Numenta’s deep experience in theoretical neuroscience research has led to tremendous discoveries about how the brain works. We have developed a framework called the Thousand Brains Theory of Intelligence that will be fundamental to advancing the state of artificial intelligence and machine learning. By applying this theory to existing deep learning systems, we are addressing today’s bottlenecks while enabling tomorrow’s applications.
Deep learning systems have demonstrated impressive achievements but face significant scalability problems that cannot be solved simply by adding more data and power. A new approach is required, and the brain offers the best path forward.
How is the brain so efficient? It stores and processes information in a highly sparse manner. At Numenta, we understand sparsity, a foundational element of the Thousand Brains Theory. We’ve created a technology demonstration that validates sparse networks on inference tasks can unlock groundbreaking performance gains of more than 100x in deep learning networks with no loss of accuracy.
Numenta’s sparse networks rely on key aspects of brain sparsity, most notably: activation sparsity (number of active neurons) and weight sparsity (interconnectedness of neurons). The combination of the two yields multiplicative effects, enabling large efficiency improvements.
This proof of concept lays the groundwork for additional scaling benefits on inference tasks with sparse networks, and the same principles can be applied to training as well. Sparsity is only the beginning. As we add more elements of the Thousand Brains Theory, we expect to see additional benefits, including continual learning without batch training, unsupervised learning, and robustness.
Numenta Co-founder Jeff Hawkins’ new book, A Thousand Brains: A New Theory of Intelligence, tells the story of the discoveries that led to the creation of The Thousand Brains Theory of Intelligence. The book also covers how the theory will impact the future of machine intelligence, and what understanding the brain means for the threats and opportunities facing humanity. Richard Dawkins, who wrote the foreword, calls the book “brilliant” and “exhilarating”.
Written in three sections, A Thousand Brains covers the brain, AI, and the future of humanity. In this video series, Jeff gives a preview of the book and some of the key ideas in each section.
INTRODUCTION: A Thousand Brains
We are a team of scientists and engineers applying neuroscience principles to machine intelligence research.
To reverse-engineer the neocortex and to enable machine intelligence technology based on cortical theory.
We believe the brain is the best example of an intelligent system. The brain’s center of intelligence, the neocortex, controls a wide range of functions using a common set of principles.
Today’s machine learning and AI have accomplished many impressive tasks but are restricted to narrow, specific goals. In stark contrast, intelligent machines continuously learn patterns in their environment without supervision, enabling them to tackle problems in entirely new ways. Intelligent machines that learn will have an enormous beneficial impact in the coming decades, and the neocortex provides the blueprint for building them.
Numenta has developed a novel theory and broad framework for understanding what the neocortex does and how it does it, called the Thousand Brains Theory of Intelligence. We have built a robust roadmap that lays out the steps to achieving truly intelligent machines based on the key principles of the Thousand Brains Theory.
We are one of the few teams that has developed large-scale theories of the brain that are biologically constrained, testable, and implemented in software.
Our deep neuroscience research provides the foundation for delivering transformational changes in machine learning and AI. We document our research in several ways, including peer-reviewed journal papers, conference proceedings, and invited talks. In addition, we place our research commits in an open-source project. We strive to be completely open in everything we do.
Our neuroscience research has uncovered a number of core principles that are not reflected in today’s machine learning systems. These principles have the promise to solve many of the known problems today’s machine learning and AI systems face. By incorporating these principles, we can overcome today’s limitations and build tomorrow’s intelligent machines. Below are some of the topics we are currently researching:
Representations in the brain are highly sparse, resulting in an extremely efficient system. For machine learning, sparsity also offers the promise of significant computational benefits, but most hardware architectures are not optimized for extreme sparsity. These limitations have hindered research into sparse models. Along with our hardware partners, we are developing methods for dramatically improving the computational efficiency of sparse neural networks.
