Is neuroscience doomed to go round in circles?
On the disturbing lack of direction in neuroscience
Neuroscience is going through another bout of soul searching. More specifically “systems” neuroscience, the field that wants to know how neurons do stuff: how they encode a view or smell a flower; how they move your arm, or plan your wedding. (Yes, there are huge chunks of neuroscience who do not care how neurons do what they do. Those areas of neuroscience care only about either (a) what neurons are made of, or (b) how neurons respond to drugs). A torrent of opinion pieces have recently poured out on the best ways to do systems neuroscience, of how to interpret the complex data it obtains. All in high profile forums, all widely read. We’ve had pieces on how best to co-ordinate efforts to tackle big questions — which seems to be leading to action in the form of the grandly-named International Brain Lab. On how to interpret neural codes. On the perils we face when interpreting causality in experiments (we’ll return to that one in a later piece).
One piece in particular has provoked much discussion. It states bluntly and boldly that “Neuroscience needs behaviour.” It is noble in principle. It raises broad issues that all neuroscientists should know about. But it is frustrating in its execution. For much of what is right is well known or trivial. And the fact someone felt it needed writing and publishing is depressing.
It leads one to wonder if neuroscience is doomed to go round in circles.
Why does neuroscience need behaviour? In their piece, Krakauer and friends warn against neuroscience going the way of molecular biology, pursuing ever smaller, finer-grained decompositions of a system, without ever approaching something that looks like “understanding”. Can’t argue there.
Their claim is that to know what neurons are doing we must know what behaviour the animal is performing. (I’d wager the guys and gals working on, for example, coding in the retina or on the pacemakers of the circadian circuits or on sleep don’t believe this, and think that they are perfectly capable of understanding what their neurons are doing without behaviour; but let’s not split hairs). Do we?
Trivially, yes. If by “understand” we mean “predict” then we need something to predict: and behaviour gives us something to predict from the activity of the recorded neurons. More accurately, it gives us something to postdict: we record both behaviour and neural activity together, then go back and attempt to predict something about the recorded behaviour from what the neurons were doing just before or during that behaviour.
(I’ll note that most of their examples don’t make this point. Their examples are observations of behaviour in one context being used as interpretation of neural activity in another. Not jointly characterising neural activity and behaviour together. Not making or testing predictions. And this is routine: we’ve done it.)
Their main thesis of neuroscience needing behaviour is not new. Here’s Matteo Carandini making that point in 2011. And a paper on the behaviour of sea-slugs from 1986, that spends much of its final section chastising neuroscientists for incorrectly interpreting what their recorded neural activity means for the sea-slug’s behaviour. And of course, David Marr in his 1982 book Vision (who Krakauer and friends quote at length). And this review entitled “The Neural Control of Behavior” from 1978. Yes, we know: neuroscience needs behaviour.
In 2014, the United States’ National Science Foundation published a white paper on the “Major Obstacles Impeding Progress in Brain Science”. They asked 78 different labs and leaders the same question: “what is a major obstacle impeding progress in brain science?” Practically all of the 78 responses came from researchers in systems neuroscience — the very people who try to relate the activity of neurons to the outside world.
Only one response chose the measurement of behaviour as the obstacle facing neuroscience. And “behavior” appears only 76 times in the entire document, with many of those in just three entries. And many of the rest are not referring to animal behaviour, but to the “behaviour of the brain”, as a metaphor. There we have 78 thought leaders, experts in relating neural activity to the outside world, and only one thought the obstacle was behaviour, and only three thought behaviour worth discussing at length. Something of a mismatch huh?
(And notably, the three who mention behaviour at length work on “simple” animals — zebrafish and leeches — not mammals. Those who observe the phenomenal complexity of the small neural circuits that control the crawling of a leech can extrapolate to the complexity of the rat brain, and shudder).
Where has this mismatch come from? Why do almost all of 78 thought leaders in systems neuroscience think the obstacles to progress are recording more neurons or tracing the wiring between neurons, and not capturing behaviour?
I’d wager it’s because advancing recording and tracing technologies is primarily a technical challenge. The goals are clear, the definitions of success self-evident: record the neurons, trace the connections. But for capturing behaviour, the goals are not clear, what defines success is not self-evident. So the prospect of capturing behaviour the same way we capture the activity of neurons is so outlandish that it doesn’t even occur to people that it’s a problem.
