ok, so maybe it’s not exactly the cat’s meow…
Two IBM project leaders are arguing about the results of the Blue Matter experiments. The critics seem to have the upper hand, scientifically speaking.
While the blogosphere and news agencies have been discussing IBM’s claim of simulating the brainpower of a cat cortex and some of us have been discussing its potential use in artificial intelligence development, Dr. Henry Markram, the lead of a similar project backed by the same company, read the press release and went into a nuclear meltdown of blistering fury around the web.
According to his rebuttals, the Blue Matter team is just playing around with brute computing force and their claims are nothing but self-aggrandizing hype. Is this a simple case of sour grapes? Well, the write-up at Fast Company included a link to a paper describing the setup behind the simulation in enough detail to see what was really being simulated and how, and as it turns out, the closest Blue Matter came to matching anything like a cat’s brain was solely in the realm of analogies.
Rather than create a feline cortex or a reasonable emulation of one, the Blue Matter team created a statistical model of how neurons would act in a hypothetical brain and fired grids of neurons as fast as they could to get simulated readings of the noise produced in the process. There were few specific structures or pathways of an actual cortex. It was all about trying to create an efficient way for dealing with highly intensive computing on a very large scale.
The logic behind how the simulation would work is detailed in the intro of Section 3, which deals not with how to accurately simulate the pathways of an actual cortex, but scalability issues in updating the status of a digital neuron on a tightly controlled cycle and digital synapses on an event-driven basis. When neurons fire, they’re updated twice (on pre and post synaptic events). As the neurons cycle, the synapses are updated through an architecture which lowers the cost of computing these events (how resource intensive the process of updating the synapse statuses will be) in a largely unspecified way.
Then, in Section 4, we get some hint of structure as the paper mentions three types of neurons connected in loops. However, we’re only taking about a few loops in which neurons should fire rather than entire structures being reproduced and exposed to a stimulus supposedly mimicking signals from visual objects or sounds. In the actual simulations, the sheets of neuron loops are fed arbitrary signals which just fire through their cycle. This is still pretty cool stuff because the team is creating a tiny snapshot of how neurons in the brain fire under several algorithms.
But without an entire, carefully charted and modeled brain, all we’re getting is just a teeny glimpse at how neural nodes go through their firing cycle according to a statistical model. We’re not getting a true sense of what happens in a feline brain when a cat sees an object or hears a sound. The vast majority of what we would need to simulate that is absent. And yet, because Modha’s team created tens of thousands of sheets of these simple loops which amount to 1.6 billion neurons and 8.87 trillion synapses in total, they use this straight-line multiplications to claim that their simulation effectively exceeds the scale of a cat’s cortex and lays the groundwork for even bigger models and ultimately, cognitive computing.
I really don’t think that’s how it works. Their simulations don’t actually compute anything or produce more than just a simulated EEG made by clouds of digital neurons as they fire through their loops. There was no actual computing being done by those neurons, no result, no actual decision making, no response to the stimuli we would consider more meaningful than the completion of a circuit. However, the whole purpose of a brain is its ability to make decisions. Brainpower isn’t just the firing of neurons en masse. It’s how we deal with the world around us and solve complex tasks. All that the Blue Matter experiments proved is that a supercomputer really can summon the brute computing force to run a scaled up loop through its paces.
As Modha notes in the intro to Section 5, the simulation is missing all the intricate details of a cortex and the connections are based on a best estimate, which offers even more evidence to Markram’s charges that simulations run by the Blue Matter team aren’t even in the same building as real feline brains, much less in the same sentence without the word “not” present. The leap between firing a lot of neurons and computerized cognition is vast, and Blue Matter has barely inched in the cognitive direction, even with this much trumpeted experiment.
The misconception was made even worse when several gadget and tech blogs started referring to the results of the experiment as cat brain simulations and representing them as a potential development in bringing AI to life, since the point of this research is to make computers that can perform cognitive tasks. And while we can’t blame Modha and the Blue Brain team for bloggers extrapolating the initial hype, we should note that their big claims are pretty daring exaggerations of their accomplishments. To call their simulations a significant step in creating models able to emulate the brain’s abilities is a little like calling your moped a potential precursor to a hyper-car that can go toe to toe with a Bugatti Veyron because you’ve been able to tie a lot of mopeds together and you’ve calculated that their combined velocity adds up to 250 mph.