Richard Granger, professor of psychological and brain sciences at Dartmouth, discussed the architecture of the brain and how it is being used to build learning-capable robots at Friday’s Jones Seminar.
Granger compared the neural structures of human brains to circuits, describing them as intrinsically parallel. Each component of brain circuitry has a separate path for current. In this respect, brain circuits are more mechanically advanced than those used in modern engineering. But by other comparisons, brains are technologically inferior to man-made circuits. Granger emphasized that the brain’s neural circuits were slow (operating on a millisecond scale), low precision (2-3 bit), and sparsely connected.
Rather than increasing the performance qualities listed above to improve functionality, the brain evolves by simply adding more circuits to existing structures. From the lowest mammalian species to humans, specialized neural regions for regulating bodily functions, described as the “reptilian brain,” remain almost unchanged.
The main difference is the size of the neo-cortex, a non-specialized region. This structure grows without bounds over the reptilian brain. Changes in its scale create new thresholds for neural processing capacity.
The thalamocortical loop may also be a key element in understanding cognitive abilities, Granger said. The loop, which connects the thalamus and the cerebral cortex, makes use of excitatory and inhibitory cell structures in order to regulate sensory data and produce hierarchies of received information.
The process starts at the eye, which sends sensory data into the thalamocortical loop. An inhibitory cell subtracts the information from the input. New data can now enter the loop via the eye without the original tag. This allows objects to be sorted into sequences of categories as the brain creates labels for each sensory aspect of a perceived object.
After this process occurs for the first time, the brain is able to not only recall memories of an object, but also to categorize and generalize. Granger described how learning requires this process of generalization so that the human brain can recognize members of categories that, while not exactly the same, still belong to the same group, such as individual cars.
The hierarchical clustering used by the brain has given rise to algorithms that can be used by robots.
“Brain circuits provide non-standard engineering approaches to intractable tasks,” Granger said.
For instance, vision algorithms derived from brain circuitry give machines enhanced visual-recognition through detection of physical relationships between objects just as the brain does with its hierarchy-building loops.
Robots now have higher-output functions, such as recognizing spindly or rotating objects that have been difficult to achieve with conventional algorithms. With these algorithms, they can even recognize distorted or changed versions of objects such as different fonts of letters.
Thus, by understanding brain circuitry, new robots with logic processing systems are being built that use categorization by relationship perception rather than holistic recognition. These may be the first steps towards development of robots that are capable of human thought.