The Neuroscience Behind Our Technology
The Thousand Brains Theory of Intelligence
Our AI technology is based on two decades of neuroscience research and breakthrough advances in understanding what the neocortex does and how it does it. We have developed a framework of intelligence called the Thousand Brains Theory and 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.
The challenge is to apply these discoveries to practical AI systems. By translating neuroscience theory to hardware architectures, data structures, and algorithms, we can deliver dramatic performance gains in today’s deep learning networks and unlock new capabilities for future AI systems.
Animation: How the Brain Works: The Thousand Brains Theory of Intelligence
Our Research Projects
Realistic Neuron Model
The world is constantly changing; therefore, our model of the world must continuously learn to reflect the changing world. In the brain, synapses on the dendritic branches of neurons give us this ability to learn something new without forgetting what we’ve previously learned – a property that is absent in today’s deep learning systems.
We have created a more realistic neuron model with active dendrites that can reliably recognize thousands of patterns, which represent different contexts for learning. Augmenting deep learning networks with the properties of real neurons enables them to learn continuously without any manual interventions.
In the brain, knowledge is stored in reference frames. They allow us to create an invariant and stable representation of an object or concept that we can navigate and manipulate. We have discovered that each cortical column establishes its own set of reference frames, and they create reference frames using cells that are equivalent to grid cells and place cells.
To be intelligent, a machine needs to learn a model of the world. The model should include the shape of objects, how they change as we interact with them, and where they are relative to each other. Incorporating reference frames into deep learning systems can enable intelligent machines that can learn quickly with fewer examples than traditional deep learning models.
We cannot sense everything in the world at once; therefore, movement is required for learning. A truly intelligent AI system requires embodiment, whether virtual or physical. It needs sensors and the ability to move them to learn, plan, and act.
We discovered that each cortical column is a complete sensorimotor system. With every movement, a column predicts what its next input will be and generates actions to explore or act on the world. Understanding and implementing sensorimotor models will allow us to build scalable intelligent systems.