Why machine learning will decide which IoT ‘things’ survive

No billion-dollar machine could replace a doctor. But a $25 machine can tell you when you need one.

In 1996, the ER at Cook County Hospital of Chicago used an algorithm to determine when a patient with chest pain was in danger of having a heart attack and was thus worth one of its scarce hospital beds. Using a systematic, flowchart-based approach of basic tests, the algorithm proved not only to be quick and efficient, but accurate: It sorted 70 percent more patients into the low-risk category, but caught a higher percentage of heart attacks (95 percent) than human doctors (75-89 percent). And this was before any deep computing was involved.

[aditude-amp id="flyingcarpet" targeting='{"env":"staging","page_type":"article","post_id":2142894,"post_type":"guest","post_chan":"none","tags":null,"ai":true,"category":"none","all_categories":"ai,big-data,cloud,entrepreneur,","session":"C"}']

Now consider that there are around 6.4 billion IoT devices in use this year — nearly one for every living human. If even one percent could analyze people for medical conditions, by collecting data on pulse, diet, or sleep, it would extend the reach of the world’s doctors by a factor of five.

But the real magic comes from machine learning. Beyond just applying singular algorithms in more places, data collected at this scale is already finding patterns in conditions that even human doctors couldn’t see after decades of experience. Imagine, for example, a Fitbit noticing fluctuations in your pulse that correspond strongly to a heart condition, sending you to a hospital for treatment. Machine learning means solving impossible problems with household devices.

But machine learning stopped being a question of potential when IBM’s Watson and Google’s DeepMind started outperforming humans in domains like Jeopardy and Go. Now the question is this: If a Fitbit could save your life and a Nike+ Fuel Band couldn’t, which would you buy?

The real value of “smart”

Among swarms of “smart” gadgets and gizmos, it’s no wonder the ones leading the way are doing so with machine learning. Take Nest, the quintessential smart device, for example. Nobody buys it so they can turn up the heat with their phone. They buy its energy-saving capabilities, the intelligent way it solves a problem that couldn’t be done before, automating a home’s temperature based on people’s presence and needs, not a simple timer.

Yet most manufacturers just chase convenience. Phillips HUE lights, while nifty, earn their “smart” label because you can control them with your phone. It’s not a problem that needs solving. You wouldn’t call a human smart because they can flip a light switch on their own. So why give that label to a device?

The lack of true smart features in consumer IoT is also what’s holding back adoption. Remote access door locks or a radio that turns on when you get home are nothing more than luxuries, consumed with the exclusivity of fine dining or cruise ship packages — earning revenues almost exclusively from the upper class.

Machine learning turns the wants into the must-haves: thermostats that keep you warm while saving you money, sleep or fitness wearables that give you personalized tips, or environmental monitors that diagnose and treat sources of pollution before they get a chance to harm your family.

Machine learning will define the winners — permanently

Products with machine learning features look sexier on the shelf than ones without. But the nature of machine learning means that among competitors, the ones that goes the furthest in machine learning get to keep their advantage for a long time.

[aditude-amp id="medium1" targeting='{"env":"staging","page_type":"article","post_id":2142894,"post_type":"guest","post_chan":"none","tags":null,"ai":true,"category":"none","all_categories":"ai,big-data,cloud,entrepreneur,","session":"C"}']

Thanks to the cloud, building machine learning into a device isn’t a design problem (all it takes on the device is connectivity) nor is it a hardware problem (the heavy processing can be done remotely). It’s somewhat of a talent problem since capable engineers are rare, but this is always solvable with enough funding. More than anything, it’s a data problem.

For a computer to reliably study patterns, the data set needs to be enormous. It needs to consider a multitude of factors, ranging from user preference to use cases, environment, and much more. But many or even most of these factors are time-dependent: frequency of use, frequency of behaviors, frequency of conditions, changes to user behavior over time, seasonal changes to the environment, data accuracy over the lifespan of the sensor, etc.

Time is doled out democratically; a hundred million devices on solid connections won’t make a company’s clock turn faster. A six-month lead on a competitor can’t be closed with more users or funding. Your data will be fundamentally better than theirs, shown in the accuracy of your readings, and the number of features you support as earlier features become reliable enough to finalize and ship. For as long as you stay active, you’ll be the leader and the competition won’t be able to catch up.

Not just for the big boys

Yet it’s still just the IBMs and Googles that ship machine learning products with any frequency. It’s as if machine learning was too expensive for startups. But this simply isn’t true.

[aditude-amp id="medium2" targeting='{"env":"staging","page_type":"article","post_id":2142894,"post_type":"guest","post_chan":"none","tags":null,"ai":true,"category":"none","all_categories":"ai,big-data,cloud,entrepreneur,","session":"C"}']

The trick is to do the heavy lifting on other people’s computers. Again, this is possible because of the cloud. Startups can pay by the hour to get access to some of the most sophisticated machinery out there with reasonable and manageable investments, most of which aren’t up front. With a few lines of procedural code, you can even queue up many batches in a row to keep things efficient.

More importantly, as there is little required hardware on the device itself to make machine learning possible, you don’t need to commit to machine learning when shipping your first batch, when design and tooling upfronts still weigh heavy on margins.

Even Nest wasn’t so smart at first; a phone-controlled thermostat roughly predicting how long it takes to heat your house through simple algorithms. It wasn’t learning much about you. But to upgrade user homes with its defining features, the company only had to send packets, not packages. Machine learning can be added at your financial convenience (as long as you do it before your competitors do).

Democratizing expertise

It’s scary to think of machine learning as a chopping block for complacent startups. But there’s much more reason to be optimistic than afraid. Machine learning adds more value than we ever imagined. It puts a doctor in every fitness band, a detective in every smart lock, a health inspector in every environmental monitor, and a butler in every luxury device.

[aditude-amp id="medium3" targeting='{"env":"staging","page_type":"article","post_id":2142894,"post_type":"guest","post_chan":"none","tags":null,"ai":true,"category":"none","all_categories":"ai,big-data,cloud,entrepreneur,","session":"C"}']

Machine learning is where smart devices stop being convenient and start being powerful. We’ve seen the early adopters, the Nests and the Echoes, rise to the top and add serious value to our lives. When hundreds of other tech companies follow suit, the world will never be the same.

Jacques Touillon is CEO of AirBoxLab, maker of consumer smart device Foobot.