Robert Hecht-Nielsen, Computational Neurobiologist (via videoconference) spoke on
The Fundamental Mechanism of Cognition.
Synopsis: Each cortical module in the celebral cortex is responsible for describing one attribute that an object in the mental world can have. These attributes are persistent terms of reference. For instance we can have the attribute of apple for a mental object. Hecht-Neilsen goes on to describe how pair-wise "knowledge links" are formed - for instance, apple can be unidirectionally linked to another neuron-collection of red. Hecht-Nielsen goes on to explain how we think (take inputs and respond with action) with this knowledge link model. He ends by discussing conversational interface research funded by the FairIsaac that employs 10B links to mimic a human conversationalist.
Robert Hecht-Nielsen titled the powerpoint, "Engines of Cognition: Cortical Modules"
Cerebral cortex is divided into modules - macroscopic, not like cortical module which is microscopic - large tissue. For a while people thought these didn't exist; does that mean don't exist? It means people didn't know what they were looking for, but now they do.
Each cortical module is responsible for describing one attribute that an object in the mental world can have. It describes it in terms of symbols. These symbols are formed in early childhood and they never are removed. You can't delete the symbol "David" - you can't do that. They're persistent. Symbols are durable, persistent terms of reference that must exist if knowledge is to be accumulated over long periods of time.
This is all the machinery of cognition that's needed.
A mental object can have an attribute of name, for instance: apple. Another symbol representing color red. Unidirectional neuron collection-to-neuron collection knowledge link between apple and red. Always pair-wise. Pairs of symbols.
How can these components think?
Each cortical module receives a thought control command input, which can cause the module to implement confabulation.
Some of these are receiving simulataneous knowledge links - very fast, highly parallel operation to determine which symbol is receiving the most activation. This "winner takes all" contest is called confabulation. Lightly excited symbols drop out - everything over in 100 milliseconds or less.
All of cognition is built on this operation called confabulation. [This has no relation to the better known definition of confabulation.] Not a digital control signal - it's analog like a muscle. The tension on the muscle must be precise.
Every time a confabulation yields a conclusion, then a set of action commands output. This happens immediately. These action commands are the behaviors we implement. Most of them are micro-behaviors, very small - but also include the macro.
The Alien Nature of Animal Intelligence: Vast quantities of knowledge are keys and billions of knowledge links are required. Thinking is exactly like moving (he uses the muscle analogy again). Learning is expensive (takes 1/3 of our life).
Language is the essential core of intelligence - that's why monkeys are stupid and we're not.
Fill-in-word example, sentence-continuation examples given.
FairIsaac has funded research on conversational machine called Vacations! mimics a human conversationalist, employs 10 billion knowledge links, exhibits powerful generalization to novel arrangements of familiar elements.
Sees a future called the "Confabulation Age". Sci Fi has already sold the vision. Envisions govt and academia playing minor role because not sufficiently nimble. Not within scope of resources for students or start-ups. Will be large IT-based companies ('elephants') as dominating.
Comments