Is AI research headed in the right direction?

artificial intelligence human brain
Image credit: Depositphotos

This article is part of Demystifying AI, a series of posts that (try to) disambiguate the jargon and myths surrounding artificial intelligence.

In 1956, researchers at Dartmouth College coined the term “artificial intelligence,” a field of science that aims to enable machines to replicate the capabilities of the human mind. AI pioneers believed at the time that in short time, “machines will be capable… of doing any work a man can do.”

But it didn’t happen.

For decades, AI scientists and researchers have been trying to recreate the logic and functionalities of the human brain. And for decades, they have dismayed themselves and the general public. Today, we’ve reached a point where artificial intelligence algorithms can solve very complicated problems, and in many cases with speed and accuracy that is far superior to those of humans. But whether contemporary AI works likes the human mind is up for debate.

In a blog post titled “The Bitter Lesson,” AI scientist Rich Sutton argues that the artificial intelligence industry owes its advances to the “continued exponentially falling cost per unit of computation” and not our progress in encoding the knowledge and reasoning of human mind into computer software.

“Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project, massively more computation inevitably becomes available. Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation,” Sutton says, adding that “the human-knowledge approach tends to complicate methods in ways that make them less suited to taking advantage of general methods leveraging computation.”

That short paragraph says a lot about the current state of artificial intelligence, but it needs a lot of unpacking.

The difference between AI and human brain hardware

Since early homo sapiens made their appearance some 500,000 years ago, the human brain wetware has not undergone any revolution. The size and computation power of the brain has remained fairly constant. But we have learned to improved our  cognitive capabilities by discovering new sciences and devising new techniques to transfer and teach human knowledge, combine human intelligence, and to tackle complexity by breaking it into smaller pieces.

More importantly, we were able to augment our intelligence by creating tools like computers, which Steve Jobs calls the “bicycle of the mind.” (Interestingly, an increasingly popular way of looking at AI is to perceive it as a complement to human intelligence, not a replacement.)

Contrary to human intelligence, artificial intelligence benefits from Moore’s law, which maintains that computing power doubles in power and halves in price every two years. Sutton generalizes the idea to “continued exponentially falling cost per unit of computation.”

Moore’s law has held for decades after Intel co-founder Gordon Moore made the prediction in the 1960s. While the speed of advances in computing power has slowed down in the past decade, it still beats the evolution of the human brain by a factor of several million (or maybe more).

Basically, what this means is that we must acknowledge the fundamental differences between the evolution of the human brain and AI. The human brain has been optimized to find techniques to use its unchanging infrastructure in new and innovative ways. AI on the other hand can be optimized to apply techniques that can scale with new and more powerful hardware.

AI’s learning and search capabilities

AI, Artificial intelligence, automation and modern information technology concept on virtual screen.

So what are the AI techniques that Sutton recommends? “One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning,” the AI scientist says.

The classic approach to developing artificial intelligence, also known as good old-fashioned AI (GOFAI), is based on creating rules for knowledge and behavior in the same way that (we think) the human brain works. An example of GOFAI is expert systems, an early AI technique that involved programmers and domain experts working together to code every single rule that goes into solving a specific problem.

This is an approach that works for many problems where the rules are few and the boundaries and logic are clear-cut and easy to codify. But in many domains, the rule-based approach quickly breaks, either because the rules are just too many and undefined, or the logic behind the way humans perform the task is just too complex to translate into computer rules.

Searching refers to computers’ ability to process and compare data and find specific information. Learning is the ability to find patterns and correlations in large data sets. Both are functions that computers can do more efficiently and faster as they’re given more computing power.

In contrast, the human brain is not very fast at processing information. Instead it relies on abstract thinking and reasoning to solve problems.

Examples of AI solving problems through search and learning

One of the examples of searching Sutton mentions to in his post is the game of chess, which scientists considered the drosophila of artificial intelligence for a long time. For decades, scientists tried to create software that tried to replicate the human way of playing chess. But in 1996, IBM’s Deep Blue, the computer that defeated world chess champion Garry Kasparov, used deep search, a technique that basically reviewed millions of moves and evaluated different outcomes before making a decision. Deep Blue did not win because it was using a very complicated and novel logic. Instead it used its massive computing capabilities to process large amounts of information at super-fast speeds, which allowed it to have decent response times when playing against Kasparov.

Learning comes into play in areas where rule-based AI has historically struggled, such as computer vision, the field of computer science that enables computers to make sense of the content of images and video. For instance, if you want to create an AI that can detect images of cats, you’ll have to develop software code that looks for curved, pointy ears, slanted eyes, fur, tail, paws.

Embedding that kind of behavior into software is extremely difficult, because there are various species of cats, they look different from different angles and under different lighting conditions and things can get complicated if they’re partially obscured. This is an example of a situation where the rules are not static and hard to define. This is exactly the kind of problem that learning can solve.

Deep learning, the current state-of-the-art AI technique, is the best example of the powers of learning. Deep learning algorithms use very little human-induced design. Instead, they rely massive computing power and huge amounts of data to develop their own behavior.

