A team of computer scientists and mathematicians at Palo Alto, CA-based Artificial Development are developing software to simulate the human brain’s cortex and peripheral systems.
As a first step along the way, the company recently disclosed that it has completed the development a realistic representation of the workflow of a functioning human cortex. Dubbed the CCortex-based Autonomous Cognitive Model (‘ACM’), the software may have immediate applications for data mining, network security, search engine technologies and natural language processing.
The first ACM computer ‘persona,’ named ‘Kjell’ in homage to AI pioneer Alan Turing, was activated last month and is in early testing stages. CCortex, Artificial Development’s high-performance, parallel supercomputer, runs the persona simulation.
Marcos Guillen, president and CEO of Artificial Development, disclosed the ACM’s emerging cognitive capabilities at a recent Workshop on Cognitive Systems, co-hosted by Sandia National Laboratories and the University of New Mexico.
Guillen commented that the ACM is intended as a test-bed for future models and is still incomplete. While the Kjell persona uses a realistic frontal cortex, motor and somatosensory areas, it still lacks the visual and auditory cortex areas, two of the most important cortical structures. Other structures, such as the hippocampus, basal ganglia and thalamic systems, are still being developed and are unable to perform most normal functions.
The ACM interacts with trainers using a text console, reading trainer’s input and writing answers back, similar to a conventional ‘chat’ program. The ACM is being trained with a stimulus-reward learning process, based on classical conditioning rules. It is encouraged to respond to simple text commands, associating previous input with rewarded responses.
The ACM uses the associative cortex to ‘evolve’ possible antagonistic responses. Large populations of neurons compete for their own associated response until the strongest group overcomes the others. The ‘winner’ response is then tested and rewarded or deterred, depending on its validity. The ACM takes into account new experiences and uses them to modify the equilibrium between the responses and the strength of the associate neural path. Thus it creates a new neural status quo with more chances to generate accurate responses.
The process is mediated by a scaled-down version of the hippocampus and the basal ganglia, and can occur up to 20 times per second.
‘In our model the frontal cortex acts like an evolution chamber,’ said Guillen. ‘Possible responses compete on the associative cortex, rounding the ‘votes’ of associated neurons. When a winner finally emerges, it takes acting control of the motor response. The response is then tested, and the neurons take note of the result for future ‘voting.’ Ultimately, only the best responses survive the process.’
CCortex itself is a system intended to mimic the structure of the human brain, with a layered distribution of neural nets and detailed interconnections. It closely emulates specialised regions of the human cortex, basal ganglia, thalamus and hippocampus. CCortex runs on a high-performance, parallel supercomputer, a Linux cluster with up to 500 nodes and 1,000 processors, 1 terabyte of RAM, and 200 terabytes of storage.
With 20 billion neurons and 20 trillion connections, CCortex is up to 10,000 times larger than any previous attempt to replicate, partially or completely, primary characteristics of human intelligence, and is claimed to be the first neural system to achieve a level of complexity rivaling that of the mammalian brain.