Henry Markram's Blue Brain Project Vies for €1 billion Prize

[Henry] Markram, a South-African-born brain electrophysiologist who joined the Swiss Federal Institute of Technology in Lausanne (EPFL) a decade ago, may soon see his ambition fulfilled. The project is one of six finalists vying to win €1 billion (US$1.3 billion) as one of the European Union's two new decade-long Flagship initiatives. _Nature

The Challenge
The Blue Brain Project in Lausanne is attempting to model a working human brain to an exquisite level of physiologic detail -- to the ion channel level. This would require an enormous amount of computer power -- to say nothing of the electrical power required to drive and cool the apparatus when fully functional. But €1 billion can presumably still buy a great deal.

Every experiment at least tacitly involves a model, whether it is the molecular structure of an ion channel or the dynamics of a cortical circuit. With computers, Markram realized, you could encode all of those models explicitly and get them to work together. That would help researchers to find the gaps and contradictions in their knowledge and identify the experiments needed to resolve them.

Markram's insight wasn't original: scientists have been devising mathematical models of neural activity since the early twentieth century, and using computers for the task since the 1950s (see page 462). But his ambition was vast. Instead of modelling each neuron as, say, a point-like node in a larger neural network, he proposed to model them in all their multi-branching detail — down to their myriad ion channels (see 'Building a brain'). And instead of modelling just the neural circuits involved in, say, the sense of smell, he wanted to model everything, “from the genetic level, the molecular level, the neurons and synapses, how microcircuits are formed, macrocircuits, mesocircuits, brain areas — until we get to understand how to link these levels, all the way up to behaviour and cognition”. _Nature...BlueBrain

Henry Markram

By the end of 2005, his team had integrated all the relevant portions of this data set into a single-neuron model. By 2008, the researchers had linked about 10,000 such models into a simulation of a tube-shaped piece of cortex known as a cortical column. Now, using a more advanced version of Blue Gene, they have simulated 100 interconnected columns.

The effort has yielded some discoveries, says Markram, such as the as-yet unpublished statistical distribution of synapses in a column. But its real achievement has been to prove that unifying models can, as promised, serve as repositories for data on cortical structure and function. Indeed, most of the team's efforts have gone into creating “the huge ecosystem of infrastructure and software” required to make Blue Brain useful to every neuroscientist, says Markram. This includes automatic tools for turning data into simulations, and informatics tools such as http://channelpedia.net — a user-editable website that automatically collates structural data on ion channels from publications in the PubMed database, and currently incorporates some 180,000 abstracts.

The ultimate goal was always to integrate data across the entire brain, says Markram. The opportunity to approach that scale finally arose in December 2009, when the European Union announced that it was prepared to pour some €1 billion into each of two high-risk, but potentially transformational, Flagship projects. Markram, who had been part of the 27-member advisory group that endorsed the initiative, lost no time in organizing his own entry. And in May 2011, the HBP was named as one of six candidates that would receive seed money and prepare a full-scale proposal, due in May 2012. _Nature

Blue Brain Model
The project is both ambitious and expensive. It is also more complex and difficult than can currently be anticipated or planned for. But kudos to Henry Markram for making the attempt.


H/T NextBigFuture

Human vs. Computer: How About a Real Challenge?

Update 18Feb2011: AI researcher Ben Goertzel presents some thoughts on Watsons ascendancy at H+Magazine online. H/T Brian Wang

Much has been made about the recently televised Jeopardy victory by IBM Watson, and earlier victories and impressive performances by specialised chess computers such as Blue Gene etc. Even in poker, computers are increasingly seen as threats to human dominance, thanks to clever human programmers. In the game world, Go is most often seen as a game where computers have not come close to experts.

What do these massive, power gobbling, ultra-pampered, spoon-fed, one-trick-pony game-playing super computers tell us about the human vs computer rivalry? Realistically, the rivalry is still human vs. human, with one team of humans utilising ultra-fast electronics devices to store, "analyse", and retrieve massive quantities of data, to gang up on a single human opponent.

