Imagine a typical day. You tell your kids’ X-Box to turn itself off again as you walk through the living room. On the way in to work you get stuck in traffic, grateful for your new car’s auto-braking feature as other drivers brake erratically. You can at least use the time in the highway parking lot to have your phone read you some personal emails, telling the voice assistant to make notes on replies you need to make later. At work, you open your laptop and check your work email, then quickly hop on Facebook to see that you have been tagged in your friend’s photo, while another mutual friend you both saw this weekend is recommended as a connection. Your phone rings with a call from your credit card company. You pick up — it has noticed dubious transactions on your account. You have a quick conversation with the interactive voice system to straighten it out. Then you open up your streaming music system and get to work.
It’s 9:10 AM and you have already used or benefited from dozens of different machine learning systems. Did you notice them all?
Most problems are still solved by traditional heuristics — what we think of as “regular programming” — but there are certain domains where Machine Learning now thrives. These domains include audio and video processing, separating or grouping data based on implicit rules, and learning behaviors.
The X-Box, phone, and the bank’s Interactive Voice Response system all use voice synthesis and voice recognition. Cameras, photo organizers, and social media applications now recognize faces. Email spam filters and personalized newsfeeds work by learning to separate or group data based on the data’s implicit rules. Web products recommend music you might like or friends you have in common using a similar approach, while fraud detection systems recognize patterns of behavior that deviate from a learned norm. Automotive technologies like self-parking, lane-keeping, and automatic braking work by modeling expert drivers and applying that knowledge to your real life driving. Intelligent agents like Siri combine approaches, learning to recognize patterns of speech or text that signal routine tasks, and then using learned behaviors to handle them.
Yet these are only the visible systems. Many other machine learning systems are embedded in the infrastructure around you, where you rarely ever notice them. Every time you use your Nest thermostat, you teach it to predict your usage patterns. At the electric utility, a load modeler performs a similar function on a industrial scale, planning a strategy for the day’s power generation. It seeks guidance from a weather modeler — a system that has learned to predict upcoming weather based on recent atmospheric conditions and historical cycles. In turn, the power generation forecasts inform financial trading systems that buy and sell fuel futures contracts to manage the utility’s price risk.
The electric grid and similar utilities, like the cable, gas, water, and telephone systems, are all maintained by work teams. Those crews are dispatched by planning software that has learned how to optimize maintenance cycles. Operations management systems monitor for service delivery problems, decide which events are important and how to react to them, and advise the schedule planner when work teams need to be dispatched to restore — or proactively maintain — service.
Your local streetlights might be timed by an adaptive management platform that learns to optimize light cycles for different traffic loads. Those products often use image recognition to determine when the traffic cameras see cars stopped at intersections. Highway transportation departments use similar systems that learn typical traffic patterns so they can detect and respond to congestion. Web-based map applications operate at a larger scale, learning to predict how accidents and road closures will affect travel times so they can recommend alternate routes.
We see similar solutions deployed at work. The office network is often protected by an intrusion detection system that has learned normal traffic patterns and notices unusual activity. Data centers produce a constant stream of notifications, such as disk failures, temperature readings, and storage capacity fluctuations. To make sense of the torrent of information, we deploy management software that learns which events are important in our environment and which are routine. More and more these systems depend on machine learning.
Although machine learning technologies have a limited role, they have become embedded in other technologies so ubiquitous that we are all using them on a daily basis. Intelligent machines are not coming — they are already here. And they are here to stay.