AI, or more specifically machine learning, can sometimes seem like an incredibly powerful thing that does all kinds of stuff.
But to understand it, it helps to break it down into its key features and capabilities. We’ve identified 6 of these, and you’ll find that more often than not, machine learning systems really implement 1 of these features.
Here’s the 6 capabilities:
- Personalization and profiling
- Predictions
- Pattern recognition and anomaly detection
- Natural language
- Object identification
- Goal achievement
Truly powerful and successful systems even combine these together, to become even more versatile and accurate.
Personalization and profiling
Personalization and profiling systems use machine learning to understand a person as a unique individual. Based on what actions a person takes, what choices he or she makes, a profile of that person is built.
Usually, the profile targets the person’s preferences, and aims to predict the person’s likes and dislikes. This profile gets continually refined each time the person does something.
How it’s being used in the world today
You’ll see such personalization and profiling systems all over the place. When you listen to music on Spotify, and Spotify recommends new songs that you’ll probably enjoy, that’s a personalization system. It’s based on the types of music you listen to, which tracks you listen repeatedly, which you stop listening halfway, and a dozen other signals.
Same with Netflix, and Amazon shopping. And even online shoe retailers. Sometimes, we call these recommender systems as well.
Another form of personalization and profiling is when you try to apply for a service. Say, a bank loan. When you apply for a bank loan, your bank is using all sorts of information about you, to build a profile on whether you are likely to be a good customer and repay the loan. If your profile says you are, you’ll most likely get that loan.
Predictions
Prediction systems, as its name suggests, is designed to use data to predict what will happen in the future. They usually draw on historical data, learn the patterns within them, and use that to make predictions.
The aim of a prediction system is to say “there is a high probability of this happening, given current data”. This helps a human to make decisions. We sometimes call these data-driven decisions.
How it’s being used in the world today
Prediction systems are used to predict stock and currency exchange prices, and manage inventory by advising the merchant when to bring in new stocks.
They’re also used to predict whether potential customers are likely to buy from you, or whether your existing customers are likely to leave you and switch to your competitor. They also help to predict whether a candidate is a good fit for a company, or whether an employee is likely to leave, and why.
With insights such as these, a human – say a HR manager, or the customer-retention department, can come up with incredibly well-targeted incentives, programmes and campaigns.
Natural language
As humans, we don’t communicate the way computers do. We use language. And our language is unique in that it is full of ambiguities and contextual meanings.
A conversation system is designed to understand human speech, and reply in natural language. These can take the form of voice, text and even images. The goal is to let us interact with machines is a completely natural way. A human way. We call this Natural Language Processing, or NLP.
Conversation systems don’t stop at trying to understand your natural language. They also dig deeper and try to understand your emotions behind what you say – are you happy, frustrated, angry, or surprised? This is called sentiment analysis.
How it’s being used in the world today
Probably the most famous example would be Siri and Google Assistant. They are voice-activated assistants which you can speak naturally to, and they understand you and carry out your requests. You can talk to them, or type to them. They also reply to you accordingly.
Chatbots in general also fall into this category. You’ve probably seen them when you visit websites that have a chat widget that you can interact with. As a point to note, some of these are powered by humans, and those are live chat systems. Chatbots are generally AI-powered and completely autonomous, with no human intervention.
Pattern recognition and anomaly detection
Pattern recognition systems tries to find consistent patterns, and then understand what is normal, and what isn’t. When something isn’t normal, it’s an outlier, or an anomaly.
Patterns are interesting, because when you’re doing something and you’re halfway through a pattern, the pattern recognition system can help you complete it. Think of when you’re searching for something on Google, and it suggests just what you wanted to search for. Or a programmer who just started writing something fairly standard, and the editor completes it for him or her.
Often, anomalies are even more interesting, because we can act on it. Anomalies may be things we need to take care of, because they shouldn’t be happening and we need to fix it. They may also be opportunities, like an advertisement that is doing exceptionally well, and we can study it, learn from it, and replicate it for more success.
How it’s being used in the world today
An example of a pattern recognition system you see everyday is Google Suggest, which was briefly mentioned earlier. Google Suggest looks at what you’re typing at the moment when you’re searching for something on Google, and helps you complete your query. We know it can be very accurate. Sometimes in a bit of an unnerving way.
Another Google example is in Gmail, when it helps you with suggestions to complete your sentence. That is a pattern recognition system, paired with a personalization and profiling system, as the suggests are tailored to your writing style as well.
Other examples include data loss prevention systems, which prevent people in a company from inadvertently or intentionally leaking data out of the organization. Leaked data is an anomaly.
Similarly, fraud detection systems work in generally the same way.
Object identification
Object identification systems use machine learning to recognize things in the world. They work with all kinds of media, including images, video, audio, and combinations of them. We often also call this “computer vision”.
Their aim is to recognize the objects that they are trained to recognize. That may be a face, or a specific face, in facial recognition. It may be dangerous objects when trying to prevent accidents or attacks. It may be detecting obstacles in self-navigating cars or robots.
How it’s being used in the world today
This is one of the most proliferant uses of AI and machine learning.
Self-driving cars, such as Google’s Waymo project, use computer vision to pilot their cars safely on roads and avoid obstacles.
Airports use them to detect left baggage, and to detect security threats. They even use them to ensure that there are enough trolleys for you to use to push your luggages.
Law enforcement regularly uses facial recognition to fight crime. It is well known that China uses facial recognition extensively, and invests heavily into infrastructure around the country.
Goal achievement
Goal-achievement systems uses machine learning to learn in an environment by using feedback from its own actions and experiences. In other words, it uses rewards and penalties figure problems out and solve them.
The beauty of such systems, is that as the machine learning algorithm tries to maximize its reward, it is quite likely to seek out very unexpected ways of solving a problem.
The fundamental technique is a sub-field of machine learning called Reinforcement Learning.
How it’s being used in the world today
Goal-achievement systems have been used to play games really well. From acing Flappy Bird to beating chess and Go grandmasters. Even hugely popular games such as Dota, AI has been used to beat even the world’s very best players.
A very common place we see them is when we play games against a computer adversary. When you’re playing against a computer-controlled opponent, the computer may well be learning what you do, and maximizing it’s reward of defeating you. Of course not all of them use such goal-achievement systems, but a good number do.
Beyond games, such goal-orientated systems have been used in myriad other places, such as when bidding for auctions.
Moving forward with AI systems
Like we mentioned in the introduction, some of the most successful machine learning systems combine these capabilities to form a homogeneous system that is even more accurate and powerful.
At the end of the day, machine learning is simply an approach to problem solving, albeit one that is rather different from the rule-based approach that we’re used to, and one that heralds much potential due to its unique “learning” abilities.
The goal of a machine learning system is to solve a problem, to achieve an outcome that is helpful to us as humans. That may be solving your internal business process challenges, or predicting what’s going to happen in the market, maximizing our return on investment when doing advertising, or identifying threats and allowing us to intervene before something (bad) happens.
And the end of the day, these 6 capabilities will help you to understand how AI and machine learning systems conceptually work, and more importantly help you to think about what AI can do for you in a structured way.
If you see any challenge you or your business faces, and you recognize it falls into 1 of these 6 capabilities, AI may potentially hold a solution for you.
