Very little attention seems to be given to the architecture of the cortical messaging protocol. Aspects like topology of the network, routing, bandwidth limitations seem to be deprioritized – or dismissed? – as implementation details.
The original term “AI bus” for the CMP, an analogy to a computer bus, suggests a rigid, one-dimensional structure that suffers from bandwidth limitations and is a single point of failure. The brain doesn’t work like that. It’s a decentralized network, and it wouldn’t be able to scale to 100,000+ columns and show the resilience to damage if it did. How do columns participating in a vote find one another? Does voting work over long distances, directly or indirectly?
“We don’t want to create a better human” is Numenta’s stated goal, rightfully so, as a “better human” would be something similar to a human in the first place – vision, hearing, language and all that – and only then an incrementally better version of that. However the more interesting applications of machine intelligence lie in more radical departures from biological intelligence: Sensors that detect chemicals, an induction coil in the road that detects passing vehicles, switching valves or traffic lights as “motor” outputs or even placing an order on the stock market.
By far the biggest departure from the biological example – and therefore the most promising aspect of machine intelligence – is a machine’s ability to extend geographically. Biological intelligence in brains is always very localized, and “natural” communication happens at low bit rates (humans talk at 38bit/s), inhibiting its ability to form larger “hive minds”. This informs our intuition about intelligence.
But the Discord forum I’m posting this on here works part in my browser here in, say, Timbuktu and part in a datacenter in Toronto, both parts tightly knit together, essentially combining to form one “machine”. And I’m only using a tiny fraction of my $40 gigabit home internet yet.
If you want to build a warehouse full of robots that are all on the same AI bus, or a citywide traffic control system, or a network of algorithmic trading systems that play the markets in New York, London, Tokyo, then you will need a protocol that adapts and optimizes for limited bandwidth and network outages. The brain, presumably, has this protocol. How does it work?
