[Article] Ways to think about machine learning

Ways to think about machine learning by Benedict Evans

Photo by Alex Knight on Unsplash

Pretty much everyone, even people outside tech/startup circle, has heard or read about machine learning and knows it is a “Next Big Thing”. Tech-savvy people might understand the theory of neural networks and how machine learning addresses a class of questions that were previously “hard for computers and easy for people”, or more usefully, “hard for people to describe to computers”.

However, knowing the math behind machine learning or watching all speeches and cool vision demos doesn’t help us to think structurally about what new things it could enable, nor what it will mean for companies in the broader economy.

I don’t think, though, that we yet have a settled sense of quite what machine learning means

Relational databases — a parallel comparison

Before relational databases appeared in the late 1970s, databases were simple record-keeping systems. Performing arbitrary cross-referenced query often would turn into a custom engineering project. Relational databases turned them into business intelligence systems. They changed what computers could do by providing a new fundamental enabling layer.

With this new layer, companies like Oracle and SAP was born. These companies gave us global just-in-time supply chain, in turn gave us Apple and Starbucks. PeopleSoft, SuccessFactors, SalesForce and dozens more new billion dollar companies were all ran on relational databases. This technology became an enabling layer that was part of everything.

Similarly, ML is a step change in what we can do with computers, a new enabling layer that companies can build on top of.

Eventually, pretty much everything will have ML somewhere inside and no-one will care.

The question is: how do you get from explaining table joins to thinking about Salesforce.com? We can do impressive demos of voice recognition and image recognition with ML, but what would a normal company do with that?

Machine learning has limited “winner takes all” effects

Machine learning is all about data, hence the saying: “Data is the new oil”. It is easy to believe that big companies likes Google, Microsoft, Amazon, BAT etc have absolute advantages because they “have all the data”. However, data is highly specific to particular applications and isn’t fungible.

More handwriting data will make a hand-writing recognizer better, and more gas turbine data will also make a system that predicts failures in gas turbines better, but the one doesn’t help with the other.

… machine learning lets us solve classes of problem that computers could not usefully address before, but each of those problems will require a different implementation, and different data, a different route to market, and often a different company.

Machine learning is *NOT* a humanoid robot with general intelligence

People often imagine we’re creating something anthropomorphic, something with general intelligence. However, what we get in reality is more like washing machines: they are only functional when given specific kinds of input, they’re not general purpose even in the narrow domain of washing.

What, then, are the washing machines of machine learning, for real companies?

Two sets of tools for thinking about the application of machine learning

1. Gather and process new types of data, ask new types of question

Machine learning allows us to:

1. ask old questions on old data, and expect better results i.e. as an analytic or optimization technique.
2. ask new questions on old data, e.g. with sentiment analysis, computers can perform tasks like “Read all the emails and find the anxious/angry ones”
3. ask new/old questions on new data
   Computers have been able to (and good at) process large amount of text and numbers since its birth. With machine learning, audio, images and video could become “machine-readable” . This opens up new data types to analysis.

2. Automation of tasks that could not previously be automated

For example, “listen to all the phone calls and find the angry ones”, “sort a pile of photos into men and women” etc. Tasks that previously required a person to perform, but even a ten year olds could do. Machine learning gives us infinite ten year olds such that we could automate one discrete task, at massive scale. (except that ten year olds do have general intelligence and common sense, unlike any neural network we know how to build)

In certain fields, machine learning could even find levels of pattern, inference or implication that human can’t recognize, by looking at enormous amount of data that no human could ever do in lifetime, and bring out new results (think about Deepmind’s AlphaGo). Again, the question is what to do with this power of automation.

But what would you do if you had a million fifteen year olds to look at your data? What calls would you listen to, what images would you look at, and what file transfers or credit card payments would you inspect?

Final

We’re just at the start of machine learning explosion. What else? What it will enable, other then low hanging fruit like optimization problems, image recognition or audio analysis questions? There are probably ten to fifteen years to go before we exhausted all the possibilities.

Summary

1. Machine learning is a new enabling layer
2. Machine learning has limited “winner takes all” effects
3. Machine learning today is far from general intelligence
4. Machine learning enables procession of new data and new questions
5. Machine learning enables new types of automation
6. The questions are: what else it will enable? what can we achieve with this power?