Numenta founder Jeff Hawkins spoke recently at the 2016 Neuro-Inspired Computational Elements Workshop at UC Berkeley about the biological nature of intelligence and what it might mean for the development of artificial intelligence.
Palm Computing and Numenta founder Jeff Hawkins spoke recently at the 2016 Neuro-Inspired Computational Elements Workshop at UC Berkeley about the biological nature of intelligence and what it might mean for the development of artificial intelligence. The video is embedded below.
Hawkins is the author of On Intelligence , a book he wrote before founding Numenta, an artificial intelligence company that applies his biologically-inspired theories to algorithm creation. Hawkins actually prefers that his company not be biologically inspired, but biologically constrained. The book describes hierarchical temporal memory (HTM), an unsupervised to semi-supervised online machine learning model that models some of the structural and algorithmic properties of the neocortex.
HTM is a method for discovering and inferring the high-level causes of observed input patterns and sequences, thus building an increasingly complex model of the world.
In his talk, Hawkins rapidly covers how learning takes place in the brain though communication and feedback in the neocortical columns (pictured above), and at the synapses. "It is an extremely robust system in every single way," he states. "Neurons, synapses, dendrites -- this is an important property if you are ever going to build these things in hardware."
Hawkins then presents 'Hawkins' List' of how the biological parts of intelligence can be organized into their base functional elements:
Hawkins’ List of the Functional Components of Intelligence
1) Networks of neurons that learn and recall sequences (required)- Continuous learning, not batch (required) - Many simultaneous predictions (required)
- Robust (required)
HTM: active dendrites, synaptogenesis, no spikes
2) Regions that use sequence memory for:- sensory inference (required)
- sensory-motor inference (required)
- motor generation (required)
3) Hierarchy of regions (required)- number of regions (parameter)
- size of regions (parameter)
- connectivity graph (parameter)
4) Embodiment (required)- sensors (parameter)
- built-in behaviors (parameter)
- emotions/motivations (parameter)
- episodic/spatial memory (parameter)
"If you're going to build an intelligent system, it's going to have to have networks of neurons that learn and recall sequences.""If you're going to build an intelligent system, it's going to have to have networks of neurons that learn and recall sequences," states Hawkins. "That is its fundamental 'tissue' or premise that is has to use. All inference, almost all inference in auditory, visual and somatosensory is inference of sequences. All motor behaviour is playing back sequences. This is not something to be added to a system it is a core fundamental principle. "
Using his theory and applications of it at Numenta, Hawkins shared his two aspirational goals of brain-inspired computing. First, he would like to see a machine that acts like a super mathematician. "You could literally build a hierarchy that's designed where part of the behaviours are mathematical behaviours."
|Hawkins points to the need for a super smart Robonaut|
Hawkins doesn't think these goals will happen in his lifetime, but expects they will come to pass soon and his theories will be part of their creation.