Watson is certainly faster to the button than its human opponents, but we already knew that electrons were faster than nerves. Once the "natural language processing" trick was mastered, Watson could more readily lock out his opponents from responding -- even when they both clearly knew the answer.

But Watson could not drive itself home after the game, could not flush away its excretions (heat) by itself, could not feed itself, etc. In the end, Watson is a very expensive gimmick which served as a showcase for various specialised programming problems.

Perhaps if Watson could master all the games mentioned above, at once, and defeat experts in all of the games, it would be impressive as a game-player. But not really. Look at all the money, mass, and energy tied up in the junkpile called Watson. How would IBM make it more capable of playing multiple games? By throwing more mass, money, and energy into the already-huge junkpile. Not very clever, really, compared with the three pound human brain and all the things it can do -- including designing, building, repairing, and programming "smart" computers.

It all points out the fact that the state of artificial intelligence is pretty pathetic, all in all. Despite over 60 years of promises to create human-level intelligence "within 10 years", AI still stinks badly, and promises more of the same into the forseeable future.

The Jeopardy challenge -- like all similar challenges -- was a huge and expensive publicity hullabaloo. It is quite likely to damage the Jeopardy brand in the long run. It certainly puts forth an entirely false idea about the modern capability of computers, vis-a-vis humans, to reason and make decisions.

What would be a real challenge for Watson? How about a spontaneous, unplanned race over an extensive, lengthy, novel, 3-D obstacle course with ladders, walls, tunnels, slides, sand, and foot-deep water traps -- against a 5 year old human child?

Let's face it: Modern life requires humans to overtly or covertly (via proxies) partner with computers to achieve optimum performance in large areas of our lives. But what will it take to get computers to the point where they are consciously setting the agenda for humans, rather than the other way around?

It will take an entirely new "substrate of thought" than the high speed digital architectures currently used to such great -- if ultra-specialised -- effect. Worse, modern AI researchers for the most part have no idea what form such a new substrate would take. Certainly they do not understand the substrate for the only proof of concept of conscious intelligence which currently exists -- the human brain.

Too much like robots themselves, too many AI researchers unwittingly plod along artificial pathways leading to nowhere but diminutive local optima. Watson is only one illustration of the kludgy phenomenon.

What will it take, and how long will it take to discover it? There are limits to pure reason and speculation. Experimentation is necessary. Hands must be dirtied and hypotheses must be generated and tested. For the luggiest of lugheads out there, we need much better challenges than chess, Jeopardy -- or even Go -- to spur the effort required.

How Can Machine Intelligence Copy the Brain?

Intelligent machines that not only think for themselves but also actively learn are the vision of researchers of the Institute for Theoretical Science (IGI) at Graz University of Technology...They have been co-ordinating the European Union research project "Brain-i-Nets" (Novel Brain Inspired Learning Paradigms for Large-Scale Neuronal Networks) for three years... _SD

The Brain-i-Nets project at Graz University of Technology is quite different from Henry Markram's Blue Brain Project in Lausanne.  While Blue Brain aims to create a silicon substitute for the brain for neuroscience and neuropathology research, the Graz project aims to create actual machine intelligences by learning from a detailed and intense study of intact brains. Brain-i-Nets is attempting to abstract the significant aspects of dynamic cortical network plasticity, by using advanced real-time imaging of intact brain networks as they are working and changing.

The scientists want to design a new generation of neuro-computers based on the principles of calculation and learning mechanisms found in the brain, and at the same time gain new knowledge about the brain's learning mechanisms.

The human brain consists of a network of several billion nerve cells. These are joined together by independent connections called synapses. Synapses are changing all the time -- something scientists name synaptic plasticity. This highly complex system represents a basis for independent thinking and learning. But even today there are still many open questions for researchers.

"In contrast to today's computers, the brain doesn't carry out a set programme but rather is always adapting functions and reprogramming them anew. Many of these effects have not been explained," comments IGI head Wolfgang Maass together with project co-ordinator Robert Legenstein. In co-operation with neuroscientists and physicists, and with the help of new experimental methods, they want to research the mechanisms of synaptic plasticity in the organism.

