Numenta recently launched an open-source research initiative called the Thousand Brains Project. Today, I sat down with Viviane Clay, Director of the Thousand Brains Project, to learn more about the guiding philosophies behind the initiative and the latest updates.
Q: Hi Viviane, to kick things off, can you tell us a bit about your background, current role and what first sparked your interest in AI and neuroscience?
Sure! I started studying cognitive science because I had a broad interest in how our brain, and intelligence in particular, works and this study program seemed to look at that question from many interesting angles. Not only did it look at what we know from neuroscience and psychology but also examined what is required to implement intelligent behavior in an artificial system at a more computational level. During my PhD, I moved more into the direction of AI research. I was driven by the idea that you only truly understand something once you know how to build it yourself. I was trying to figure out the principles needed to build a truly intelligent system.
One key principle I focused on during that time was sensorimotor learning and the effect that learning through movement has on what is being learned. I looked at this effect in both humans and machines and concluded that the way we learn (i.e. through interaction with the world) has a huge effect on the representations we form. It was then that I came across Numenta’s research which aligned really well with what I was working on, so I reached out to them and we ended up collaborating during the last part of my studies. After completing my doctorate program, I joined Numenta full-time and have been happily working here ever since!
I started out as a researcher when we first started working on the Thousand Brains Project idea almost three years ago. After a while I started leading the project and this summer, when we decided to scale the project and make it open-source, I transitioned into the role of director for this new initiative. It’s been an exciting journey watching the project grow and take on a life of its own!
Q: For those who might not be familiar with the Thousand Brains Project, can you give us a brief overview?
The Thousand Brains Project is part of Numenta’s broader vision that neocortical principles will drive the creation of machine intelligence. Instead of applying individual principles to existing AI systems (as we have done with sparsity and dendrites), this project starts from the ground up. We are building a whole new framework for a new type of AI, designed with neocortical principles from the start. It is built in a fundamentally different way, focusing on sensorimotor learning, internal models structured by reference frames, and a repeatable modular structure, as the key building blocks. We believe that developing genuine human-like intelligence demands a bold, innovative departure from the norm, not just incremental improvements upon existing algorithms and benchmarks.
Over the past two years we have developed a new framework that we believe can be the basis for this new type of AI, one that works on the same mechanisms as our own intelligence. Of course, there are still many open questions and the code is in an early beta version, but we are so excited about this approach that we want to share it with the community and gather feedback and input from researchers worldwide.
Thousand Brains Project Explainer Video (1:53)
Q: The project recently announced its first code implementation, can you tell us a little more?
Yes, we recently published all of our code under an MIT license, which is quite permissive. Anyone can check out the code, test it in their applications, contribute to it, or even use it commercially. It’s still in a very early beta version, which means that if you plan to use it for a particular application, you will need to do significant engineering work yourself. It is not an out-of-the-box solution yet, and there are still many capabilities we plan to have in the future that have not been implemented.
That said, we’ve spent the past few months putting together extensive documentation and other resources to make it easier for people to get started and to learn more about our approach. The code is functional, and with a bit of time and effort, you can do some cool stuff with it already! We hope that people who are interested will help us make the code as user-friendly as possible, contribute ideas or even build awesome demos with this beta code base.
Q: What was the primary motivation behind making this project open source?
Numenta has a long history of open-source projects, and we’ve had great experiences with it. By making this project open source, we are hoping to return to that practice of making our research very public and generating interest and excitement around our approach. We are also hoping to get early input and feedback from the research community and from those who have real-world challenges that they are trying to solve. This collaborative approach will help us refine and improve our framework, ultimately enabling us to build a great sensorimotor learning AI platform.
Q: There seem to be other research initiatives that are also working towards the same goal of achieving AGI. What makes the Thousand Brains Project different?
While there are several AGI-focused research initiatives in the AI space today, our project is distinct in its approach. Most AGI efforts rely on scaling up existing architectures like large language models in hopes that increasing data volume or model size will eventually lead to human-like intelligence. We believe that this approach is valuable in specific domains but has inherent limitations. In contrast, the Thousand Brains Project is grounded in a sensorimotor framework of intelligence, the key to how humans interact with and understand the world. It is a fundamentally different approach that will unlock potential that is unreachable with today’s solution, no matter how much you scale them. After all, the brain is the only truly intelligent system we know.
How the Brain Works: The Thousand Brains Theory of Intelligence (2:36)
Q: What do you think are the most exciting new industries the Thousand Brains Project will support in the near future?
With a huge innovation like this, we are not aiming for one specific application. We are planning to transform the entire field of intelligent systems and all its applications by creating an incredibly flexible, versatile, extendable, and scalable framework. Just as an operating system or programming language supports countless applications, this new framework will provide the tools necessary to address many of today’s current problems as well as completely new and unanticipated ones without being specific to any one of them.
Some areas that we currently think the system will be useful in are any applications where internet scale data for training is not readily available or where quick adaptation and extrapolation of knowledge are necessary. Any area where you have moving sensors that need to quickly learn how to model incoming data and make decisions and actions based on it would be a good application for our system.
Q: What type of contributors are you looking for? And what kind of collaboration opportunities does this project offer?
People who are interested can contribute at many different levels. The most direct way is to contribute to our code base. Since our code is open source, we welcome anyone who has ideas on how to improve our implementation or who wants to pick up one of our existing issues or tasks and contribute to it. Another way someone could contribute is by helping us improve our documentation and educate others about this approach. I, for instance, am so deep in this material that it is hard for me to explain it to someone who is completely new to it. Other people may have some great ideas for metaphors that helped them understand or tutorials that could give a nice learning experience.
Even if you are not an engineer or researcher, there are plenty of ways to stay engaged and support our project, such as posting questions and ideas on our discourse server. We also welcome researchers who write scientific publications and include our approach in one way or another. People who want to build applications on top of our framework are encouraged to share these with us so we can include them on our project showcase page and they can inspire others and convey all that is possible.
Lastly, simply sharing our project with friends or colleagues who might be interested is a great way to help us! We are also open to closer collaboration opportunities with other research labs or application teams. If you are interested, you can contact us directly by email.
Q: For developers and researchers new to this project, what resources would you recommend to help them get started quickly?
The best place to get started is our documentation, which covers every aspect of this project in detail. It outlines the project’s vision and core principles, how to run the code, how to develop the code, how the algorithm works on a more detailed level, how to contribute, and much more.
Another place to get more information is our YouTube channel. Here we post our past and present research meetings and more succinct presentations on specific topics.
We are also having a meetup on December 4th where we will give a quick-start overview of the project and the algorithm. And then, if you are like me, you probably also want to have a look at our code to get a more in depth understanding of how everything works, so definitely check out our GitHub repository.