From time to time we’re asked what the difference between Artificial Intelligence (AI) and Machine Learning (ML) is. There is no clear-cut definition, but this is how we (and many with us) see it.
The short answer is that Machine Learning is a way of accomplishing Artificial Intelligence.
Artificial Intelligence is usually divided into two categories. Weak, and Strong AI. Weak AI is a system or computer program that can solve a narrow set of problems with some level of perceived intelligence. An example could be a system that plays chess or Jeopardy! at or above human expert level, or that can categorize images of different kinds of objects really well. There are plenty of examples of this today.
The other category, Strong AI, is a system that is of superhuman intelligence in a broad range of tasks, or ultimately, in every aspect. This has, to common knowledge, never been achieved, and many industry experts and researchers fail to agree if this is possible, and in that case in how distant a future it is likely to happen. Some believe it might be possible in 20-30 years time, others say 50, and some say it will not be achievable in less than a 100 years.
Whereas Artificial Intelligence is the abstract idea of a system (e.g. a computer program) that behaves seemingly intelligently, Machine Learning is a way for a computer program to learn how to make intelligent decisions based on historic data, instead of being explicitly programmed.
Let’s apply this in our own industry with a simple example. Assume that a user reads a lot of culture related articles and never or rarely reads about sports. Then this fact suggests that it might be a good idea to recommend culture related articles and not sports related articles in the future. This way of making use of information about the past to predict the future, just like humans do, is what machine learning is all about.
The recommendations of course becomes much more interesting when you couple the information about the individual user with what interests people at large, and what people with similar interests are reading right now. It’s a bit like how a good librarian can recommend books based on what you have read before, what is generally popular at the moment, and what other people with similar reading habits likes. This can of course be done at an individual level by a human, but when it comes to making millions or billions of recommendations per week to millions of different people it stops being feasible to use humans to provide recommendations. Computers on the other hand are great for this. They only need a split second to provide tailored recommendations for you, and they have no problems keeping track of the current reading habits of millions of users at the same time. That is what makes machine learning for article recommendations so exciting!
The alternatives to this would be to manually write specific rules for each and every user or group of users that interact with your product. With many millions of users this would hardly be plausible.
So what does it mean that Machine Learning systems learn from the data? And why is it a big deal? Early attempts to create AI made use of hand written rules and decision trees. It was a slow, error prone and very costly process that required many domain experts to encode their knowledge in the system. Simply put, it didn’t scale and the result was often mediocre at best. Learning from data, i.e. from looking at records about the past in order to predict the future, scales comparably very well, since all you need is data to learn from and processing power to update the machine learning model. It is a completely automatic process. Processing power is quickly becoming cheaper and cheaper, data is becoming more and more abundant, and machine learning algorithms are quickly becoming more accurate and efficient. Hence the opportunities to use machine learning instead of manually writing rules are also becoming greater and greater.
This is why these are such exciting times for publishers. Machine learning really allows them to find the best and most relevant content for any given user at any given time, in order to offer them the best possible experience. This is of course is equally true for native advertisers, which are essentially in the same business of finding the right reader to the right piece of content.
At Strossle we are committed to pushing the boundaries of how well article recommendations can be made with machine learning, and we will not rest until every user is truly amazed by how relevant and inspiring recommendations they are provided on our customers’ websites.
Fredrik Skeppstedt, Strossle