When the chips are down, should I ask a neuroscientist?
The way we do neuroscience may be fundamentally flawed if we can’t understand a simple microprocessor using the cutting edge tools of modern neuroscience
Let’s start with a cliche: the human brain is the most complex machine in the known universe. 87 billion neurons; a trillion connections. Numbers so large that your favourite media outlet has to compare them with other, literally astronomical, numbers: “as many neurons as galaxies in the observable universe”; “more connections than there are stars in the Milky Way”. Great. Now we have four sets of numbers we can’t grasp. And a slight urge to buy a telescope (“ooo shiny”).
This ridiculous scale of the human brain gives a flavour of its complexity. And this complexity raises a very sensible, but seldom articulated question: are the methods we use in neuroscience able, in principle, to understand the brain? In other words, if we can scale what we do now to the size of the human brain, would they tell us how it works?
One reason this question hasn’t been asked much before is that it is not clear how to test the answer. How do we know if we’ve understood it, if we don’t already understand it?
Jonas & Kording recently offered a clever solution in a paper posted to bioRxiv (1), using a simple analogy. The brain is an information processor, one built of many parts with distinct functions. What else is an information processor, built of many parts with distinct functions? A microprocessor. So they reframed the question as: can we use the methods of neuroscience to understand a microprocessor?
Answering this question is made possible because some heroic team has written a complete software simulator of the classic MOS6502 processor. It accurately simulates the voltage in every wire and the state of every transistor, on every simulated processing cycle. And as we understand how a processor works – we could hardly build a note perfect simulator if we didn’t – so we know exactly what we’re aiming for.
The MOS6502 was at the heart of some classic home computers – the Atari VCS/2000, the C64, and the BBC Micro. No surprise then that its simulator can play some truly classic games. So Jonas & Kording had their MOS6502 “brain” produce three “behaviours”: Pitfall, Donkey Kong, and Space Invaders. They then applied neuroscience methods to analyse the activity of the transistors during each behaviour, to uncover how the processor worked.
They completely, gleefully, failed.
In one attempt, they took a textbook neuroscience approach to uncovering what bits of the brain do: they cut them out. (This approach already has known issues, articulated beautifully by Olveczky and colleagues). By eliminating transistors one at a time, they could observe if each of them disrupted a behaviour. For about 45% of the transistors, removing each disrupted all three behaviours; a neuroscientist observing this effect of cutting out neurons may well conclude “so these are for core functions – arousal perhaps; or memory. Or breathing.” For about 5% of transistors, removing each only disrupted a single behaviour. The classical neuroscientist interpretation: some of these transistors are for Donkey Kong; and some of these are for Space Invaders. Which would be barking mad. It’s general purpose processor – there are no transistors “for” a specific game.
In another attempt, Jonas & Kording used a causality analysis common to neuroscience to try and uncover the information flow across the processor. This is of course literally hard-wired: instructions are read into the decoder, which updates the register, whose values are then used by the Arithmetic Logic Unit to do simple operations – adding, subtracting, and so on. But the causality analysis, though it got elements right, could not find this information flow. Worse, it added pathways that do not exist. Even worse, it claimed the key routes of information flow were different between the three “behaviours”. Barking, barking mad. It’s a general purpose processor – there are not hardwired information flows for a specific game.
However simple or sophisticated the approaches they threw at the processor, they drew a blank. Is all neuroscience an utter waste of time?
Not quite. What Jonas & Kording are offering us here is caution. The brain is not like a processor in any deep way; and if we wanted to understand a processor, we probably wouldn’t use these methods to understand it. But their work does tell us that we do not think enough about what the answer – what “understanding the brain” – will look like. And the moral lesson: don’t ask a neuroscientist to fix your computer.
If you liked this, please click the 💚 below so other people can read about it on Medium.
Reference: Jonas, E. & Kording, K. (2016) Could a neuroscientist understand a microprocessor? bioRxiv. DOI:10.1101/055624
(1) the brave new world of the biology pre-print server. Physicists have been posting pre-prints – full papers ready for submission to a scientific journal – since the early 90s. But for biologists, the idea of showing anyone your work before it was published has been considered so crazy that the mere existence of bioRxiv prompted the New York Times to write: “Handful of Biologists Went Rogue and Published Directly to Internet”