Here’s why True AI Won’t Come From Deep Learning

true ai
Sergey Tarasov |

As promising as Deep Learning is, and despite its initial success stories, it might actually slow the development of true AI.

If you’ve been keeping tabs on our recent AI coverage, you are probably aware that AI is the umbrella under which exists many concepts and methods of development. In the end, these different but not exclusive approaches seek to bestow machines with more human-like reasoning and intelligence.

Deep Learning is an obstacle to the development of true AI.Click To Tweet

Deep Learning is one type of reasoning that AI platforms adopt. Deep Learning, as a representation of that concept, is receiving significant research efforts, investment, and even media buzz.

Deep Learning, the “True” Path to Human-Like Intelligence?

Since the dawn of computing technology, developers created programs and algorithms by writing code that machines translate into precise instructions.

Computers, however powerful and versatile they might be, are often times incapable of carrying out tasks that humans perform effortlessly. Complex problems that take real-time experience into account can’t yet be reduced into code lines, hence the need for more novel approaches.

Deep Learning proposes a different way of solving problems.

Instead of code-writing the way a program solves a problem, the program “learns” to solve it on its own.

That’s the broad concept behind Deep Learning, which relies on multi-layered neural networks, where each layer starts where the last layer has left off to solve the problem.

It should be noted that neural networks are separate but linked nodes that simultaneously run calculations, yet they only resemble our own neural system.

Some experts argue that Deep Learning isn’t the true path to create human-like intelligence for machines, and that it might be hindering “true AI” progress.

All Deep Learning systems currently in use are “supervised”–meaning they need pre-determined data that they will, basically, classify–which take up huge resources that otherwise would be directed to the development of AI with real potential, referred to as “unsupervised AI”.

In a not too dissimilar way than the human brain, unsupervised AI would recognize new patterns, label them on its own and classify them without human prior input. This is the “true AI” MIT scientists were referring to as “unsupervised”.

Big Corporations Make do With “Supervised” Deep Learning

Deep neural networks have been around since the 1960s, but their development has really only taken off in recent years when two conditions have come together: Big Data and computing power.

Deep Learning systems must have at their disposal large amounts of data, and sufficient computing power to continuously update themselves, learn from their experience and keep improving.

Thanks to Deep Learning, several data crunching applications are made possible, such as voice recognition, image recognition, and mapping.

These applications and many others represent for Google, Facebook, Apple, Microsoft, Amazon and the likes the stakes of deep learning.

Big corporations are already cashing on the constant flow of data, dedicating their colossal resources to supervised Deep Learning.

Besides some initiatives with timid progress (like Google’s Artificial Brain), no “unsupervised AI” big project has gotten off the ground.

Per MIT Technology Review, Quoc Le (one of Google’s Brain research scientists) has identified “unsupervised learning” as the biggest challenge to developing true AI that can learn without the need for labeled data.

Under what circumstances would you unleash unsupervised AI into your life?

banner ad to seo services page


  1. Surprised I do not see a comment just yet here. Even so, the writer makes a good point – deep learning isn’t the “unsupervised” learning we would expect in a human-like AI. I can see it as creating a potentially very subservient “thinking” machine but not one that was truly independent with thoughts & conclusions of its own not guided by the reins placed upon it. Though I might point out even humans have constraints from our environment, upbringing, etc. but we have the ability to imagine outside them and even come to turn and pull on those reins to suite our desires (aka manipulate our situation).

    • I thought I had responded to you earlier, Eli, but alas. I really like that you pointed out that we have to acknowledge whether we are truly “unsupervised.” Perhaps some of us are able to think for ourselves, but aren’t there people out there that challenge this idea?

      • Among humans it would seem to me that most fall under the “unsupervised” category. However, even then there is a sort of “supervision” that comes from the hole nature vs. nurture realm. We are each of use molded by who/what we naturally are as well as the environment that raised us. Granted, it is not an “on rails” sort of supervision but more like strings tugging us along the way in our lives. Of course we still have the ability to choose how we respond to such tugging but this is typically at least influenced by the previous “tugs” on us in life.

  2. There are certainly AI experts that argue deep learning isn’t the correct path to “true AI” (I reject their claims), but this article’s claim that “unsupervised learning” is what is missing, is not the argument those experts are making. In the fact, machine learning is a very old idea and neural networks have been a form of machine learning for 50 years. The problems run into during the 60’s was a lack of good unsupervised learning algorithms for use with neural networks. Due to the poor algorithms, neural networks were all “shallow” — few layers. Adding more layers didn’t improve their ability to learn,. There were no “deep” networks that worked any better than the sllow networks (like 3 layers). A key breakthrough in machine learning has more recently occurred with the work of people like Jeffrey Hilton and algorithms such as Restricted Boltzmann Machines (RBMs). These algorithms are 100% UNSUPERVISED. It’s the creation of these newer algorithms that have allowed neural networks with more layers to out perform the shallow networks. They first layers of the “deep” networks are being trained by unsupervised algorithms and these new “deep” networks created by new unsupervised algorithms, are breaking all the old performance records and generating all the new buzz about “deep learning”.

    The “deep” in deep learning, only exists because of new and improved UNSUPERVISED algorithms.

    So to suggest that deep learning isn’t a valid path because it’s a supervised algorithm, when the only reasons “deep” networks exist at all is because of advances in unsupervised learning algorithms, really misses the mark by a mile.

    The true arguments are far more complex than can be explained or debated in a short blog post.


Please enter your comment!
Please enter your name here