In the brain, cortical networks are sparsely connected and extremely dynamic. As many as 30% of the connections in the neocortex turn over every few days. We are investigating ways to create highly sparse networks that learn their structure dynamically through training.
Our neuroscience research has shown that sparse representations are more robust and stable than dense representations. We have developed a cortically inspired sparse algorithm that can be applied to deep learning networks trained through backpropagation. These networks have both sparse activations and sparse connections. Sparse networks achieve accuracy competitive with the state of the art dense models, but are significantly more robust to noise.
Truly intelligent machines must have the capability to learn and adapt continuously, a property that is absent in today’s deep learning systems. In the brain, sparse representations plus a more complex neuron model, enables us to continuously learn new patterns in an unsupervised manner. Incorporating these ideas into artificial neural systems can enable systems that learn continuously from streaming data without any manual interventions.
Long term, our focus is on enabling the creation of truly intelligent machines that understand the world. Future research projects may include:
Vernon Mountcastle proposed that because every part of the neocortex has the same complex circuitry, then every part is doing the same thing. Therefore, if we can understand a cortical column, we will understand the neocortex.
Through our focus on cortical theory, we seek to understand the complex circuitry of a cortical column. We want to understand the function and operation of the laminar circuits in the brain, including what the neurons are doing, what the layers are doing, how they interact, and how a cortical column works.
We are often asked how we do neuroscience research at Numenta and what it means to focus on cortical theory. We read many neuroscience papers and published results to get a deep understanding of neuroanatomy, neurophysiology, and cortical function. We propose theories about how the detailed architecture of the neocortex implements these functions. We ensure our work satisfies biological constraints and test our theories via simulation. Since we do not do any wet lab work, we collaborate with experimentalists that run investigations and give us feedback.
We document our research in several ways, including peer-reviewed journal papers, conference proceedings, research reports, and invited talks. In addition, we place our daily research commits in an open source project and answer questions about our research posted on our forums . We strive to be completely open in everything we do.
Although there is much to do, we have made significant progress on filling in the pieces of the common cortical circuit that underlines all of intelligence. We have proposed a novel theory and broad framework for understanding what the neocortex does and how it does it. We have made discoveries on how neurons make predictions, the role of dendritic spikes in cortical processing, how cortical layers learn sequences, and how cortical columns learn to model objects through movement.
Building on Mountcastle’s proposal that the neocortex is made up of nearly identical cortical columns, we are working on filling in the pieces of the cortical circuit – understanding each layer, what the neurons are doing, and how a cortical column works. As part of this research, we have proposed a framework, consistent with anatomical and physiological evidence, that explains how cortical columns function. The framework is based on how cortical columns learn sensorimotor models of the world by combining sensory inputs with location signals.
Find more information on cortical columns.
Memory and recall of sequences is an essential component of inference and behavior. We believe that sequence memory is occurring in multiple layers of the neocortex. We’ve shown how a layer of pyramidal neurons with active dendrites, arranged in mini-columns, will learn transitions of patterns and form a robust sequence memory.
Find more information on sequence learning and prediction in cortex.
Sparse Distributed Representations (SDRs) are a foundational aspect of all of our theories. Everywhere in the neocortex, information is represented by distributed and sparsely active sets of neurons. We have shown through mathematical analysis and simulation that SDRs enable semantic generalization and robustness.
Find more information on sparse distributed representations.
We have a growing collection of published peer-reviewed papers, supplemental white papers and research manuscripts. You can search our publications by category or by year. Some are currently under review at journals/conferences, but we have made all manuscripts freely available on preprint sites, such as arXiv or bioRxiv. Our goal is to document all of our discoveries in scientific journals.
HTM (Hierarchical Temporal Memory) is an algorithmic implementation of the Thousand Brains Theory. HTM builds models of objects and makes predictions using sensory input. It generates motor commands to interact with its surroundings and continuously test its predictions. This continuous testing allows HTM to update its predictive models, and thus its knowledge.
Learn more about our HTM technology, from application to forums, and more here.
