Self-driving cars that crash, psychopathic robots reared on Reddit—these are the kinds of stories related to Artificial Intelligence that capture public (and naysayers’) attention. But instead of reading about machine experiments gone awry, CTOs, CIOs, and CEOs alike need to know about existing AI applications; ones that have been tried and tested, can be easily (and safely) implemented, and can help them grow their businesses today. Headspring’s Dean Goodman provided some foundational AI definitions to demystify the subject and help business leaders understand why AI actually does apply to them. We sat down with Dean to discuss the specific ways that things like machine learning, deep learning, and natural language processing can help businesses to improve their processes, connect with customers, and stay ahead of their competition.
What makes Artificial Intelligence a potential growth driver for businesses?
A lot of the headlines are really about the bleeding edge kind of stuff: self-driving cars and robots that are going to take over the world and that kind of thing. I see fewer articles about the applications of machine learning that can benefit regular, maybe not so high-tech, businesses. I think that’s because it’s not all new research and it’s ground that’s already been broken: things like analyzing past sales data to find trends or product recommendations, which is something that’s become pretty standard for big retailers these days. Or classifying certain types of data. Image and voice recognition are somewhat newer functionalities, but they’re being made available by the big providers via APIs. These are things that developers can start to integrate with, bringing new value to businesses without having to fully understand the data science or machine learning behind it. I think those areas all represent good potential growth because they’re becoming more accessible to a wider variety of businesses.
What are some legitimate resources for those wanting to understand AI and machine learning and their real-world applications?
Start by finding trusted individuals who know a lot about the subject. Andrew Ng is a great person to follow. Actually, the first course I ever took in machine learning was one that he launched out of Stanford in 2013. You can also search Twitter to find active users with “data science,” “machine learning,” or “AI” in their profiles. I also use Flipboard to read news from various sources and aggregate content on the topic. On the business side, Forbes has been posting recently about the business value of AI, and I’ve seen some great articles out of Harvard Business Review. And then on the technical side, there are some good feeds to follow—DataCamp is a really great organization with a strong Twitter feed. If you’re trying to learn data science, they have a lot of good interactive exercises. And then there’s KDnuggets that has a lot of content and is constantly tweeting.
What is it that business leaders need to understand about AI in order to make smart decisions?
AI and machine learning can sound very magical. It’s reported on in this very breathless way at times. But at the end of the day, there are some very real applications, some specific problems that can be solved that previously required people to kind of eyeball the data or have a gut feeling about that data just from past experience. AI gives us these automated tools that can extract further insights that we as people didn’t necessarily see before. There’s a lot of what was previously called data mining or statistical analysis at the basis of machine learning. There’s a ton of business value that can be mined out of the data that we’re already collecting every day. Machine learning is the next-generation toolset being used. These areas are supercritical to our understanding of the capabilities from a business perspective. The subject is a lot deeper than self-driving cars and robots and that kind of thing.
What are some job-specific problems that can be solved with machine learning and how do you determine if they’re worth investing in?
Machine learning is a very natural fit for some specific problem areas. A lot of the big players are providing new APIs to give businesses to access things like voice recognition, image recognition, and OCR (extracting text from image). Specifically, on machine learning front, there are a few different sets of problems:
One of them is classification. A good canonical classification problem is something like identifying a spam email. An email carries a bunch of information with it: who the sender is, the actual content of the message, who the recipient it is, when it was sent, etc. Machine learning can be used for spam detection by training on an input data. We tell the computer, “this is spam,” “this is not spam” and at the end, the output of the machine learning algorithm is the classification: spam or not spam. It can be more than a binary classification, there can be a whole range of classes. That’s the same family of problems that image recognition is solving. Given an image, what is represented in the image itself? A bird, a flower, a building, and so on. It’s a classification.
There’s another set of problems that we solve with something called regression. This is what you might think of as standard statistical analysis. It’s been around for a long time, we just have some new tweaks for machine learning. Whereas with classification there is a set number of outputs (we’d say, of the six possibilities, this is possibility one, etc.), regression gives us a numerical output on a continuous range. A good example that’s often used is real estate: I have all this data about real estate in my area—square feet of a house, what the asking price is, the year of the house, etc. All of that is fed into the machine learning algorithm, with the output being an expected value. That expected value is going to fall into a huge range. What we’re doing is trying to predict with all of those inputs the sale price of our new house so that we can be well-placed in the market.
