Henry Markram has become famous as the creator of the world's most expensive brain simulation, but neuroscientists know him best for his pioneering experiments on synapses. Markram was one of the first to investigate the sequential version of Hebb's rule in a systematic way, by varying the time delay between the spiking of the two neurons when inducing synaptic plasticity. (Changes in the synapses, the connection points between cells. One scientist reduced Hebb’s rule to: “Cells that fire together, wire together.”) When I first heard Markram speak at a conference, I also encountered the chain-smoking and charming Alex Thomson, another prominent neuroscientist, who lectured about synapses with bubbling enthusiasm. She was in love with them, and wanted us to love them too. Markram, in contrast, came across as the high priest of synapses, summoning our awe and respect for their intricate mysteries.
In a 2009 lecture Markram promised a computer simulation of a human brain within ten years, a sound bite that traveled around the world. If you view the video of the lecture online, you might agree with me that his handsomely sculpted face looks a bit fierce, but his manner of speaking is gentle and inviting, with the quiet conviction of a visionary. He didn't sound so calm later that year. His competitor, the IBM researcher Dharmendra Modha, announced a simulation of a cat brain, after having claimed a mouse brain simulation in 2007. Markram responded with an angry letter to IBM's chief technology officer:
You told me you would string this guy up by the toes the last time Mohda [sic] made his stupid statement about simulating the mouse's brain.
I thought that ... journalists would be able to recognize that what IBM reported is a scam - nowhere near a cat-scale brain simulation, but somehow they are totally deceived by these incredible statements.
I am absolutely shocked at this announcement .... I suppose it is up to me to let the "cat out of the bag" about this outright deception of the public.
Competition is great, but this is a disgrace and extremely harmful to the field. Obviously Mohda would like to claim he simulated the Human brain next-I really hope someone does some scientific and ethical checking up on this guy.
All the best,
Markram didn't keep his indignation secret. He sent copies of the letter to many reporters. One of them blogged about the controversy with a story wittily headlined "Cat Fight Brews Over Cat Brain:”
The letter marked a new low point in Markram's relationship with IBM. They had started out allies in 2005, when IBM signed an agreement with Markram's institution, the École Polytechnique Féderále in Lausanne, Switzerland. The goal of the joint project was to showcase IBM's Blue Gene/L, at that time the fastest supercomputer in the world, by using it to simulate the brain. Markram called the project "Blue Brain;” an allusion to IBM's nickname, "Big Blue:” But the relationship soured when Modha started a competing simulation project at IBM's Almaden Research Center.
Markram tried to defend his own work by accusing his competitor of fakery. But actually he cast doubt on the whole enterprise. Anyone can simulate a huge number of equations and claim it's like a brain. (You don't even need a supercomputer these days.) What's the proof? How do we know that Markram isn't a scammer too?
His glitzy supercomputer should not distract us from a potentially fatal flaw of his research: the lack of a well-defined criterion for judging success. In the future, Blue Brain could be evaluated with the specific Turing test described earlier, but this test only becomes useful when the simulation approaches the real thing. These purported mouse and cat brain simulations are not even in the ballpark yet. No "Mouse-tin Guerre" is going to fool you any time soon. The Turing test [a test of computer intelligence] will tell us when we've reached our destination, but until that day comes, we need a way of knowing if we're going in the right direction.
Are these researchers really making progress? The full text of Markram's letter was too long to print here, so I'll just summarize the science behind his vitriol. In short, Blue Brain is composed of model neurons that are highly sophisticated in their handling of electrical and chemical signals. They are more faithful to real neurons than are the model neurons of Modha's simulation, which in turn are more realistic than the weighted voting model discussed in this book.
There is plenty of empirical evidence that the weighted voting model approximates many neurons well. But we also know that the model is not perfect, and can even fail badly for some neurons. Markram is correct that real neurons have many complexities that are not captured by simple models. A single neuron is an entire world in itself. Like any cell, it's a highly complex assembly of many molecules, a machine built from molecular parts. And each of these molecules in turn is a minuscule machine made of atoms.
As I mentioned earlier, ion channels are an important class of molecule, because they are responsible for the electrical signals in neurons. Axons, dendrites, and synapses contain different types of ion channels, or at least have them in differing numbers, which is why these parts of neurons have distinct electrical properties. In principle, every neuron is unique in its behavior, owing to the unique configuration of its ion channels. This is a far cry from the weighted voting model, according to which all neurons are essentially the same. But it sounds like bad news for brain simulation. If neurons were infinitely diverse, how could we ever succeed at modeling them? Measuring the properties of one neuron would tell you nothing about another.
There is one hope for escaping the morass of infinite variation: neuron types. You may recall that Cajal [neuroscientist pioneer Ramón y Cajal] classified neurons into types based on location and shape. You can think of these properties as being like an animals’ habitat and appearance. When a neuroscientist speaks of the double bouquet cell of the neocortex, it reminds me of the way that a naturalist speaks of the polar bear of the Arctic. The naturalist might also point out that polar bears, unlike brown bears, hunt for seals. Likewise, neurons of the same type generally exhibit the same electrical behaviors. This is presumably because their ion channels are distributed in the same way.
If this is the case, then neural diversity is actually finite. We should compile a catalog of all the neuron types, a “parts list” for the brain, and then construct a model for each type. We’ll assume that each model is valid for all neurons of that type in all normal brain, much as we assume that all resistors behave the same way in any electronic device. Once all neuron types have been modeled, we’ll be ready to simulate brains.
Markram's laboratory has characterized the electrical properties of many neocortical neuron types through experiments in vitro, and constructed a model for each. Each neuron type is modeled as hundreds of interacting electrical "compartments;” which is an approximation to simulating the millions of ion channels in a neuron. Markram deserves credit for the realism of the multicompartmental model neurons used in Blue Brain.
