Deep Learning: A ‘Type’ of AI?

AI applications are everywhere. Some are ‘based on machine learning’. Some ‘use deep learning’. Other ‘types of AI’ exist as well. But terms like ‘AI’ and ‘deep learning’ can mean different things to different people. What do these terms mean, and how are they related to each other? Most sources that clarify the definitions of these two terms (AI and deep learning) explain that deep learning is a subset of AI — a special kind or flavor of AI, if you will. A Venn diagram representation of these interpretations is not uncommon.

Artificial Intelligence is the broadest (and oldest) of the concepts in this diagram. Some people say AI started with Turing's work in 1945; others may even claim it started with Euclid c. 300 BCE! Whatever the origins may be, AI is the attempt to get machines to behave intelligently.

Data science is then a specific kind of AI, using data in various ways to predict and control outcomes. A particular subset of data science is now known as machine learning: computer algorithms that can teach themselves (using the data provided to them) and improve their performance as they are fed more and more data. And finally, deep learning is a special kind of machine learning. The Venn diagrams describing these categories of AI can become even more complex, with finer distinctions.

These diagrams are not necessarily wrong, of course. Nevertheless, especially if you're thinking of building your own AI tool, I would suggest looking at things a bit differently.

Instead of thinking of deep learning as “a subset or a type of AI”, let's look at it as “a tool for AI”. This change in approach can help us appreciate that “AI” is more than a group of different technologies, it is something we create using one or more different tools. Deep learning is then to the AI practitioner what the hammer is to the master woodworker. Useful, even necessary; but a hammer is not wood craft. Wood craft is sometimes created using a hammer, along with many other tools.

Just like the master carpenter must be well acquainted with all the specifics of his different tools, the AI practitioner has to be comfortable with all of hers. She has to know the exact capabilities and limitations of each of her tools, to be able to use the right one at the right place in her workflow. It's also worth appreciating that this workflow we're talking about is going to be different for every project. Every carpentry project isn't going to “use the hammer” as its third operation, and designing an AI workflow also demands more ingenuity and care than just following a predetermined order of chained algorithms.

This means that to get the most out of deep learning, an incredibly powerful tool in our arsenal, we need to use it with more finesse than just throwing it at any problem that comes our way. A case in point is AlphaGo — the AI application that beat Go master Lee Sedol in 2016. AlphaGo was the first computer program to play the game of Go at a championship level. Examples like these are quite the testament to how powerful AI tools really are. Tools, though, not tool. Did AlphaGo use deep learning? You bet it did! Did it use only deep learning? Not a chance!

The above block diagram depicts AlphaGo's basic algorithm. The algorithm does use deep learning in various places. But AlphaGo is driven by an algorithm called Monte Carlo tree search, a strategizing algorithm which uses various deep learning algorithms as “advisors”.

Here is the clincher: each of the four deep learning algorithms shown above is capable of playing Go on its own. How strong do you think the deep learning algorithms are when taken alone? Left to their own devices, the four deep learning algorithms manage to play at a very amateur level! But a committee of such weak deep learning players, with the non-deep-learning Monte Carlo tree search sitting at the head of the table “discussing strategies” beats the human world champion! (Intrigued? We will be describing AlphaGo and game playing AI in a future blog post. Stay tuned...)

The AlphaGo example clearly shows how deep learning acts best as a tool in a larger work flow. Another example is the upcoming technology of self-driving cars. Self-driving cars are not fully functional yet. They're almost there, but a painful truth about technology is that what seems ‘almost there’ can sometimes take a whole lot of time and effort to actually get there. We don't know the time and effort the future holds for this particular achievement. What we do know is that whatever the final algorithm may be, it won't exclusively use deep learning. The algorithms being tested by autonomous car companies today are a mixture of deep learning, statistical inference, data fusion, path prediction, geometric path planning algorithms and even expert systems.

In fact, just as in the field of playing board games, pure deep learning algorithms drive cars at a very amateur level. The genius of the AlphaGo developers and the autonomous driving engineers is figuring out where exactly to incorporate deep learning in a complex workflow of diverse AI tools.

Our own experience at Oneirix mirrors these observations. We've delivered many AI algorithms: for medical devices, consumer applications, enterprise applications and more. We've used deep learning often, but never has a pure deep learning approach worked the best! Don't get me wrong, deep learning is a powerful tool. It has often been an irreplaceable part of our AI algorithms, when used in the right place. But the trick is in knowing when (and when not) to use it!

This series of articles on the details of AI algorithms will help the reader get acquainted with the specifics of the different tools available in her AI tool chest. And we already know that's step one in becoming the ultimate AI craftsman...

Like, share or comment on this article: