New artificial intelligence project seeks to go beyond neural networks

An ambitious new endeavor called the Thousand Brains Project aims to develop a new artificial intelligence system that, according to its founder, will operate on the same principles as the human brain, but will be fundamentally different from the principles underlying the deep neural networks that dominate artificial intelligence today. With funding from the Gates Foundation, the open-source initiative will collaborate with electronics manufacturing companies, government agencies and university researchers to explore potential applications for its new platform.

In modern artificial neural networks, components called neurons receive data and interact to solve a problem, such as recognizing images or predicting the next word in a sequence. Neural networks are called "deep" if they have multiple layers of neurons.

Deep neural networks currently match or exceed human performance in many tests such as skin cancer detection and complex games, but they face a number of challenges. For example, as they grow in size and power, they become increasingly energy-intensive - to train OpenAI's GPT-3 in 2022, the company spent $4.6 million to run 9,200 GPUs for two weeks, according to a Nature study. Neural networks are also often unstable: small changes in the data they receive lead to wild changes in the results. For example, a previous study showed that changing a single pixel in an image can cause the AI to mistake a horse for a frog.

To overcome these challenges, the Thousand Brains Project aims to develop a new AI platform by re-engineering the neocortex, which accounts for about 80 percent of the mass of the human brain.

"Today's neural networks are based on fundamental neuroscience from 80 years ago. We've learned a lot about neuroscience since then, and we want to use that knowledge to develop artificial intelligence," says Jeff Hawkins, one of the inventors of the Palm Pilot in the 1990s. Hawkins is co-founder of Numenta, an artificial intelligence company in Redwood City, California, that launched the Thousand Brains Project June 5 at Stanford University's Human-Centered Artificial Intelligence Conference.

Objectives of the project "A Thousand Brains

The project's name is inspired by the structure of the neocortex: it consists of thousands of so-called cortical columns, each of which is divided into several layers of neurons. "There are about 150,000 cortical columns in the human brain, and each of them is essentially its own learning machine," Hawkins tells IEEE Spectrum.

Deep networks essentially generate a single model of the world by processing data step by step from simple features to complex objects, Numenta researchers claim. In contrast, the company's "thousand-brain theory of intelligence" suggests that multiple cortical columns generate multiple maps of the world, as if each human brain were actually a thousand brains working in parallel at the same time.

"Once we learn how to build one cortical column, we can build as many of them as we want," Hawkins says.

The goal of the project is to mimic this neurobiological structure in AI by using multiple cortical columns, each of which can perform a sensorimotor task, such as controlling a robotic finger. These units can then communicate with each other using connections that look a lot like the long-range connections observed in the neocortex. Hawkins believes the modular structure will make his approach easily scalable.

"The human brain has grown very rapidly through evolution by repeatedly replicating the cortex, and we hope to be able to do the same," Hawkins says.

The role of sensorimotor learning

The project is also based on the role of the neocortex in sensorimotor learning. While deep neural networks currently learn from giant static databases, the neocortex learns dynamically: It perceives its surroundings using the senses, learns how objects work using body movements, and builds models of the world based on sensory and motor feedback.

Hawkins argues that the difference between AI strategies is stark. Creating and labeling the datasets on which deep networks learn is expensive and time-consuming, and these systems cannot continuously learn on new data; instead, they have to retrain on the entire database. In contrast, the neocortex is able to learn actively and adapt quickly to new data.

"We can create machines that work similar to the neocortex in terms of sensorimotor learning, and yet they will be inherently robotic," Hawkins says. "I think our work is the future of not only AI, but of robotics."

In addition, the project is developing AI based on the neocortex's use of a coordinate system. In the mammalian brain, so-called place cells help encode location memories, and grid cells help map location in space. The neocortex uses these reference frames to store and understand the constant stream of sensorimotor data it receives.

"The way the brain structures data in two- and three-dimensional reference frames mimics the structure of objects in the real world," Hawkins says. "When you look at deep networks, they don't understand the nature of the world, so if you just change one tiny feature of an image, they often don't recognize it. In contrast, reference frames can help the brain understand how its models of objects might change under different conditions."

According to Hawkins, potential applications for this new AI platform could include sophisticated computer vision systems that can use multiple cameras to understand what is happening in a scene, or advanced sensor systems that help robots manipulate objects. "The Gates Foundation is interested in sensorimotor learning for global health. Think of an ultrasound that moves a transducer in space to build a model to visualize a fetus, for example. It's essentially a sensorimotor task."

The Thousand Brains open source project is developing a software development kit so that others can use its work. The initiative also pledges not to assert its patents related to the Thousand Brains approach.

Funding from the Gates Foundation

The Gates Foundation will give at least $2.69 million dollars over two years to the Thousand Brains Project. (The Gates Foundation declined to comment for this article). "We also hope to soon announce agreements with government agencies around the world," Hawkins says.

The goal of the project is to create a complete software version of the cerebral cortex, and then connect multiple units for a complex process such as vision or hearing. "Then we want to combine modalities - say, combine vision and touch - and finally build hierarchies, with a model of the world made up of objects that are made up of objects, and so on," Hawkins says.

While the Thousand Brains Project focuses on creating software, it is collaborating with researchers such as John Shen, a professor of electrical and computer engineering at Carnegie Mellon University in Pittsburgh, who is developing hardware based on the Thousand Brains concept.

"Since the first neuromorphic hardware was created to mimic the brain, neuroscience research has made significant progress in understanding how the brain works and how it is organized," Shen says. "We want to take the best of what we know about neuroscience and combine it with the best advances in silicon to create a new kind of computer. We've really embraced the thousand-brain theory, and we're talking to chip companies this summer to see if they have any interest in partnering with us."