But Blue Brain [which has now morphed into the Human Brain Project] is severely lacking in one respect. Since no cortical connectome is known yet, it's not clear how to connect the model neurons with each other. Markram follows Peters' Rule, a theoretical principle stating that connectivity is random. The accidental collisions of axons and dendrites in the tangled "spaghetti" of the brain lead to contact points. At everyone of these, a synapse occurs with some probability, as if it were the outcome of tossing a biased coin.
Peters' Rule is conceptually related to an idea introduced earlier, the random synapse creation of neural Darwinism. The ideas are not equivalent, however. Neural Darwinism includes activity-dependent synapse elimination, which makes the surviving connections end up nonrandom. Violations of Peters' Rule have already been discovered. I suspect that many more will be found, and that the rule has managed to survive only because of our ignorance of connectomes.
As computer scientists like to say, "Garbage in, garbage out:” If the neural connectivity of Blue Brain is wrong, the simulation will be too. But let's not be overly critical. In the future, Markram could always incorporate information from connectomes into Blue Brain. Then wouldn't his simulation become truly realistic?
To answer this question, let's again consider the roundworm C. elegans. Its connectome is already known, unlike that of the neocortex. It may come as a surprise that only small parts of its nervous system have been simulated. These models have been helpful for understanding some simple behaviors, but they are piecemeal efforts. No one has come close to simulating the entire nervous system.
Unfortunately, we lack good models of C. elegans neurons. As I mentioned earlier, most of them don't even spike, so the weighted voting model isn't valid. To model the neurons, we'd have to measure from them, but this turns out to be more difficult for C. elegans than for mouse or even human neurons. We also lack information about C. elegans synapses. The connectome did not even specify whether the synapses were excitatory or inhibitory.
So Blue Brain lacks a connectome, while C. elegans lacks models of neuron types. Both elements are needed to simulate a brain or nervous system. Thus the earlier claim should be revised to say, "You are your connectome plus models of neuron types' (Let's assume that a connectome is defined to specify the type of each neuron.) But the models of neuron types are likely to contain much less information than the connectome, as most scientists agree that there are far fewer neuron types than neurons. In this sense, "You are your connectome" would remain a very good approximation. Furthermore, we assumed above that all neurons of one type behave in the same way in all normal brains, just as all polar bears hunt seals under normal circumstances. If we uploaded multiple people, all the simulations could share the same models of neuron types. The only information unique to a person would be his or her connectome.
It's worth noting that the balance of information content is quite different in C. elegans. Its three hundred neurons have been classified into about one hundred types, which is not that much smaller than the number of neurons. Essentially every neuron (along with its twin on the other side of the body) is its own type. Every neuron may end up requiring its own model, and the total information in these models might exceed that in the connectome. So "You are your connectome" would be a terrible approximation for a worm, even though it might be almost perfect for us.
To put it another way, the C. elegans nervous system is like a machine built from parts that are all unique. The individual workings of the parts are just as important as their organization. The opposite extreme would be a machine built from a single type of part. (You may be old enough to remember old-fashioned Lego sets, which contained only one type of Lego block.) The functionality of such a machine would depend almost entirely on the organization of its parts.
Electronic devices are close to this extreme, as they contain only a few types of parts, like resistors, capacitors, and transistors. That's why a radio's wiring diagram determines so much of its function. The parts list for the human brain is longer, so it will take many years of effort to model every neuron type in the human brain. But the parts list is still far shorter than the total number of parts. That's why the organization of the parts is so important, and why connectomes are more crucial for humans than for worms.
There's one more important aspect of connectomes to include in brain simulations: change. Without it, your uploaded self would not be able to store new memories or learn new skills. Markram and Modha have included reweighting using mathematical models of Hebbian synaptic plasticity. But it's also important to include reconnection, rewiring, and regeneration. In general, our models for the four R's are much less refined than those for electrical signals in neurons. It will be possible to improve them, but it will take many more years of research.
These are all important caveats, but models of neuron types and connectome change still fit into the overall framework of connectome-based brain simulation. Is there anything about the brain that is fundamentally incompatible with the framework? One difficulty is that neurons can interact outside the confines of synapses. For example, neurotransmitter molecules might escape from one synapse, and diffuse away to be sensed by a more distant neuron. This could lead to interactions between neurons not connected by a synapse, or even between neurons that do not actually contact each other. Because this interaction is extrasynaptic, it is not encompassed in the "wiring diagram;” the connectome. It might be possible to model some extrasynaptic interactions fairly simply. But it's also possible that the diffusion of neurotransmitter molecules in the cramped and tortuous spaces between neurons would require complex models.
If extrasynaptic interactions turn out to be critical for brain function, then it might be necessary to reject the hypothesis "You are your connectome:' The weaker statement "You are your brain" could still be defensible, but this would be much more difficult to use as a basis for uploading [your brain into a computer, as some futurists have proposed]. We might have to throwaway the abstraction of the connectome and descend still further to the atomic level. One could imagine using the laws of physics to create a computer simulation of every atom in a brain. This would be extremely faithful to reality, much more than a connectome-based simulation.
The catch is that a huge number of equations would be necessary, since there are so many atoms. It seems absurd to even consider the enormous computational power required, and is completely out of the question unless your remote descendants survive for galactic time scales. At the present time, it's difficult to simulate even those modest assemblies of atoms called molecules. Simulating all the atoms of a brain is almost beyond imagining. Limited computational power is not the only barrier. There is also the difficulty of obtaining the information to initialize the simulation. It might be necessary to measure all the positions and velocities of the atoms in the brain, which is far more information than in a connectome. It's not clear how to collect that information, or how to do it in a reasonable amount of time.