David Marr: neuroscience’s Einstein

A case-study on the invisible hand of theory in biology

Not David Marr. Source: http://silverbased.org/2008/01/

Your challenge: name a scientist from the 20th century. 
“Well, Einstein obviously. His name’s in the title”.

Smart-arse. How about one more?
“OK: Feynman”.

Any more? Tick all that apply: 
“Schrodinger” (didn’t like cats)
“Heisenberg” (unsure of where he was and how fast he was going) 
“Dirac” (Auntie matter) 
“Hawking” (Wheels)

“What about that Higgs guy?”
Thanks, I’d forgotten about him.

What do all these have in common? White, male, mostly dead (as in, Higgs and Hawking are not dead; not the Princess Bride sense: we’re not about to revive Dirac by sticking a chocolate-coated ball down his throat. No: all we can do now is go through their pockets and look for loose change). (Loved the film? Read the book; it’s beautiful. Read the book? Watch Stardust; a noble attempt to capture the same adult fairy-tale aesthetic).

Where were we? Ah yes: White, male, mostly dead. And theoretical physicist. Every single one.

A modern lament in science is “Where are the Einsteins?” Which roughly translates as: “where are the lone geniuses who revolutionise research?” We’ve just met the answer. Lone geniuses who revolutionise research are all theorists. But no field other than physics takes theory as the bedrock of its research. So no other field recognises its lone geniuses — all our Einsteins are lost.

Neuroscience had one, who flamed brightly but all too briefly. His body of work is jaw-dropping. I’d never met him, and you’ve never heard of him.

He was David Marr.

Let’s do the Einstein checklist. Miraculous achievements in rapid succession at a young age? Check.
In three years Marr published three extraordinary sole-author papers, on computational models of the cerebellum (1969), cortex (1970), and hippocampus (1971). Each proposed a ground-breaking theory: that cerebellum learnt to correct motor errors; that cortex can act as a general purpose learning device; that hippocampus can act as a temporary storage for events, and recall them even if some information is missing. All these theories have found their way into the very foundations of research on these brains regions, even if the current practitioners have no idea where they came from.

He published these papers between the ages of 24 and 26. (Writing this piece is doing nothing for my self-esteem, I’ll tell you that now).

Moved to working on a grand problem of the field? Check.
After these papers, Marr turned to work on vision. “Seeing” to the rest of the world. Vision is the Goliath that bestrides neuroscience. Partly because it is our primary sense, so we find it inherently fascinating. And partly, prosaically, because the region of cortex where signals from the eye first arrive is the easiest bit to record from. He wrote a series of seminal, influential papers. His work culminated in the major book simply titled Vision.

Changed an entire field of research? 
Check. 
Marr’s true legacy was establishing a way of thinking for the entire fields of computational and theoretical neuroscience. The brain is a ridiculously complicated machine. We can only ever hope to sample from a tiny fraction of its activity, in a tiny fraction of a bit of brain, during a tiny fraction of all the things the brain will ever do. In the face of this hopeless mismatch between our data and reality, Marr proposed that the only to make sense of brain data was to break any brain problem into three levels: Computation; Algorithm; Implementation.

Or: the problem being solved; the steps to solve it; and the machinery that actually does all that.

An example: I want some cake. (There are times when I don’t want cake. Those times, I want crisps). The computational problem: make a cake. The algorithm: the recipe. The implementation: me, the cook, using a bowl, spoon, knife and gas oven; or mixer; or electric oven; or spatula; or….

Marr’s key insight was that each level, on its own, is not much use. For any given computational problem, there can be different algorithms to solve that problem. If I want a sponge cake, there are many different recipes. In turn, for a given algorithm, there are many possible implementations. I can use different tools to do any one recipe, hand blenders or mixers, spoons or spatulas, or bung it all in the food processor and make a cup of tea. (Those who watched the Great British Bake-Off 2016 final may recall the technical challenge: make a Victoria Sandwich. The recipe said, only, “make a Victoria Sandwich”. Lo and behold, left unconstrained, each contestant used a different set of tools to make the exact same cake.)

(The Great British Bake-Off is the quintessential British TV experience. Lots of people in a marquee tent, baking competitively, and drinking tea to calm the nerves. A huge audience tunes in, lapping up its genteel ambiance. That, and the fact that during the first final, in 2011, one cut-away shot panning over a lawn purposely zoomed in on a giant squirrel scrotum. [Warning: contains a picture of a giant squirrel scrotum])

Marr said that we need to pay attention to all three levels. Each level constrains the others. If the computational problem is “make cake”, I don’t then look up a recipe for casserole. Or buy some bricks. If the recipe is cake, I don’t use a cement mixer to make it. (Well, I could, but some have complained about the state of the kitchen afterwards).

But just as important, and what Marr stressed, is that if we attempt to understand just one level, without thinking about the others, we are hopelessly lost.

Imagine if you motion-captured my cake-making performance, but didn’t know I was making a cake. You can then watch hours of wire-frame, stick-man video of me moving about. Occasionally one of my arms picks up an object, and the other arm rotates furiously above it. What does that mean? Am I pretending to be a helicopter? Am I having a seizure? By just looking at the implementation, without any idea of the algorithm, or even the computational problem, understanding what I’m doing is nigh-on impossible.

Similarly, if I look to solve a computational problem about the brain — like “how do we decide” — without thinking about the algorithm and its implementation by neurons, then I am going to waste a lot of my time on solutions that are literally impossible for the brain to implement.

This mistake Marr identified, of looking at one level of the brain without thinking about the others, is made over and over and over again. It’s costing us a fortune in wasted research money.

Marr’s legacy lives on thanks to a group of researchers who worked with him or drew directly on his influence. A new book was just released to mark what would have been his 70th birthday. Marr died in 1980 age 35 of leukemia. I turn 40 next year. It would be amazing if any of my work to date ends up being 1/100th as influential as anything Marr did.

“Are there other theorists in neuroscience at this remarkable level?” Yes. For example, Warren McCulloch, Walter Pitts, Wilfrid Rall.

“Who?” Exactly.

(For an accessible discussion of Marr’s paper on hippocampus, its context, and legacy, read the lovely piece by Willshaw, Dayan, and Morris).

If you liked this, please click the 💚 below so other people can read about it on Medium.