What is: machine learning?
Everybody is talking about artificial intelligence (AI) these days. Products like Siri and Alexa have shifted the way that humans interact with technology forever. Algorithms are being used to optimize even the most mundane of digital tasks. Automated marketing tech and big data are giving brands unprecedented access to our wants, needs and desires.
Pundits across mainstream media and in the blogosphere are predicting that AI tech will come to play an even more important role in our lives in the decades to come.
And yet hardly anyone we know has a robot butler yet!
Because, as we discussed in our previous post “What is artificial intelligence?”, higher-level AI technology (like robot butlers) is still a long way off.
What we have instead – for the time being at least – are highly specialized programs capable of completing specific tasks with amazing efficiency and accuracy.
The reason for this amazing efficiency and accuracy? Machine learning.
We’re pretty sure of the following facts:
- Alexa didn’t understand everything the people talking to her said right away.
- Google’s self-driving cars had an accident or two along the way.
- Netflix was wrong when it first predicted that we would be excited to see Adam Sandler’s latest flick.
- If the machines behind these technologies didn’t improve at the tasks they were programmed to complete, those technologies would most likely suck.
Thank heaven for machine learning…
What is machine learning?
As Tech Target defines it, “Machine learning is a type of artificial intelligence (AI) that provides computers with the ability to learn without being explicitly programmed.”
Which is half true.
In reality, explicit programming is actually still necessary, and no machine has yet been designed that is capable of learning anything that falls outside of that programming.
Put simply, machine learning is a subset of AI in which a machine is designed to perform a specified task and programmed to get better and better at completing that task over time.
While there are a few different types of machine learning, each with a slightly different method of teaching an AI to complete its specified task, the teaching process is always essentially the same. Through an exhaustive process of trial and error known as “training data”, a machine can “learn” what works and what doesn’t. Unsuccessful strategies and methods are discarded, successful ones are honed and developed. In this way, the machine’s ability to complete its task becomes more efficient and effective with each attempt.
Whether a task is simple or complex, whether it involves many variables or very few variables, a properly programmed machine will (if given enough time) figure out the best and most efficient way to complete it.
Seven examples of machine learning you’ve probably already met
#1: Your Facebook News Feed
Each and every time you perform an action on Facebook, your experience is personalized just a little bit more. This is because Facebook’s News Feed algorithm looks at your past behavior to predict what you’ll engage with next.
For example, let’s say you add a new friend on Facebook. If you comment on that friend’s statuses frequently, Facebook will assume that you enjoy interacting with that person. Then, Facebook will show you more of their status updates, photos, etc. in your News Feed.
However, if you scroll past all of that new friend’s updates in your feed without reading them, Facebook will assume you’re not interested in their updates. They’ll show you less and less from that friend as time goes on.
This process is constantly happening in an effort to make Facebook as enjoyable as possible for you to use so they can keep you logging in.
#2: Google’s self-driving cars
Some people may be put off by the concept of self-driving cars, but this technology has the potential to greatly reduce the number of traffic fatalities that occur each year due to human error.
As Dmitri Dolgov, Head of Waymo (a self-driving car project formerly part of Google) says on Medium, “We know that human error is involved in 94% of all crashes. That’s why we’re working harder than ever to bring self-driving cars, that don’t get tired or distracted, to our roads.”
Waymo uses machine learning to improve Google’s self-driving cars’ ability to handle complex situations. The cars are tested by being driven automatically in various locations with a driver at the ready to take over in any dangerous moments. These moments are called ‘disengages’. The number of disengages for each 1,000 miles has proven to be quite an effective metric for measuring the performance of self-driving cars.
As the following graph shows, over a year of machine learning, Waymo’s cars required only a quarter as many disengages in 2016 as they did in 2015.
This shows that we’re getting closer and closer to a future where self-driving cars are an everyday reality for many people, and we have machine learning to thank for that.
#3: Your Netflix recommendations
The more you use Netflix, the more finely tuned its recommendations for you become.
For example, let’s say you binge watch an entire season of a comedy show on Netflix.
Then, you go on to watch some stand-up comedy.
Netflix will start recommending other comedy series and stand-up performances that other users enjoyed.
But if you start watching romantic dramas, Netflix will start giving you some rom-com recommendations as well.
This all seems incredibly basic, but these “recommendation engines” are such a part of your daily life that you may not even notice them.
Netflix’s recommendations feature is a key part of its business model, designed to keep you engaged. For this reason, Netflix has employed finely-tuned machine learning programs. They become increasingly accurate in identifying which films/shows its users will want to watch over time.
#4: Your Spotify mixes/radio stations
As you listen to more and more music on Spotify, the platform learns what genres and artists you enjoy. Spotify uses this information to make daily and weekly mixes with new music its algorithm thinks you’ll enjoy.
Machine learning is also used on the Radio Station feature of Spotify. You can upvote or downvote individual songs that come up on the radio, and Spotify will adjust the station accordingly.
For example, let’s say you put on a general “rock” station and the first song is by the Rolling Stones, which you dig. Then, a song by a hair metal band like Guns N’ Roses comes on. If you downvote that song, your rock station will have less hairspray in it as time goes on.
#5: Your email spam filter
Spam filters like Gmail’s use machine learning to predict which messages should go to your main inbox and which messages should go straight to the spam (AKA “promotions”) folder.
They do so by comparing the content of emails with the number of users who label emails as spam and identifying patterns. For example, if an email mentions:
- wiring money abroad,
- adult content,
- or whatever singles in your area are supposedly doing,
then it’s likely to be marked as spam. Other red flags Gmail has learned to look for are emails with lots of misspellings and an abundance of spammy words and phrases.
While spam filters are sometimes wrong, overall they work quite well… and this is thanks to machine learning.
#6: Your email marketing subject lines
While spam and email marketing are two completely different things, they both face the same problem: how do they get into your inbox and convince you to open their emails?
To accomplish this, Phrasee employs machine learning algorithms to optimize the language used in email subject lines.
Through the testing of millions of subject lines and measuring the human responses they generate, Phrasee’s learning machines can predict the language that will produce the highest number of opens and clicks. The end result, of course, is more conversions and more ROI for brands.
Phrasee’s machine learning has become so effective that we now provide optimized language for many of the world’s biggest brands. The subject lines written by our tech outperform those written by humans in over 90% of cases.
#7: Your internet access in remote places
Google has a pretty ambitious plan to provide the internet to remote locations by using hot air balloons that are synced up with satellites. They have dubbed this effort “Project Loon”. These balloons can provide internet service to any area, sort of like a floating cell phone tower.
However, a huge issue with the balloon WiFi project has been that of profitability.
Since the balloons can only rise or fall, and can’t navigate their environment otherwise, they must rely on using air currents and drafts in the stratosphere to stay in roughly the same location. This is necessary to provide consistent internet coverage to end users.
In the past, it was difficult for Google to maintain internet coverage in a location without using hundreds of balloons. This proved to be cost prohibitive. By using machine learning, Google’s X division has found it possible to keep balloons in place much more easily than they thought.
As Astro Teller, the rollerblade-wearing director of Google’s X division states in an interview with Wired magazine: “We can now run an experiment and try to give service in a particular place in the world with 10 or 20 or 30 balloons, not with 200 or 300 or 400 balloons,” Teller said. In the process, Teller states that Project Loon “… has a much better chance of ultimately being profitable.”
So as you can see, machine learning is already all around you, quietly giving you a helping hand. What other problems can machine learning solve for you?
Want to know more about AI?
Head over to our Learning Hub!