Plans to build the “world’s biggest spiking neural network” to mimic the brain were announced by Mountain View, Calif.-based Artificial Development at the Accelerating Change Conference on Sunday.

The CCortex system will be a “massive spiking neuron network emulation and will mimic the human cortex, with 20 billion layered neurons and 2 trillion 8-bit connections,” according to AD’s President and CEO Marcos Guillen, listed in the Guardian’s “The Young Rich” for his former position as Director of Red Internauta of Spain, valued at 29.6 million pounds.

The network will run on a 1000-processor supercomputer cluster operating at 4.8 teraflops, with 1.5 terabytes of RAM and 80 terabytes of storage, he said.

“Most neural network models to date have been based on the Hebbian network, a simplified version of the real neural networks based exclusively on connectivity properties between neurons,” Guillen said. CCortex adds a time-sensitive, analog representation of the shape of spikes that allows CCortex to “tune vast populations of neurons and the information they hold to complex spiking patterns, adding a much higher level of complexity to a highly realistic simulation” using a “database that has a representation of the layered distribution of neural nets and detailed interconnections in the brain.”

Some conference attendees were skeptical. “The major problem with CCortex’s estimates of the amount of CPU power needed to simulate the brain is that at 10^15 ops/sec. they drastically underestimate the amount of computation that a single neuron does,” Ramez Naam, CEO of Apex NanoTechnologies and author of the forthcoming More Than Human, told KurzweilAI.net.

“Neurons don’t just add up the inputs they receive. They have their own memory, they change gene expression and protein synthesis in response to previous activity, they’re very very sensitive to the timing of signals, and in response to all of this they behave in complex ways.

“Simulating a single neural spike accurately requires the mathematical solving of partial differential equations, which is quite computational-intensive. The lowest estimates that are credible with neuroscientists are 10^21 or 10^22 ops/second to simulate the brain, which may actually be low by many orders of magnitude.”

“It’s an interesting effort, but they seem to have no particular insight into the operation of the human brain and how it may give rise to intelligent behavior,” said computer scientist and AI expert Dr. Ben Goertzel. “Their focus so far has been on the mechanics of building up the distributed neural-net simulation. They have no neuroscientists involved with the project, just programmers.

“It seems very likely that their initial efforts will not lead to anything very interesting. But, afterwards, if they ally themselves with some folks with more insight into the neuroscience side of brain modeling, they might come up with something interesting or useful, who knows.”