AI for enterprises: Start small and choose projects wisely

Despite years of overblown stories about artificial intelligence, there has been surprisingly little mainstream discussion about how the techniques we classify as AI actually work, and how they might apply to enterprises trying to get started with it. Everyone has heard grandiose predictions about AI adding trillions of dollars in economic value, or dire predictions about super-intelligent machines ending civilization, but how many business leaders have even an inkling of how to meaningfully get their companies started down the path of artificial intelligence?

In this episode of Cloud Native in 15 Minutes, AI expert Andrew Ng—whose past includes leadership roles at Google, Baidu, Coursera, and Stanford, and who is currently leading Landing AI, deeplearning.ai, and the AI Fund—offers up just that type of advice. Ng explains the most effective technique for doing AI today (supervised learning, often powered by deep learning), and gives examples of how it has been applied across numerous industries. Among other things, he also discusses how to scope early AI projects; when to build versus buy; and the connections between AI, data science, and automation.

Here are a few quotes from the episode, where Ng explains the importance of starting small, choosing projects wisely, and understanding supervised learning.

Start small and grow with success

“I’ve seen a lot more companies fail by starting too big than by starting too small. So, I recommend to most companies to start with a few pilot projects, deliver a quick win in 6 months, and then use that to ratchet up to bigger and bigger projects. …

“My first internal customer of the Google Brain team was Google’s speech recognition team. Speech recognition was not, maybe still is not, Google’s most important AI application—it’s not web search or advertising—but by delivering a quick win there, it helped other teams within Google gain more faith in what my team and I could do. And then we got our second customer, which was to help Google Maps use AI to read house numbers to more accurately geolocate houses and buildings on Google Maps.

“Only after delivering this second win to this second internal customer did I then go on and start the more serious conversation with the advertising team.”

Choose your AI projects wisely

“Scoping the right AI projects is really hard, and there are a few gotchas on identifying the right thing to work on. One of them is just unrealistic expectations about what AI can and cannot do. For example, a few years ago, a lot of people thought chatbots could soon have fully general-purpose conversations and talk about almost anything with anyone. That turned out not to be true, and companies that tried to build fully general-purpose chatbots did not succeed because the technology was not there.

“So I think it’s important for executives to have enough of an understanding of what AI can and cannot do to scope the valuable and feasible projects, and then also to have the appropriate mechanisms for staffing them, tracking them, and helping set up the engineering team for success.”

Supervised learning and input-output mapping

“Deep learning is maybe one of the hottest technologies in AI right now, and what we’ve found over the last several years is it’s a wonderful technique for doing supervised learning. For example, if you’re running a factory and you want an AI to look at the smartphones going down your manufacturing line and figure out if they’re scratched—being able to to automatic defect inspection—then it turns out that deep learning is a very effective way to learn that input-output mapping. You can input, say, a picture of a smartphone, and the output will be, ‘Does this smartphone have a scratch on it or not?’

“A lot of the recent rise of AI and supervised learning is driven by discoveries in deep learning, allowing us to make … very accurate input-to-output mappings.”