A Skeptical Reaction To IBM's Cat Brain Simulation Claims

kreyszig writes "The recent story of a cat brain simulation from IBM had me wondering if this was really possible as described. Now a senior researcher in the same field has publicly denounced IBM's claims." More optimisticaly, dontmakemethink points out an "astounding article about new 'Neurogrid' computer chips which offer brain-like computing with extremely low power consumption. In a simulation of 55 million neurons on a traditional supercomputer, 320,000 watts of power was required, while a 1-million neuron Neurogrid chip array is expected to consume less than one watt."

Brain Power

[Trevin]

The cat's brain is made up of 1 BILLION neurons and 10 trillion synapses. So with the nuerogrid chips, it will require at least a kilowatt to simulate.

Re:long ways to go yet

[slashchuck]

... A real cat's brain also fits inside a tiny furry space the size of a baseball...

The brain size of the average cat is 5 centimeters in length and 30 grams. [wikipedia.org]

Re:long ways to go yet

[Anonymous Coward]

More than this, their simulated neurons aren't anywhere close to the real thing. A real neuron, an individual cell, has tremendous computing power due to the distribution of a bunch of different ion channel types (active conductances) in a highly complex dendritic tree. Simulating a few seconds of just ONE neuron accurately can take several minutes to several hours of supercomputer time. I know this because I do it for a living.

Re:Except

[eldavojohn]

Except that individual neurons have tens of thousands of possible connections to other neurons, and continually morph and change those connections. That's impossible to do on a rigid piece of hardware.

I'm no expert and I've just been reading the second link's project site on Stanford's page [stanford.edu] but your assertion to continually morph and change those connections seems to be mitigated by this strategy:

Neurogrid simulates six billion synaptic connections by using local analog communication, another choice motivated by cortical studies. Cortical axons synapse profusely in a local area, course along for a while, then do it again. Thus, nearby neurons receive inputs from largely the same axons, as expected from the map-like organization of cortical areas. Local wires running between neighboring silicon neurons emulate these patches, invoking postsynaptic potentials within a programmable radius. Using a patch radius of 6 lets us increase the number of synaptic connections a hundredfold—from 600 million to six billion—without increasing digital communication.

If they connect most (if not all) possible connections that the morphing/changing synaptic channels can take, then they use a sort of addressing technique and RAM strategy to continually morph and change:

Instead of hardwiring the silicon neurons together, as Mead did in his silicon retina, we softwired them by assigning unique addresses. Every time a spike occurs, the chip outputs that neuron’s address. This address points to a memory location (RAM) that holds the synaptic target’s address, or to multiple memory locations if the neuron has multiple synaptic targets. When this address is fed back into the chip, a post-synaptic potential is triggered at the target. An extremely efficient technique, as the same post-synaptic circuit serves all the synapses that neuron receives—virtual synapses! Encoding, translating, and decoding an address happens fast enough to route several million spikes per second, allowing a million connections to be made among a thousand silicon neurons. These softwires may be rerouted simply by overwriting the RAM’s look-up table, making it possible to specify any desired synaptic connectivity.

Although their page is really hard for a lay person like myself to get through, it's very informative [stanford.edu]. Having read it, I'm considerably more optimistic about the future of biological tissues and nervous systems being translated to hardware. At least people are starting back at the simple components and adding new twists.

Re:long ways to go yet

[toppavak]

He's not arguing that it didn't work, he's arguing that they essentially ran a simulation of a large Artificial Neural Network, a relatively trivial task as long as you have a big enough computer behind it. ANNs are essentially points that connect to each other and learn by assigning weights to these various connections- this is essentially the simplest possible way to simulate the behavior of a neuron. The argument is being made that to claim an ANN, regardless of its size, approaches the capabilities of any mammalian brain is simply wrong, and that a true attempt to create such a simulation would need to factor in the stochasticity of ion channels, branchings in neurons and various other biological phenomena that have a tremendous impact on how our brains work.

Without reading more details on the original work, I'm inclined to say that he has a very valid point if they were indeed only running a large ANN model.

Markram's for real

[bellwould]

My research recently took me to some of Markram's work - the guy is brilliant and REALISTIC. His research goals are simple and attainable and any claims of success he has are *well* within the real world. He's incrementally worked his way up from a few neurons - the way a *real* scientist works; and to him, the simplest "brain simulation" of any sort is definitely possible, but far off in the future.

Off-topic correction...

[hiryuu]

Our world increasingly looks like Fredrick Pohl's story "The Marching Morons"...

Not saying this because I know better, but because your mention of the story intrigued me and I hoped to find it or at least find out more about it. It appears it was written by Cyril M. Kornbluth, a contemporary and good friend of Pohl's.

link [wikipedia.org]

I think I must find this story, as the premise of "Idiocracy" was interesting but the execution seemed, to me, quite flawed.