Because we want our technology to be broadly adopted, we make it widely accessible in an open-source project. You’ll find our algorithms, source code, and our latest work on applying HTM to today’s machine learning platforms there. For example, we’ve created libraries to create sparse deep learning networks in nupic.torch.
Anyone is welcome to use our technology for free, under the AGPLv3 open source license. If you are interested in seeing, developing or working with our technology, you’ll first have to sign the Contributor License. For more on our licenses, see the Licensing & Partners section.
We have an active discussion forum with HTM community members covering a variety of topics. We welcome members of the HTM community who want to translate our documentation into languages other than English.
To help you learn about our theory and technology, we have created a number of videos, podcasts and educational resources. They are designed for anyone who wants to learn more about our cortical theory and machine intelligence research.
Numenta’s licensing and intellectual property (IP) strategy is to create an active research community and build a foundation for future intelligent applications. Unless it involves proprietary information regarding our work with other companies, we openly publish our scientific findings, software, intellectual property, and business strategy.
We have several software libraries available for anyone to use, along with the associated IP, under an AGPLv3 license at no cost.
If you are interested in seeing, developing, or working with our technology, you’ll first have to sign the Contributor License.
For scientists and researchers who want to use our intellectual property without our software, or whose work may be covered by our patents, we make a clear statement of non-assertion: as long as your work is for non-commercial use, we will not assert our patents.
We have a growing list of more than 40 U.S. and international patents that we believe will be critical for machine intelligence. The list of issued U.S. patents can be found here. In addition, we have pending U.S. and international patents that are not included in this list.
Commercial licenses are currently only available under the terms of the AGPLv3 license. Although Numenta previously offered non-AGPL commercial licenses, we have discontinued that practice. The work we were licensing is now many years old and we are not maintaining it. Our new technology, pertaining to machine learning, is not yet available for licensing. We will announce a commercial licensing program when available.
If you are interested in using our intellectual property without our software in a commercial application, or have questions about our patents and licensing, contact us at email@example.com.
We may be busy contemplating cortical theory, but we’ve got a work-hard, play-hard attitude. At the heart of the peninsula, our downtown Redwood City location is a short walk from the Caltrain station. When we are in the office, our kitchen is stocked with snacks, and we enjoy weekly catered lunches from a variety of local restaurants. Outside the office, we enjoy getting together for company events, happy hours, and other fun activities. In the past, we’ve cheered on the SF Giants, baked pies at Pie Ranch in Pescadero, and do-si-doed through the night.
In addition to our full-time positions, we are always looking for strong research candidates to join us through our research internships.
Numenta Work Environment
Numenta is based in the San Francisco Bay Area with a physical office in downtown Redwood City. We are currently requesting that employees work from the office on Tuesday and Wednesday. The remaining days, employees can choose to work from the office or from home. Although we favor employees who are based in the Bay Area, we can make exceptions. In those cases, we ask that employees travel to the Bay Area at least two times per year at specified times. The Numenta office is restricted to people who have been fully vaccinated against COVID-19. Prospective employees should be prepared to submit documentation of vaccination.
Donna is a serial entrepreneur best known for her work as CEO of Palm Computing and then Handspring, pioneers of the first successful handheld computers and smartphones. Previously, Donna spent 10 years in a multitude of sales, sales support, and logistics functions—both at Apple and at Claris, an Apple software subsidiary. She founded Numenta with her long-time business partner, Jeff Hawkins, in 2005.
Donna earned a B.A. from Yale University, and an M.B.A. from Harvard Business School. In addition to chairing Numenta’s board, she currently serves on the boards of Stanford Health Care (Palo Alto, CA), and Twilio (NYSE: TWLO). Donna also served on the board of Yale University from 2006-2018, including two years as Senior Fellow.