Take a mouse exploring a box. It’s a dull, drab, grey enclosure, the walls broken only by a light, a lever, and a spout for water. The sort of box in which a mouse is trained to press the lever after the light goes on — and if it does, to get rewarded with a drop of water. A report of this experiment will likely only mention how long it took the mouse to learn, reliably, that it needed to press the lever. Days, compressed down into a single number.
Real behaviour is complex, multi-faceted, full of ifs, buts, and maybes. Prevarication and indecision. Sudden leaps and diversions. The mouse sniffs the lever, the spout, the corners; it scrabbles at walls, rearing on its hind legs; it pauses and twitches its ears; it grooms and scratches; it follows the walls and sits in the corners. It tentatively licks the flashing light.
If we want to know how the mouse’s brain is producing these behaviours, what must we do? Record everything, every little twitch and sigh? Even if we could, how do we turn all that into numbers — numbers that we can predict using the activity we recorded from the mouse’s brain?
Say that we could do all that (and there are tentative reasons to believe we can — for example, we have some good data now on what a mouse does in its spare time). That’s not even the hardest problem, the running-into-a-brick-wall-face-first problem. No, the hardest problem is we don’t know what the brain of the mouse is actually tuned for. We can observe all the fine-grained, hyper-detailed behaviour we like: but we have no idea of whether the mouse’s brain actually has any representation of that bit of behaviour. To find out, we need a theory for how the brain would produce such a behaviour — an algorithm for action — which then predicts how the brain could represent that behaviour. And then we can search for clues when we record both neural activity and behaviour together.
And what the brain of a mouse is tuned for will be rather different from what the brain of a ferret is tuned for. And that will be rather different from what the brain of a panda is tuned for. Because they evolved in separate niches, with different things trying to eat them, and different things they try to eat. If they care about different things in the world, you can be damn sure their brain cares about different things in the world. So an understanding of how an animal’s brain represents behaviour is intimately tied up with understanding how that animal makes a living. Algorithms for action need not be universal.
(The same is true for inputs to the brain. To know how the brain represents one single, solitary input — like a picture of a single line — we need to know everything about that input. We need to know the colour of the line and of the background the line is on. We need to know how bright the room is. We need to know what angle the line is at, and where it is in relation to the eye. Then we can record from many neurons while the eye looks at this dull picture, and take apart which neurons care about what. And even then, we don’t know what the brain is actually tuned for — what things the brain cares about. Because what a mouse finds most valuable for its survival is not exactly what a dog find most valuable for its.)
So we’ve reached full circle. Neuroscience needs behaviour. It needs behaviour because, without it, we have nothing to predict. Which means we lack perhaps the most basic demonstration of understanding. And accurately measuring the relevant behaviours is the biggest obstacle we face in neuroscience, because we don’t know how to measure it, nor what is relevant.
There are labs out there who know that we need to understand behaviour deeply. Some labs have made it their mission to deeply quantify the behavioural task they use before they report their neural data from that task. For instance, Carlos Brody’s lab deeply analysed how animals made decisions based on how many clicks they heard — and compared two species of animal, mice and humans (an analysis so thorough it was a Science paper). Joe Paton’s lab did a detailed analysis of how mice performed their task of waiting a set amount of time before pressing a lever — and with that thorough understanding, they then produced a rapid sequence of insightful papers on how the brain represents time. And very recently, Tiago Branco’s lab have taken this a step further, analysing in detail how a mouse’s natural behaviour — to flee to safety at the threat of a looming predator, at the imminence of an owl — adapts quickly to wherever safety happens to be. So some corners of neuroscience know we need behaviour. But only some.
That Krakauer and friends’ piece has achieved its primary aim is demonstrated simply through the fact that I’ve written this, and you’re reading it. They have provoked conversation. But their main points are conversations we’ve had before. For decades. And we’re dodging the big conversation: what would it mean to understand the brain? If we keep recording ever more neurons, and we keep recording ever more behaviour, will it help? We don’t know. The fact that we don’t know even what direction we should be heading in suggests we are missing some vital ingredient in neuroscience. We have the “neuro”; the science part, perhaps?
If you liked this, please click the ♥ so other people can read about it on Medium