Your thoughts on simulating a human brain?

Ed Yong recently published a great article on BBC Future entitled, “Will we ever…simulate a human brain?”, summarizing the debate over the Human Brain Project. I was interviewed for the piece, and explained why I am skeptical that the project will accomplish its goal of realistically simulating a human brain within 10 years. The article presents several different viewpoints, ranging from researchers who think it won’t happen to others who say it may but question both the approach and the relevance of the results that could be produced. I highly recommend you read Ed’s article.

And then come back here, because I’d like to know what you think. BBC Future only allows people to comment by logging on to their Facebook page or messaging them on Twitter,   neither of which will produce a very coherent comment stream. So, I’d like to have the discussion here.

  • What do you think about the Human Brain Project, and the recent announcement that it will be awarded 1 billion euros by the European Commission?
  • Do you think the HBP’s goal to simulate a human brain in 10 years in realistic?
  • What do you think will be the biggest challenges in simulating a human brain?
  • What do you think could be some of the most exciting results, if the HBP is successful?

I’d love to hear opinions from people of all backgrounds, so please leave me your comments!

Advertisements

12 Comments

  1. I think most of the claims for the HBP are a little silly but I also think it is important to be funding neurons more than hadrons at this point. Let he or she with a better plan than Markram step up then. I think we need a decent CAD model to explore what even the macroscopic anatomy really is, maybe $1 MIllion?

    1. Thanks for your comment, John! For others who are interested, you can read more on John’s views about the HBP here. I agree that other researchers should step up to the plate with their ideas; it is easy to criticize, but not so easy to present viable alternatives. However, I take issue with the idea that “it is important to be funding neurons more than hadrons”. First, I’m reluctant to value one type of science over another. The emphasis should be on whether the science is solid and can further our understanding. One big difference between the HBP and the Large Hadron Collider projects relates to what Zen mentions below. While LHC researchers have formulated testable hypotheses, the HBP seems to have none (or at least none they have shared up to this point). In general, I’d place my bets on the project with testable hypotheses.

      1. Good points Erin. I suppose a physics person I should show a little but of loyalty to big physics. I only used the Neuron simulation environment briefly in the late 90s so don’t know much about today’s extensions in common use. At that time I could use the Genesis simulator to model Ca diffusion and then “grow” new dendritic compartments to crudely represent some kind of plasticity. I think a key point of the brain that needs to be part of any simulation is that there is so much motion going on as far as vesicles, spines and organelles. Any good model of the brain should be able to take an electron micrograph picture and come up with a pretty good guess of what that same anatomical slice would evolve to look like seconds or minutes later as well as previous. In other words, which direction are things moving, expanding and contracting? To Phillepe’s point about glial cells, I don’t know of any electrical model that has ever incorporated them. Yet who among us has any idea if every myelin sheath down any given axon wraps the same direction, if every sheath of any given oligodendrocyte wraps the same way, or if all axons in one hemisphere wrap the same direction, or if axons torque when they fire to assist in myelin wrapping? To the questions regarding the biggest challenges in simulating the brain I would point to a couple further thoughts http://www.extremetech.com/extreme/141507-how-to-create-a-mind-or-die-trying

  2. “1. What do you think about the Human Brain Project, and the recent announcement that it will be awarded 1 billion euros by the European Commission?”

    I’m currently unimpressed, but then, I haven’t read the full proposal. Maybe it’s more measured and reasonable than the summary sounds.

    “2. Do you think the HBP’s goal to simulate a human brain in 10 years in realistic?”

    Depends on what level of detail they hope to achieve. One problem is how do you establish what a successful simulation is?

    “3. What do you think will be the biggest challenges in simulating a human brain?”

    Capturing the variation between, and dynamic physiological changes of, single neurons.

    “4. What do you think could be some of the most exciting results, if the HBP is successful?”

    I don’t know. There doesn’t seem to be any question or hypothesis here, so I can’t even figure out what a result would look like.

    1. Thanks for your comment, Zen! A few notes:

      “I haven’t read the full proposal. Maybe it’s more measured and reasonable than the summary sounds”

      What I understand to be the full proposal submitted to the EC (unless I’m missing something) can be found on the HBP’s website (to the right under “HBP Report”). In my reading, I found no new significant information to indicate that project summaries have been leaving something out. In fact, I was surprised and disappointed that the report does not contain any preliminary data – no recordings, which will be the bases of the information they will put into the model, nor simulations. True, there are publications from the Blue Brain Project, but I would have expected some additional preliminary data in any grant proposal, especially one of this size.

