According to one reknowned scientist, IBM's recent claim is all hype.
After IBM's announcement that they had developed a neural network that exceeded the processing power of a feline cortex, Henry Markram sent several scathing e-mails and comments debunking the claim. Markham heads up the Blue Brain project at Lusanne, Switzerland-based EPFL, where his team is attempting to reverse engineer the mammalian brain. According to Markram, the cat brain claim is little more than hype from Dharmendra Modha, the manager of IBM's Almaden Research Center.
Science blogger and tech analyst Greg Fish caught up with Markram to ask him what problems he sees in IBM's simulation, what kind of neural models he envisions for the future and what practical applications they may have to computing in general.
Greg Fish: How far short do you feel IBM has come up in their claim of simulating a structure exceeding the scale of a feline cortex and why?
Henry Markram: They claim to have simulated over a billion neurons interacting. Their so called "neurons" are the tiniest of points you can imagine, a microscopic dot. Over 98 percent of the volume of a neuron is branched (like a tree). They just cut off all the branches and roots and took a point in the middle of the trunk to represent a entire neuron. In real life, each segment of the branches of a neuron contains dozens of ion channels that powerfully controls the information processing in a neuron. They have none of that.
Neurons contain tens of thousands of proteins that form a network with tens of millions of interactions. These interactions are incredibly complex and will require solving millions of differential equations. They have none of that.
Neurons contain around 20,000 genes that produce products called mRNA, which builds the proteins. The way neurons build proteins and transport them to all the corners of the neuron where they are needed is an even more complex process which also controls what a neuron is, its memories and how it will process information. They have none of that.
They use an alpha function (up fast, down slow) to simulate a synaptic event. This is a completely inaccurate representation of a synapse. There are at least six types of synapses that are highly non-linear in their transmission (i.e. that transform inputs and not only transmit inputs). In fact you would need a tens of thousands of differential equations to simulate one synapse. Synapses are also extremely complex molecular machines that would themselves require thousands of differential equations to simulate just one. They simulated none of this.
There are complex differential equations that must be solved to simulate the ionic flow in the branches, to simulate the ion channels biophysics, the protein-protein interactions, as well as the complete biochemical and genetic machinery as well as the synaptic transmission between neurons. Hundreds of thousands of more differential equations. They have none of this.
Then there are glia -- ten times more than neurons -- and the blood supply, and more and more. These "points" they simulated and the synapses that they use for communication are literally millions of times simpler than a real cat brain. So they have not even simulated a cat's brain at one millionth of its complexity. It is not even close to an ants brain.
Fish: The official IBM press release did not explicitly state that the Blue Matter team modeled a cat brain, however many blogs and news sites did. Do you feel the distortion of the announcement is the fault of IBM, the press, or both and why?
Markram: I know very well how the press can blow up things, but this time the press is not to blame. They are the victim of an outright false announcement. This is IBM's Modha false claim, no one else is to blame. Here is an excerpt from the abstract of his non peer-reviewed "Cat Out of the Bag" Gordon Bell Prize paper:
This is not a media blow up at all, the media was unethically mislead. It is a serious case of misconduct and any scientists should not stand by and listen to such nonsense and watch the public being deceived. That is what can happen if papers are not peer-reviewed. Well, this is his peer review.
Furthermore, Eugene Izhikevich [of the Brain Corporation in San Diego, CA,] performed a simulation that is actually 60 times larger than Modha's already several years ago (100 billion neurons) and he did it on desktop computers. Modha, even used Izhikevich's equations, so he knows very well that his simulation was not even the largest performed. Izhikevich should get the Bell prize not this guy.
Ovidiu Anghelidi [of the International Neuroinformatics Coordinating Facility in Toronto, Canada] actually holds the world record with a 700 billion neuron simulation. In fact, his neurons are much more complex neurons as he uses what is known as the Hodgin-Huxley equations which can capture the subtleties of complex ion channels on neurons. Their simulation took days so it is not optimal, but if they had a supercomputer available this would be really straight forward -- no big technical feat.
So, this pretty much discredits the Gordon Bell committee for not even checking what has been done previously. The fact that he had a supercomputer to do this makes it technically even more trivial, not more remarkable. Try doing it on a GRID or on desktops.
Fish: What is your take on cognitive computing projects and how do you see models of cortical structures or entire brains advancing the concept? How would it be applied in practical uses?
Markram: These are not the kind of simulations that I think will help you understand the brain, but I fully accept that neural networks can be extremely powerful for computing problems that normal AI cannot easily solve. In fact, the neural network industry is huge and simulating larger networks does get you new kinds of computing capabilities. What I object to is the false and misleading claim this is brain. It is an artificial network, not an artificial brain simulation. If you can use new learning algorithms then you can train such artificial networks to solve very complex problems, but you will probably need a trillion trillion such neurons to get close to the computational power of a small insect.
Fish: In your conception, what would be an accurate model of a brain or a brain structure and how would it be set up?
Markram: Well, you need to at least mathematically abstract every single interaction at the molecular level and build in all the constraints that took evolution billions of years to discover. You need to have the right composition of neurons, the right numbers and the right synaptic connectivity between the neurons etc etc. Blue Brain uses a strategy to reach this goal, following biology rather than theory.
Fish: Do you believe that such models could lead to artificial intelligence or sci-fi technologies like mind uploading often mentioned in the popular science realm? If so, how could such systems be set up?
Markram: My prediction is that we will understand the neural code way before we finished building the brain - if we build it bottom up (not just points interacting). The neural code is the code that the brain uses to represent information. Once you have this code many things will be possible such as virtual reality that stimulates your brain to experience certain perceptions. The technology to stimulate the brain in this way will probably take longer, but eventually it will come. Now to record is much the same. If you can record from billions or trillions of locations in the brain, you probably could catch a neural state that represents a thought or a memory.
If you can stimulate the brain with these patterns then the other person may see something similar. However, our brains are calibrated differently so you would have to learn how to interpret the stimulation and then you should be able to see what someone else saw. So downloading is far in the future, but there is no fundamental reason why it will not happen eventually. I kind of allude to it, better not to speculate too far into the future.