In the case of the cat-detector AI, instead of trying to teach the computer what a cat looks like, developers create a deep learning algorithm and provide it with millions of pictures labeled as “cat” and “not cat.” The algorithm runs the photos through artificial neural networks, which basically compare the pixels of the images, find common patterns and transform those patterns into very complex mathematical equations with thousands of dimensions. They then use those equations to evaluate new images and determine whether they have enough similarities to the cat examples they have previously seen.

“Early methods conceived of vision as searching for edges, or generalized cylinders, or in terms of SIFT features. But today all this is discarded. Modern deep-learning neural networks use only the notions of convolution and certain kinds of invariances, and perform much better,” Sutton says.

In the past few years, deep learning algorithms have helped bring great advances to fields such as cancer diagnosis, self-driving cars, face and voice recognition, online translation and more. They have also helped master some of the most complicated board, card and video games.

The concept of neural networks and deep learning has existed for decades, but it has mainly been thanks to the explosion in the availability of compute resources and data that it has become a reality in the past few years. Deep learning is another example of Sutton’s “methods that continue to scale with increased computation.”

Artificial intelligence can’t replicate the human mind

Artificial Intelligence

“[T]he actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries All these are part of the arbitrary, intrinsically-complex, outside world,” Sutton says.

There’s been a lot of debate about whether we will ever be able to create AI that can function like the human brain. In his post, Sutton dismisses such notions, stressing that computers should create their own intelligence, not mimic ours. “We want AI agents that can discover like we can, not which contain what we have discovered,” he says.

There’s a historic precedent to back Sutton’s arguments. In creating tools, we often get ideas from nature, but end up taking a different course. An interesting example are airplanes, which were inspired by birds, but work in a different manner. Neural networks and deep learning also take cues from some of the workings of the human brain, but also work in fundamentally different ways.

The divide on artificial intelligence

Rich Sutton is a longtime proponent of reinforcement learning, the subset of deep learning that helped AI conquer the ancient game Go. Therefore, it’s natural to expect Sutton to have a bias toward this blend of artificial intelligence.

But not everyone subscribes to Sutton’s view on the current state of AI. A few days after Sutton published his post, Rodney Brooks, a pioneer in AI and robotics, wrote an essay titled “A Better Lesson,” in which he debunked some of the claims Sutton makes and outlines some of the fundamental flaws that deep learning models suffer from. The entire post is very short and a worthy read, but here are some of the key ideas Brooks touches on:

  • Neural networks’ lack of understanding the context of images make them prone to making stupid and dangerous mistakes.
  • Deep learning hasn’t obviated the need for human knowledge and logic. It has just shifted the human mind labor into designing networks instead of hardcoding logic.
  • The amount of data and computation needed to train deep learning models makes it hard for individuals and small organizations to take full advantage of them.
  • The slowing of Moore’s law is creating a major hurdle for Sutton’s premise that we should leverage larger computing power to enhance AI models. Also, more computing power means more power consumption and a larger carbon footprint.
  • Specialized AI hardware is helping make up for the lack of computing power required to run deep learning models, but also makes AI models very rigid. Instead of using general-purpose hardware to develop and run AI models, engineers will be forced to create specialized hardware for each new AI application. Even then, we need human knowledge to design and create those hardware.

Deep learning has many critics, including New York University professor and Gary Marcus, who has written an in-depth critical appraisal of deep learning and regularly writes about the shortcomings of neural networks and connectionist AI models.

Deep learning has distinct limits that many AI experts and scientists regularly point to:

  • Deep learning algorithms are data hungry need a lot of training data to perform the simplest of tasks.
  • Deep learning algorithms are good at performing specific tasks, terrible at generalizing their knowledge and carrying their capabilities to other domains. A new problem domain usually requires a new model and training from scratch.
  • Neural networks, the underlying structure of deep learning algorithms, are hard to understand. Often, even the creators of deep learning algorithms are hard-pressed to make sense of their innerworkings.

To be clear, this is not the final word on deep learning and the current state of artificial intelligence. In fact, other AI experts debunk the criticism thrown at deep learning. Talks over the limits and capabilities of deep learning models often trigger heated debates between the best path to create efficient AI.

But what many experts agree on is that artificial intelligence will require the combination of many different techniques and approaches, including neural networks, neuroscience and GOFAI.

“I think a better lesson to be learned is that we have to take into account the total cost of any solution, and that so far they have all required substantial amounts of human ingenuity. Saying that a particular solution style minimizes a particular sort of human ingenuity that is needed while not taking into account all the other places that it forces human ingenuity (and carbon footprint) to be expended is a terribly myopic view of the world,” Brooks concludes in his post.


  1. At Daisy, we’re excited about A.I.’s potential but the landscape is still confusing as people try to determine how to define A.I. and how it can be successfully leveraged.

    We see reinforcement learning as the “True A.I.” The ability for reinforcement learning to continually improve through trial and error will have impact on how a business operates.

    Reinforcement learning also uses simulations to predict future behaviour, as well as the past. This is what separates it from predictive analytics, which uses historical information.

    Here’s a recent blog post that did about reinforcement learning:

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.