...The three-year project is financed by the EU funding framework "Future Emerging Technologies" (FET), which supports especially innovative and visionary approaches in information technology. International experts chose only nine out of the 176 applications, among which was "Brain-i-Nets."
_SD

Brain-i-Nets Website

The overall long-term vision of this project is

to develop new design principles for adaptive, reconfigurable very-large-scale hardware systems implementing novel learning rules inspired by biological neural networks in vivo.

Learning mechanisms implemented in the brain appear to be much more robust and flexible than those currently used in neurally inspired computing systems. To confer the superior adaptive and computational capabilities of biological neural systems to large-scale recurrent neural hardware systems and other novel massively parallel computing devices, new and more sophisticated learning rules are needed.

Our long-term vision is that the learning rules for global gating of local learning, identified and explored in this project, will become ideal candidates for implementation in hardware. Conceptually the interaction of local factors that can be monitored and stored at the site of each connection with one or a few global factors is very attractive for hardware implementation. Previous collaborations of several partners of the project have shown that networks of spiking neurons can be implemented in a truly large-scale, parallel, mixed analog-digital hardware system. The inclusion of learning rules that go beyond the classical Hebbian or STDP rules for unsupervised learning, by including a third factor representing for example information on saliency or reward, will advance the hardware into a regime where a much broader class of learning tasks can be solved by these ultra-rapid machines. __Brain-i-Nets

Besides the Graz University of Technology, scientists from University College London, the University of Heidelberg, the University of Zurich, Ecole Polytechnique Federale de Lausanne, and the Centre National de la Recherche Scientifique of France are also partnering in the project.

Al Fin neuroscientists consider this project to be one of the most sophisticated approaches to the biomimetic creation of machine intelligence to this date. For such a project to be truly successful, it is likely to require the expertise of scientists, technologists, and engineers from North America and East Asia, as well as those at the European centres mentioned above.

Blue Brain Tour in Lausanne: What Is Happening?

Bluebrain | Year One from Couple 3 Films on Vimeo.

Henry Markram brings us up to date on the aims of Blue Brain.  H/T MachinesLikeUs

The Swiss Blue Brain Project is meant to provide a "machine model" of the brain, for purposes of neuroscientific study.  It does not mean to build a machine intelligence.  Instead, Blue Brain is a massive simulation of brain function down to the level of ion channels within individual neurons.

Markram expects the project to provide a means for solving the diseases of the brain -- including Alzheimer's, Parkinson's, Huntington's, and others -- at a cellular and molecular level.

Clearly, Blue Brain is a massively ambitious and expensive project.  How soon will it begin to pay for itself by solving the critical problems of brain degeneration and other deficits of brain function?  It is impossible to say.

The video segment above is the "year one" report out of a "ten year project" to document the Blue Brain project on film.

The Brain Is Far More Complex Than Believed

We have been bombarded with predictions that human-level artificial intelligence will be developed "within ten years." These predictions inevitably come from persons with a computer science, engineering, or artificial intelligence background. Such  prognosticators understand algorithms and / or electrical circuits, but do they understand how consciousness is created? Do they comprehend the basis for the only human-level intelligence that exists: the human brain? Clearly not.

This respected Harvard team of neuroscientists has its hands full studying a nematode nerve network of 4 measly neurons! They are hoping to expand their study to include more than 4 neurons soon.

Meanwhile, there is the problem of "The Other Brain", the glial network.

Glia communicate by broadcasting chemical messages. Moreover, glia can sense information flowing through neural circuits and alter the communications between neurons at synapses! Glia, we now know, have receptors for detecting the flow of ions generated by neurons firing electrical impulses and for sensing the neurotransmitters neurons release at synapses. Glia intercept these signals and act upon them to increase or decrease the transmission of information across synapses and speed or slow the transmission of electrical information through axons.