      I agree with you that a huge problem with the HBP is a lack of criteria for what will constitute successful simulations, as well as a lack of specific questions and testable hypotheses. It makes it very difficult to judge what understanding could come from the project.

  3. It’s a little outside of my area of expertise so I have more questions than meaningful comments. I study artificial neural networks but as a machine learning tool, not as a way to simulate the human brain.

    Is it possible to model the brain? One thing I learned by doing science is that, while science opens a lot of doors, it closes a great deal of doors too. There is a possibility that modelling the brain is impossible, for example if we cannot capture the essence of a neuron without modelling everything down to the molecule.

    Do we know enough about neurons? I’m not fundamentally opposed to “bottom-up” approaches but they require a very good understanding of how the “units” work. AFAIK, efforts to model cells are in their infancy, it’s just very hard to find the right abstraction for such complex systems of molecules. We have little margin or error here: what if our model for the neuron fails to capture some subtle behaviour that happens to matter a lot to model the brain correctly?

    Do we know enough about the brain’s architecture? Real neurons are in a dynamic network where they interact with glial cells and other cells. I suspect that, as it is often the case, we underestimate the complexity of these interactions and I wonder how they plan to deal with them (directly or indirectly?).

    Do we have the computational power to simulate the brain within 10 years?. I seriously doubt it. Even if we had a “yes” of the previous three questions, I doubt we have the computational power to solve this kind of problem (yet).

    Impact on funding. This is the thing I’m most scared of: the project fails miserably because we are not ready for it and lead to another rough period for funding on projects related to intelligence.

    1. Thanks so much for your comment, Philippe! Well put. Your concerns echo many of my own. I’ll try to address them best I can:

      “Is it possible to model the brain?”

      I think the answer to this depends on what type of model you want and what level of detail. I have no doubt it is possible to model something that looks similar to a brain and reproduces some of its behaviors. Attempts have been made, see for example Izhikevich’s work . But to my knowledge, most of these models are phenomenological, meaning simplified models which reproduce the basic changes in a neuron’s electrical activity, but are missing crucial physiological details (e.g. integrate-and-fire neurons). These model neurons can mimic many firing patterns, but it is questionable whether we can extract mechanistic information from them given that many of the parameters do not correspond to measurements from real neurons. You can connect many of these neurons together and argue you have built a brain, but what new understanding will you get?

      Do we know enough about neurons?

      Short answer? No. We are missing a lot of information about the full complement of ion channels expressed in a neuron, how channel proteins interact, how expression changes over time, experience, etc. And that’s just if we want to understand the basic electrical activity of neurons. As you and John point out, there is much more going on in neurons than just the movement of ions.

      Do we know enough about the brain’s architecture?

      Also no. Glial cells continue to be relatively ignored by many, despite increasing evidence that they are crucially involved in many complex processes, including memory formation. As far as I’m aware, the HBP has no plans to incorporate additional cell types in their model. And even if they do, modeling glial-neuron interactions would be a significant challenge considering how little we know. It will be challenge enough to accurately model neuron-neuron interactions.

      Do we have the computational power to simulate the brain within 10 years?

      I think this will probably be the least of the HBP’s problems. They have significant computing power and now the funds to increase it substantially. Within the 10-year time frame, I also imagine computing capabilities will increase much faster than our neurophysiological knowledge.

      Impact on funding.

      Agreed. Funding is a huge problem. First, irrespective of the project, I don’t agree with concentrating so much money in one place, especially in an age of limited funding. I would rather see wealth spread around so that many labs with great ideas can carry out research. And yes, if the project fails, it could have severe consequences. I find many neuroscientists are skeptical enough about modeling as it is. A high profile failure could seriously damage the funding prospects for future modeling projects.

  4. Article: http://www.newscientist.com/article/mg21729036.800-why-were-building-a-1-billion-model-of-a-human-brain.html

    Stated goals in this article (paraphrased):

    “We could make a robot learn, and trace the chain of events from molecules to cognition.” That’s speculation, not an hypothesis.

    “Build a database of the biological signatures of disease.” Not an hypothesis. Not clear how will a humongous model help that quest.

