How to spend a billion euros on the brain
At the core of the Human Brain Project is a model of cortex. Its flaws show what the project should have been.
Say I wanted to spend a billion euros to understand the human brain. I could spend it on crisps and beer for all my friends, while we sit in a bar thrashing out theories, scrawling intensely on the back of napkins. I’m told that’s not allowed under current EU funding rules (the beer, not the napkins). I could spend it finding a way to record many single brain cells in humans without needing to insert anything from the outside – non-invasively, as it’s called – and skip the animal step. Though not without raising a host of difficult ethical questions (“You want to stimulate what part of the brain? And that will make your participants collapse uncontrollably to the floor? No way!” Hypnotists don’t have these issues.)
The Human Brain Project chose to base their billon euros on a computer model. But a recent paper in Science raises awkward questions about this choice. It suggests the core model has a fatal flaw.
The Human Brain Project has been exceptionally controversial every since it was proposed. Some controversy was inevitable. Any project that proposes to spend up to one billion euros towards understanding the human brain will get criticised. But this caused a civil war. Much has been written about its flaws, both in its scientific aims, particularly the lack of them, and in the way it is run.
Or, rather, was run: thanks to the unceasing criticism, the Human Brain Project has been repeatedly scrutinised and deeply restructured. Things are looking up. (And note that, despite the name, the press releases, the media coverage, and its own website, the Human Brain Project is not really about understanding the human brain. But that’s a story for another time).
Strangely little has been written about the flawed scientific choices at the heart of the project. Its centrepiece is a giant computer model of a small bit of cortex. It is also the most mature part, originating in the preceding Blue Brain Project. Late last year, 9 years since the launch of that 5 year project (yep, not a typo, but an accurate indication of how tough science is), we finally got to see the fruits of its labour.
What it built was a tiny piece of rat cortex, a little stack of 30,000 neurons. Specifically, the part of rat cortex that activates when you touch its left back leg. Of a 14 day old rat – a young teenage rat, not fully developed. Not the obvious choice, I grant you. But you have to start somewhere.
Why there? The model aims to represent every neuron in all its glory – every twist of its branches, every little bump emanating from them; and every connection between them. And then to represent every bit of activity inside them. To do that, they needed to record every flicker of electrical activity throughout the neuron – in its input branches, its body, and its output cable. The techniques needed to do this only work well in young bits of brain. (Called “patch-clamping” – but as that’s neither evocative nor memorable, we could just as well call it “Susan”). So it had to be 14 day old, teenage rats.
Or did it? In doing this, they took a big gamble – and one that did not pay off.
At about the same time that the Human Brain Project paper appeared in Cell, a paper appeared in Science describing the shape, size, wiring, and electrical activity of all neuron types in visual cortex. Sound familiar? And here’s the kicker: this team cracked the problem of using the recording technique (“Susan”) in adult bits of brain.
Their data show big differences between adult and teenage rats. For example, they report that in adults some local “inter” neurons no longer connect to the main neurons of cortex; meaning: these main neurons are missing some inhibitory signals. Another: they report much sparser connections between the main neurons in adults; meaning: these neurons are less able to communicate with each other. Both potentially mean big differences between teenage and adult rats in how the whole circuit responds to its inputs. Both certainly mean big differences in which connections between neuron can appear, disappear or change strength during learning.
The core model of Human Brain Project, the model claimed to be used for studying Alzheimer’s among other disorders, is then not even a model of the adult rat, let alone a model of the adult human.
Let’s be clear that the Human Brain Project is a behemoth. Within it are many good, talented scientists and engineers doing quality work. There’s a whole project devoted to “neuromorphic” chips, computer processors that use model neurons instead of binary transistors. There’s one devoted to embodied robotics, where models of brains are tested in real environments to test the limits of our understanding. There’s a large ethics section, very welcome in the current climate of increasingly hysterical claims over the capabilities of AI. The core model itself is a technical marvel. Building it required gathering a vast amount of data on the detailed shape, size, and wiring of neurons, alongside all the electrical activity measurements. To synthesise all that data into a model required developing many new techniques in using and extrapolating from experimental data. But the Project has a potentially fatal weakness at its core: a model of a teenage rat brain being used as the foundation for understanding an adult human brain.
One might ask why it did not spend its mountain of euros on the basic science question. Why it did not make its very first aim cracking the problem of applying its core data gathering technique (“Susan”) to adult tissue.
Moonshot programs reach for unreachable targets (the moon) by solving unsolvable problems (how to fire people in a tin can across space onto a revolving rock). Such programs can fail; are expected to fail. Huge funding on huge time-scales allows risk taking that is way beyond the reach of individuals or individual labs.
The Human Brain Project should have been a moonshot program, but it is not. It has up to one billion euros, and more than ten years. It has an unreachable target – understand the human brain. But it did not, and does not, have a core set of “unsolvable” problems that would unlock the future of neuroscience. Here are some.
- Get “Susan” – “patch-clamping” – to work in an adult rat, in any bit of brain, at any age.
- Find out how connections between neurons change in adults. Changes to these connections are how we learn. But all we know for sure is, again, about teenage rats. The few times we’ve looked at adult brains, the rules for how connections change seem to be different. So what are they?
- The dark neuron problem. Most neurons in cortex do nothing most of the time. Literally nothing. When we watch hundreds of neurons in visual cortex, and show them a picture, most of them do not respond at all. So why are they there? What are they for? They must have a function, even if it is just “back-up” (in the hard drive sense, not the cop sense) – evolution does not tolerate waste.
(Incidentally, the dark neuron problem means that most of the neurons in the Human Brain Project’s model must do nothing, in order to accurately reproduce rat cortex. Its a bit odd spending all that time and money building something that we know does nothing, without knowing why it does nothing).
It can still be a moonshot program. It has the time and the money. In fact, it now gives away much of its money as research grants to labs outside of the Project. So if it’s too late to do the hard work of restructuring the entire Project around the moonshot challenges, then make those grants the moonshot: offer them for specific, tough challenges; offer them with failure welcomed, embraced, celebrated.
(For a great account of the Project’s political crises and civil war, see Leonid Schneider’s account here.)
If you liked this, please click the 💚 below so other people can read about it on Medium.