These recent discoveries open an entirely new dimension into brain function. Glia are involved in all aspects of nervous system health and disease. They can control neuronal communication, development of the fetal brain, generation of new neurons in the adult brain, participate in epilepsy, Alzheimer's disease, mental illnesses such as depression and schizophrenia, and they provide a new mechanism of learning that operates beyond synapses. _RDouglasFields

Henry Markram's Blue Brain project is the one supercomputer project that seems to be taking into account much (but not all) of the brain's complexity.   But Markram's project is not trying to create human-level AI.  It is trying to create a simulated brain that can be used to better understand brain function and diseases of the brain.   Markram is a neuroscientist, so his goals are focused more on the realities of the brain than on the goal of "human-level AI."

The best of the "brain simulations" by AI workers is the simulation by Dharmendra Modha's IBM team. It barely simulated one of the most basic functions of the visual cortex, at average speeds less than 1 / 100th that of a mammalian brain. It required over $1 million a year to power its processors, and much more to cool the supercomputer. For something with the same number of neurons as a cat brain, Modha's simulation was a hugely inefficient use of energy -- compared to a cat.

There are some interesting new hardware tools coming along, such as Chua's memristors, meminductors, and memcapacitors. Alice Parker's BioRC project at USC may hold some future promise.

The point is not that a human-level AI would have to emulate the human brain in great detail. That would be a rather pointless extravagance, when crafty abstraction can save space, time, energy, and effort. But new hardware and software tools are needed, obviously. One cannot build a slow-functioning silicon retard occupying a skyscraper-sized building requiring a nuclear reactor to power and mega-tons of refrigeration to cool, and call it a human-level brain.

Anyone wanting to build a human-level machine intelligence will need to understand much of what consciousness entails, and how the human brain achieves it.

It is possible that an insect-level machine brain of appropriate size and energy-consumption might be developed within 5 to 10 years. A rodent-level machine brain of appropriate dimensions might be developed within 10 to 20 years, if we are lucky. A human-level brain may not take more than 5 years to achieve after the rodent-level brain, but shrinking it to the size and energy-efficiency of a human brain may take longer.

Humans are stupid. But they are also the most powerful general-purpose intelligence known to us. Machine intelligence of human-level or better would generate significant changes to human societies. Modern conventional human governments do not want to lose control of this particular revolution. If they do, it might be the last mistake they ever make.

More On IBM's BlueMatter "Brain Simulation"

Popular Mechanics provides information on a recent IBM "brain simulation" of 1.6 billion simulated neurons -- an attempt to "simulate the human visual cortex". Here are some fascinating details from PM:

Modha's billion-neuron virtual cortex is so massive that running it required one of the fastest supercomputers in the world—Dawn, a Blue Gene/P supercomputer at Lawrence Livermore National Laboratory (LLNL) in California.


Dawn hums and breathes inside an acre-size room on the second floor of the lab's Terascale Simulation Facility. Its 147,456 processors and 147,000 gigabytes of memory fill 10 rows of computer racks, woven together by miles of cable. Dawn devours a million watts of electricity through power cords as thick as a bouncer's wrists—racking up an annual power bill of $1 million. The roar of refrigeration fans fills the air: 6675 tons of air-conditioning hardware labor to dissipate Dawn's body heat, blowing 2.7 million cubic feet of chilled air through the room every minute.


Dawn was installed earlier this year by the Department of Energy's National Nuclear Security Administration (NNSA), which conducts massive computer simulations to ensure the readiness of the nation's nuclear weapons arsenal. Modha's team worked with Dawn for a week before it was transitioned to NNSA's classified nuclear work. For all of its legendary computing power, Dawn still ran Modha's 1.6 billion neurons at only one-six-hundredth the speed of a living brain. A second simulation, with 1 billion neurons, ran a little faster—but still only at one-eighty-third of normal brain speed.


These massive simulations are merely steps toward Modha's ultimate goal: simulating the entire human cortex, about 25 billion neurons, at full speed. To do that, he'll need to find 1000 times more computing power. At the rate that supercomputers have expanded over the last 20 years, that super-super computer could exist by 2019. "This is not just possible, it's inevitable," Modha says. "This will happen." _PM

Can you imagine $1 million a year just to power the processors? 6675 "tons" of air conditioning to keep the hardware cool? All to run a very poor simulation of a human visual cortex at 1 / 600 th the speed of a human brain?