    “We want to build neuromorphic computers that learn the way the brain does.” Not an hypothesis. Why we want things to learn the way a human brain does? Human brains are crappy at a lot of tasks. Let’s not hobble our machines.

    “One hundred billion neurons is numbers, not complexity.” First, that’s about 15 billion neurons too high, according to recent estimates, which doesn’t fill me with confidence you know what you’re doing. Second, I agree that’s not complexity IF all those neurons are interchangeable widgets. THEY ARE NOT.

    Comparisons to the Human Genome Project are not convincing. We knew what the end product would be; a sequence. Still not at all clear what a “model” of the brain will be.

    1. Couldn’t agree more. Thanks for this break down! These are definitely NOT hypotheses. And variations on these statements are about as specific as I have seen the HBP get, including in the proposal they submitted to EC. I can easily imagine researchers submitting to NIH with such a lack of testable hypotheses and well-defined aims receiving the “This is a fishing expedition” kiss of death. So, why not the HBP? It’s not clear to me what the EC saw in the proposal that led them to fund it.

  5. Hi Erin–this comment is off-the-cuff–have only about 3 minutes to write–but I enjoy the topic! I haven’t read the article, but have kept-up with the story. No simulation is a complete simulation just as no map is of one-to-one size as the geography (see Borges story). They are making choices in what to implement and what not to (perineuronal nets, anyone?) as they build their sim-brain from the bottom-up. But Markram knows his cortical connections and cell types very well, and I am excited because I think that this will become a great resource for putting experimental results into context, and for asking new questions about the effects of receptor X, Y, Z, or neuron type W, on cortical network processing. I agree that you can’t always get this from small models, because qualitative changes can emerge from quantitative additions. Ultimately, we want a model of everything, to be able to predict how physiology, cognition, and behavior happen as they do. But a model of everything, even if accurate, won’t grant understanding unless we remove components and see whether they have any effect on particular processes of physiology, behavior, cognition. And if we don’t know which components are important, then how do we know what to leave-out of the simulation (do we need to include temperature dependence of membrane dynamics?)? The Blue Brain project has the potential to provide a very substantial foundation on which to ask specific questions about brain function and operations to compliment experimental results, but that will only be the case if researchers have easy access to the shazillion supercomputers that support it and can easily add or remove specific receptors, neuron types, connectoin strengths, etc. So: I’m cautiously excited. And I am way over my allotted 3 minutes!

    1. Thanks for commenting, Nathan! (Sorry for the delayed reply.) You make some great points and I’m glad to hear from a neuroscientist who is optimistic (even if cautiously) about the project. I agree that a big challenge for HBP will be how to share the fruits of their labor with other researchers. In the HBP’s report to the EC, they talk a lot about the importance of data sharing, but I didn’t see any details on how this will be implemented. There is also no mention, as far as I can see, as to exactly how they will make the model usable by other researchers. Assuming they are successful in creating a functional model brain, researchers throughout the world should be able to run simulations and test hypotheses. But to do so, they will need the supercomputing power of the HBP. How will this be made accessible? How will simulations be controlled? How will results be stored? I would like to see a very detailed data management plan, as is usually a part of any NSF/NIH proposal. Maybe the HBP has such a plan. If any other readers know whether this exists and where we can access it, please leave a link in the comments.

      But this is the cart before the horse. Can they build an accurate model of the brain? As you say:

      Markram knows his cortical connections and cell types very well

      Absolutely. No argument there. And certainly data emerging from Blue Brain so far has increased our understanding of cortical cells and network organization. But what about sub-cortical regions? Will the principles of organization, rules of gene expression, firing behaviors, etc. studied within small columns of cortex hold when we need to represent other areas of the brain? I’m not so sure.

      And even if a model brain is constructed, it’s not clear how much understanding we will get from it.

      a model of everything…won’t grant understanding unless we remove components and see whether they have any effect on particular processes of physiology, behavior, cognition. And if we don’t know which components are important, then how do we know what to leave-out of the simulation?

      Absolutely. A brute-force approach, taking out components one-by-one, is not realistic in a model with the proposed complexity. And in any case, I don’t think it will be as simple as removing single components and seeing what happens. If something that looks like learning, intelligence, consciousness (however you define those) emerges from the model, it is likely to be due to the interaction of many different components. Testing all the potential interactions won’t be possible.

      I could go on, but now I’m way over my allotted time, by at least an order of magnitude :).

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s