There is so much we can do with data, but the first question to ask is, “do I have data available for the problem I’m trying to solve?” Without good data, you’re not going to get anything out of a machine learning algorithm. Or worse, if your data is low quality, you’ll get answers out of your model, but those answers are likely to be inaccurate. Certainly, you’ll also want to ask yourself about the business value of the thing you want to model—that calculus is the same as ever.
Can you call to mind any specific examples of companies that are implementing AI really well, or industries and verticals in which it’s becoming more applicable?
Rolls-Royce is an interesting one: The company recently deployed its IntelligentEngine vision that draws on over 70 trillion data points from its in-service aircraft fleet. The program applies machine learning to all of this data to facilitate predictive maintenance on airplane engines and improve their reliability. The goal is to create engines that are connected, contextually aware, and comprehending.
Another great example is a company called Clover Technologies that specializes in recovery, remanufacturing, and recycling technology assets. It partnered with Microsoft to build out a new sales forecasting model that leverages AI. The old model was all in Excel, extremely slow (it took 20 hours to run), and was not that accurate. By investing in a machine learning model, they developed a much more accurate sales forecast that adjusts in real time as new data comes into the pipeline.
Industry-wise, healthcare, marketing, manufacturing, oil and gas are all sectors that already have a ton of data to leverage and that are really starting to see the benefits of AI and machine learning.
Of the many providers that have built out platforms for AI, do you have opinions on which are the most useful?
It’s kind of a big arms race right now. For most of the big providers—Amazon, Google, Microsoft, IBM—the core offerings are the same. They all tend to have speech recognition, vision recognition, some sort of chatbot helper for natural language processing: these are all core areas of AI and they all are starting to provide algorithms that do these things. For those solutions, it comes down to the environment that you’re used to and the toolset that you’re used to. There’s not a lot of differentiation right now, but there’s starting to be some.
For instance, there’s an AI Machine Learning tool developed by Google called TensorFlow that’s becoming really popular for building AI models. Amazon has a TensorFlow hosting platform in the AWS cloud that’s tailored specifically to that utility, so that’s an area that it’s trying to differentiate in. And IBM Watson arguably kicked off the big commercialization of this kind of technology, so they still have an interesting lead in the expert knowledge systems. But at the moment, there’s a lot of work being put into trying to improve the standard APIs that everyone’s had available—none of the APIs are perfect. Machine learning is never perfect, so there’s kind of an arms race over who can make the better model for a given task. For example, an image classifier: we’re waiting to see whose is going to work best.
We referred earlier to recent news highlighting the failures of AI. What would you say to business leaders who may be put off by these types of headlines?
I think it’s important to remember that all of that stuff is really kind of bleeding edge technology. When something new comes along, we’re always going to have to work the kinks out. But some of the classic problems that we talked about a little bit earlier—those are all well solved. You don’t see as much talk about it, just because it’s not sensational. Occasionally you’ll find an article in a news outlet that talks about how a business implemented an AI system in its supply chain and it had a huge benefit. The power of AI and machine learning is not just the sci-fi and futuristic stuff. It can really be applied to standard business problems that we’ve had for quite a long time. It can help us answer things like:
- Which products are performing the best and should we push those?
- Who is most receptive to some specific tailored marketing, or how can we tailor our marketing to our client list?
- What equipment do we have out in the field? Can we monitor and figure out if the equipment is in need of preventive maintenance rather than pay for downtime and repairs because something catastrophic went wrong?
Those are areas that may not make for splashy headlines, but where real business value is realized.
So why invest in AI today?
There are areas of machine learning that are pretty well-understood at this point. If your problem is one that fits well with the things that we know how to solve with machine learning right now (and there are many of them), there’s not a huge benefit to waiting to invest. Generally, having some model for solving your problems is better than having no model for it. Maybe something new and better will come along, but who knows how much longer and how much of your competition is already starting to make use of AI.
It’s also becoming more and more cost effective to take on these kinds of projects. There are a lot of new tools online, and more and more there’s a push to make standard tools available that software developers are using to do things such as make API calls. The availability of these solutions means that you don’t necessarily need to understand the behind-the-scenes of how the technology works to implement it. There’s plenty of opportunities right now to harness the power of AI in a cost-effective way.