You have to admire Modha's optimism when he claims that it is inevitable that the human level artificial cortex will be operating in real time by the year 2019. As I said a couple of days ago, the estimation is preposterous.

First of all, the current simulation is of the most basic cortical neural architecture. It is the easiest, least complex type of neural simulation problem to solve. In fact, an insect brain is more complex -- with far fewer neurons, at much faster speeds, in a far smaller space, with a far lower energy budget -- than the monstrosity that IBM has strung together at Livermore.

It looks as if Modha's team is in a race with Henry Markram's Swiss team, and other teams around the world to simulate the human cortex the soonest. Perhaps there is an X prize for that achievement. But it seems to Al Fin engineers that this brute force approach is best adapted for burning up research funds, rather than actually coming close to simulating a human brain.

Yes, you must crawl before you can walk, walk before you can run.  But it seems as if what the IBM team is doing is moving two fingers in the air in a crawling motion, and calling it crawling.  Time will tell.

As I said before, the achievement of simulating some basic quasi-physiologic responses to simulated input stimuli -- on this scale -- is remarkable.  This platform will provide for some fascinating lab experiments in artificial neuronal networks.  There is a lot to be learned.

But please go easy on the claims.  The field of artificial intelligence has left a junkyard of exaggerated claims and absurd unfulfilled predictions across the landscape of the past 50 years.  Stick to the facts.

Actually, No, IBM Did NOT Simulate a Cat's Brain

Cognitive computing seeks to engineer the mind by reverse engineering the brain. The mind arises from the brain, which is made up of billions of neurons that are liked by an internet like network. An emerging discipline, cognitive computing is about building the mind, by understanding the brain. It synthesizes neuroscience, computer science, psychology, philosophy, and mathematics to understand and mechanize the mental processes. Cognitive computing will lead to a universal computing platform that can handle a wide variety of spatio-temporally varying sensor streams. _Source

Image via Brian WangAt a supercomputing conference in Portland, Oregon, IBM announced a successful simulation of a "cat brain scale" silicon neuro-computing platform. You can read an engrossing account of this achievement from Dharmendra S. Modha (via Brian Wang).

The achievement is a fascinating one, and worth celebrating. But it would be inaccurate to describe the accomplishment as the simulation of a cat brain. IBM's simulation is nowhere close to a simulation of a cat's brain -- and we should have no illusions regarding the comparison. Consider this:

...using Dawn Blue Gene / P supercomputer at Lawrence Livermore National Lab with 147,456 processors and 144 TB of main memory, we achieved a simulation with 1 billion spiking neurons and 10 trillion individual learning synapses. This is equivalent to 1,000 cognitive computing chips each with 1 million neurons and 10 billion synapses, and exceeds the scale of cat cerebral cortex. The simulation ran 100 to 1,000 times slower than real-time. _ Modha

Some of the neuronal level phenomena observed in the simulation (in response to stimuli) vaguely resembled actual neurocortical electrophysiological activity observed in a mammalian brain. But the parallels are quite loose. And considering the time scale of the simulation was between 100 and 1,000 times slower than real-time, tweaking the stimuli and response to better match mammalian equivalents may take some while.

IBM's simulation of 1 billion spiking neurons was interesting, but a very pale shadow of a genuine cat brain, with all its specialised structural cortical and sub-cortical components -- and accessory (but very influential) glial and vascular infrastructure.

IBM is estimating that around the year 2018, it will be able to simulate a human-scale "brain" in real time. That is preposterous. Only if you accept that IBM has simulated a cat's brain in 2009 would you accept that IBM may simulate a human brain by 2018. But as IBM readily admits, it has not even simulated a cat's multi-specialised cortex -- much less a cat's entire brain.