Jeff is a scientist whose life-long interest in neuroscience led to the creation of Numenta and its focus on neocortical theory. His research focuses on how the cortex learns predictive models of the world through sensation and movement. In 2002, he founded the Redwood Neuroscience Institute, where he served as Director for three years. The institute is currently located at U.C. Berkeley. Previously, he co-founded two companies, Palm and Handspring, where he designed products such as the PalmPilot and Treo smartphone. Jeff has written two books, On Intelligence (2004 with Sandra Blakeslee) and A Thousand Brains: A New Theory of Intelligence (2021).
Jeff earned his B.S. in Electrical Engineering from Cornell University in 1979. He was elected to the National Academy of Engineering in 2003.
Subutai brings experience across real time systems, computer vision and learning to Numenta. He has previously served as VP Engineering at YesVideo, Inc. where he helped grow the company from a three-person start-up to a leader in automated digital media authoring. In 1997, Subutai co- founded ePlanet Interactive, a spin-off from Interval Research. ePlanet developed the IntelPlay Me2Cam, the first computer vision product developed for consumers. He has served as a key researcher at Interval Research.
Subutai holds a B.S. in Computer Science from Cornell University, and a Ph.D in Computer Science from the University of Illinois at Urbana-Champaign. While pursuing his Ph.D, Subutai completed a thesis on computational neuroscience models of visual attention.
Christy brings more than two decades of technology marketing and communications experience to Numenta. Previously, she launched analytics programs for the Retail and Healthcare industries as the Global Product Marketing Director of Analytics at Actian. Christy held a number of software marketing roles during her 13 years at IBM, where she managed user groups, produced live demos and developed big data video tutorials. She was also one of the founding members of IBM’s thought leadership group: the IBM Institute for Business Value.
Christy holds a BA in Economics from Princeton University.
Ed has been a part of the core team of five Silicon Valley start-ups. Ed’s first big success was Radius, Inc. where he was instrumental in building products and the brand. After Radius, Ed was the first vice president of marketing for Palm where he helped develop the original Palm Pilot, the Palm brand and Palm’s product strategy. He moved on from Palm to found Handspring where Ed and his partners created the forbearer of all future smartphones; the Handspring Treo. Ed drove the transaction that reunited Palm and Handspring into a single Palm again. He established relationships with wireless carriers globally and drove Palm’s annual smartphone business to more than $2 billion. As the CEO of Palm, Ed spearheaded the transformation that created the WebOS platform and Palm Pre line of smartphones.
Ed now spends his time investing in and mentoring entrepreneurs. Ed is a board member of Numenta, Inc., Active Mind Technology, and POPS Worldwide, and is an investor and on the board of advisors of six other start-up companies. Ed holds a B.S. from the University of Oregon.
Mike is a successful serial entrepreneur. He has founded many companies with breakthrough technologies including FTL, a super computing company that merged with MIPs, Epigram, which was acquired by Broadcom, Rambus and Matrix Semiconductor, a creator of 3D integrated circuits.
Mike currently sits on the board of Rambus (NASDAQ: RMBS). He is participating on the Numenta board as an individual, rather than as a representative of Skymoon Ventures. Mike holds a B.S. degree in Mathematics from Purdue University and a Ph.D. in Computer Science from Stanford University.
In 2002, Dr. Harry J. Saal was chosen by the US Department of Justice to lead the Technical Committee charged with monitoring and enforcing the Microsoft Antitrust case, and he served as Chairman of the Committee through the May 2011 expiration of the Judgment.
Harry founded Nestar Systems, a pioneer in local area network systems, in 1978. In 1986, Harry became the founder and CEO of Network General Corporation, the first company wholly dedicated to the area of network diagnostics. From 1993 through 1995, Harry served as founding CEO and President of Smart Valley, Inc., a non-profit organization chartered to create a regional electronic community based on an advanced information infrastructure and the collective ability to use it.
Harry is active in philanthropy and community affairs, and serves on the board of the American Institute of Mathematics, among others. Harry holds a B.A., M.A. and Ph.D. in Physics from Columbia University.