New Electricity


Jeremy Howard


October 11, 2016

I’ve been saying for some time that Deep Learning is going to be even more transformative than the internet. This view is shared by the always insightful Andrew Ng (Chief Scientist at Baidu, former CEO of Coursera, and head of Google Brain–and perhaps the only person I’m aware of who understands both business strategy and deep learning). This month’s Fortune magazine has Deep Learning as their cover story, and in it they quote Ng as saying: “In the past a lot of S&P 500 CEOs wished they had started thinking sooner than they did about their Internet strategy. I think five years from now there will be a number of S&P 500 CEOs that will wish they’d started thinking earlier about their AI strategy.

In fact, Ng goes even further, saying “AI is the new electricity. Just as 100 years ago electricity transformed industry after industry, AI will now do the same.” Fortune discusses this in a commentary titled The AI Revolution: Why You Need to Learn About Deep Learning, which is a most timely reminder, given that applications for the first deep learning certificate close in two days!

I remember how many of my colleagues and clients reacted when I was at McKinsey & Co in the early nineties, and I was telling everyone I could that the internet was going to impact every part of every industry. At that time, as a very new consultant, I very little success in getting heard. (In hindsight, I clearly should have left consulting and started a company based on my conclusions!) I hope that this time around I am in a better position to help organizations understand why they need to invest in deep learning as soon as possible.

I have had many opportunities to discuss this issue with the S&P 500 executives who have attended my data science classes as part of the executive program at Singularity University. Many execs have gone on to develop data driven organization initiatives at their companies - but for those that don’t, these are some of the excuses that I’ve heard:

  1. As a big company, we focus on competing, and our competitors aren’t doing this now
  2. We run on expertise–we don’t trust models, but trust our instincts
  3. Our data is too messy; our data projects aren’t ready yet
  4. We can’t hire the right experts

Let’s look at each of these in turn.

1. You need to lead, not follow, on massive industry transformations

The lesson of the internet shows us the danger of being a follower when there’s a massive industry transformation going on. Whether it is Kodak vs Instagram, Amazon vs Borders, or any of the other pre-internet companies that got destroyed by new competitors, there are more than enough examples of the danger of waiting until you see your competitors’ completing transformation projects. You won’t know about your new competitors until it is far, far too late. And it’s much easier to get started early, when there’s time to build up the infrastructure and capabilities you need.

We can also see from the internet example that companies that are amongst the first into a space are the ones that win in the long term. Look at some of these examples:

  • Thomas Edison’s original electricity company today is GE
  • The company created for the first punch card data collection (in the 19th century for the US census) today is IBM
  • The first significant e-commerce company was Amazon.

2. Data and instinct work together

There is nothing wrong with trusting your instincts as an industry leader–for most execs, it’s your instincts that have gotten you to where you are today. But today’s data-driven companies are powering ahead on every metric that matters; the best role model is surely Google, which has nearly all leadership positions filled by computer science and math PhDs, and has used data to become one of the world’s largest companies in quick time.

Data and models should not be used to make decisions on their own, and neither should instinct. The best execs use a combination of both. Deep learning models provide deeper insight and greater accuracy, make existing products better, improve operations (e.g. Google used deep learning to reduce data center cooling requirements by 40%!) and make new classes of product available.

3. Using the data and infrastruture you have now is the best way to start

Every large organization I’ve even worked with has been in the middle of a major data infrastructure project going on at all times. If you wait until your data infrastructure is perfect, you’ll never start actually using that data to create value! There’s a lot of benefits to starting to use the infrastructure you already have to start creating value now:

  • You learn quickly what data is the most valuable in practice, and can focus your development efforts there
  • You create role model project results you can use to evangelize data-driven projects throughout the organization
  • Your further data infrastructure work can be funded by the value from your initial projects
  • You find out which of your team is most effective at delivering value from data, and can identify your recruiting needs more accurately

Deep learning is particularly effective at handling noise in data, and in handling unstructured data - so if your data infrastructure is not in a good state, it’s even more important that you invest in deep learning.

4. Rather than hiring experts, develop them internally

The people that best understand your business are the people who are in your business. Looking externally for deep learning experts, rather than developing deep learning expertise within your existing staff, means that you will be creating a gap between your domain experts and your new data experts. This gap can be nearly impossible to fill, and can lead to many organizational problems.

Furthermore, deep learning experts are like unicorns at the moment–there are very few available, and they are very expensive ($5m-$10m acquihire value, according to VC Steve Jurvetson). But any reasonable numerate coder can develop deep learning skills within a few months; in fact, we’re trying to teach the best practices in just seven weeks in our deep learning certificate!

The best approach, of course, is to do both: hire existing deep learning experts if you can, whilst developing your own team’s skills at the same time.

In closing

If you think that your organization should heed Andrew Ng’s advice, please send this article to the manager of every team that you think could benefit. My talk Deep Learning Changes Everything is included on USF’s deep learning certificate site, and has more information, including a sample deep learning lesson.