In terms of details in our simulations, we are currently working on differentiating our cortical region into specific areas (such as primary visual cortex or motor cortex) and providing the long-range connections that form the circuitry between these areas in the mammalian brain. For this work, we are drawing from many studies describing the structure and input/output patterns of these areas as well as a study recently performed within IBM that collates a very large number of individual measurements of white matter, the substrate of long-range connectivity within the brain. _Modha

Give them time. I expect some amazing results from this approach to brain simulation, eventually. Unfortunately, neither the hardware nor the software tools are anywhere close to providing scientists with a semi-realistic mammalian brain at this time.

Even a realistic insect brain would be a marvelous achievement. Particularly one that operated in real time, and was sized on roughly the same scale as an actual insect.

So while Al Fin cognitive neuroscientists, engineers, and cognitive computing specialists congratulate IBM on its latest achievement, they also caution IBM not to overstate what they have actually accomplished.

25 Nov 2009 Update: Be sure to read Henry Markram's response to the IBM claims at Brian Wang's site (and in comments here).

Rat Cortical Column Simulation Update: From Here to an Artificial Brain?

One of the holy grails of neuroscience is the creation of an accurate simulation of mammalian brain cortex. Swiss researchers have been working on the "Blue Brain" project since 2005, and are collaborating with IBM researchers to simulate the neocortex.

By mimicking the behavior of the brain down to the individual neuron, the researchers aim to create a modeling tool that can be used by neuroscientists to run experiments, test hypotheses, and analyze the effects of drugs more efficiently than they could using real brain tissue.

The model of part of the brain was completed last year, says Markram. But now, after extensive testing comparing its behavior with results from biological experiments, he is satisfied that the simulation is accurate enough that the researchers can proceed with the rest of the brain.

"It's amazing work," says Thomas Serre, a computational-neuroscience researcher at MIT. "This is likely to have a tremendous impact on neuroscience."

TechReview

The neocortical column is considered the functional building block of the mammalian cortex--a logical unit of brain organisation to begin a useful brain simulation project.

The project began with the initial goal of modeling the 10,000 neurons and 30 million synaptic connections that make up a rat's neocortical column, the main building block of a mammal's cortex. The neocortical column was chosen as a starting point because it is widely recognized as being particularly complex, with a heterogeneous structure consisting of many different types of synapse and ion channels. "There's no point in dreaming about modeling the brain if you can't model a small part of it," says Markram.

The model itself is based on 15 years' worth of experimental data on neuronal morphology, gene expression, ion channels, synaptic connectivity, and electrophysiological recordings of the neocortical columns of rats. Software tools were then developed to process this information and automatically reconstruct physiologically accurate 3-D models of neurons and their interconnections.

The researchers now think they have their neocortical column model well enough perfected to begin working on an entire mammalian "brain." They think they can model a mammalian brain realistically within 3 years, but respected neuro-researcher Christof Koch says "not so fast!"

However, none of these results have so far been published in the peer-reviewed literature, says Christof Koch, a professor of biology and engineering at Caltech. And this is by no means the first computer model of the brain, he points out. "This is an evolutionary process rather than a revolutionary one," he says. As long ago as 1989, Koch created a 10,000-neuron simulation, albeit in a far simpler model.

Furthermore, Koch is skeptical about how quickly the brain model can progress. Any claims that the human brain can be modeled within 10 years are so "ridiculous" that they are not worth discussing, he says.

Rat brains have about 200 million neurons, while human brains have in the region of 50 to 100 billion neurons. "That is a big scale-up," admits Markram.

source

The simulation is at a cellular level, and the researchers want to go deeper to the molecular level. This will put a tremendous strain on the computational infrastructure of the system. And it is not clear what is to be gained at this early stage by going to molecular resolution. Particularly when the cortical function appears to be at least partially based upon oscillatory phase-locking of assembles of neurons, such as columns and columnar groups.

Perhaps the Swiss researchers' "bottom-up" approach, combined with "top-down" approaches by people such as Jeff Hawkins, will begin to simulate some of the function of the human neocortex within the next 15 years. Perhaps.