of emerging tech so that more the benefits flow to their communities. I think the same is true for AI.
As I mentioned in an earlier post, robots aren’t ready for ordinary folks to do real work: the tech is still too primitive. But a lot of robotics will be driven by technology where the baby version of it is available now. Most notably, although AI is ridiculously underpowered compared to where it’ll be in 20+ years, you can actually do something useful with it today — and the field is really taking off. Forrester predicted that investment in AI would grow by 30% this year. Gartner predicts that “by 2020, AI technologies will be virtually pervasive in almost every new software product and service,” and IDC estimates that by 2020 revenue from AI-based systems will hit $47 billion. There is also an almost insatiable demand for people who are fluent with current AI tech.
As a result, Google, Microsoft, and others are in a race to see who can make their AI programming libraries and tools more accessible. Today, a typical programmer could learn how to write a 15-line program that will, say, recognize pictures of a cat or answer a moderately impressive range of questions that only advanced practitioners could accomplish a decade ago (see more at the end of this post). Becoming really skilled at this tech still takes a lot of work, but the barriers to entry are far lower than they used to be.
There’s also been a flurry of work online to make it easier to learn these AI tools. For example, Siraj Raval has created a charming, funny YouTube series aimed at hackers who want to do cool stuff with AI; Josh Gordon at Google and Brandon Rohrer at Microsoft (Microsoft: whole course) have also produced really helpful YouTube tutorials. And there are first-rate free courses at Fast AI that make it much easier for someone with some programming background to do some really impressive work.
Almost all the major players in this space they want to “democratize” this technology. Recently, for example, Google launched the People and AI Research Initiative PAIR), whose goal is
“to focus on the “human side” of AI: the relationship between users and technology, the new applications it enables, and how to make it broadly inclusive.”
Based on what they’ve done so far, I think the major tech players are very sincere in wanting to make AI more accessible.
But when these players use the word “democratize,” what their work shows they really mean is making it fully accessible to people like me, who are programmers. There are more and more types of analysis that you can do using AI by pushing a few buttons. But once you go beyond basic work, there’s a pretty big gap between what most tech players they want to do and what they are doing.
That’s why I think a network of community groups could have a real impact on AI. And unlike just a few years ago, there are a lot more opportunities for some interesting partnerships.
For example, fast.ai’s terrific courses say they are aimed at “anyone with at least one year’s coding experience.” If any network of community groups wanted to work with fast.ai to create courses and a community-oriented ecosystem for learning that would make this knowledge accessible to even more people, I’m sure the people at fast.ai would be excited by the idea.
Playing with AI today isn’t as fun as playing with augmented reality. But there are some incredible opportunities here to help shape the direction of the tech industry and who will benefit from this emerging tech.
NOTE: If you are one of the readers of this blog who likes geeking out with tech and want to get your feet wet with AI, I’d recommend playing with Keras. Keras is an open source library developed by a someone at Google which makes coding “deep learning” AI much easier. Through Keras, you can work with a variety of AI libraries, including Google’s Tensorflow and Microsoft’s CNTK, using the same code. The other fabulous thing about Keras is that it bakes in a lot of best practices. When I first tried to start learning machine learning and AI a few years ago, figuring out how to write the code wasn’t that hard. But to properly use the code, you needed to make decisions about a bewildering number of options, and trying to figure out what those options meant was very painful. Keras sets a lot of those options for you, see you don’t have to worry about tweaking them until you really know what you’re doing.
If you do decide you want to play, I would strongly recommend starting with Fast AI‘s Practical Deep Learning for Coders. It’s a very impressive piece of work. And in the first hands-on lesson, you learn how to do some really cool stuff in just a handful of lines of code.