Wired for memory: how your brain remembers by completing patterns
CA3, you complete me
You’re a beautiful machine. Watch:
“To be or not to __”
“Take a look at the lawman, beating up the wrong ___, oh man…”
“You can take your ___ and shove it up your ____”
Automatically, your brain filled in all those missing words. Whether there’s a right answer (‘be’, ‘guy’), or a psychologically insightful one (who got “cantaloupe” and “iguana” for the last one? Just me then), your brain fills in the missing bits. It completes patterns.
Your brain is the best pattern completion machine in the known universe. A fragment of input – a snatch of music, or a few words – and your brain fills in the rest. Think how often you turn hearing a fragment of a song, just a few notes, into immediately recalling the next bits: the instant attack of Beethoven’s Fifth (1); the picked notes of Stairway to Heaven; the sound like a monkey with its head stuck in a wah-wah pedal that David Draimen makes at the start of Down With the Sickness (altogether now: OH-WA-AH-AH-AH).
But how does it do this? (Your brain, not David Draimen – some things are beyond science). A fabulous new study has unravelled the machinery, right down to the level of individual neurons, and shown that an ancient theory is correct.
(Ancient in neuroscience being anything that predates Pop Tarts).
Here’s what we already knew. We know the hippocampus is important for some types of this pattern completion. It deals with episodic memory, the memory of stuff that happened to you. Especially recent stuff. So glimpsing a photo of the Eiffel Tower might immediately remind you of a wistful day walking by the Seine, or an evening stroll across Invalides, or being studiously ignored by a Parisian waiter when you wanted the bill sometime in the next hour or two. Or arriving in Leicester Square, and remembering there’s a great pub hidden just behind it, blissfully, mercifully free of tourists shouting “Hey Bob! HEY BOB! THEY’VE GOT A TACO BELL!”. Again, these are your brain filling in reams of information from a small fragment of input.
Within the hippocampus, we already knew the important bit was likely to be area CA3. Many theorists had noted this seemed to have just the right wiring. (The first to write it down in some form was perhaps David Marr, in 1971). The main neurons in CA3 have the unique property that they not only connect onward to the next bit of hippocampus (CA1), but also they connect to each other. And these connections are excitatory.
The theory then, is simple: if a few of these neurons fire, then they will tend to excite the neurons they are connected to; which are also connected together, so will tend excite each other, and keep the whole thing going. So a small input, making a small number of neurons fire at first, can get turned into a large set of neurons firing together. This is pattern completion! So if all those neurons are storing a memory (2), then a small, fragmentary input is recalling the whole memory.
Guzman, Jonas and colleagues’ latest paper in Science asked a simple question: great theory, but is the real CA3 actually wired to do this? For example, does it have enough connections between neurons in the first place? And if it does, are they strong enough for the neurons to make each other fire? And if not, what else is going on?
They found that, no, there are not enough connections. In fact, there are hardly any connections at all: less than 1% of the pairs of neurons they checked were connected together. Finding this out was nothing short of an act of scientific heroism. To find out if the neurons were connected, they used patch-clamp recording (“Susan” to regular readers): tiny glass electrodes that attach (“patch”) to the surface of the neuron, and allow the recording of every flicker of activity in that neuron. Neurons are tiny, microscopic – the neurons they recorded from have cell bodies that are typically 15 microns across. Attaching a glass electrode to something that tiny takes extraordinary skill and patience. In some recordings they did this for up to eight different neurons at the same time. And they did 1102 separate recording sessions.
In each of these extraordinary number of recordings they could test if their patched neurons were connected. They stimulated one neuron, and saw if any of the other neurons responded. The vast majority of the time – 99.08% of the time – they did not. Only 0.92% of the time was there a response. How crazily stubborn do you have to be to find the very first response? That, my friends, is one reason why this paper was selected for the rare privilege of being published in Science – out of sheer respect for the mental toughness of the authors.
Hang on, you said that wasn’t enough connections – so after all that work, it failed? Yes, at first. They built a computer model of how CA3 does pattern completion, a type of model that has been built many times before but with two key differences: they made it enormous (the size of a rat CA3 – 330,000 neurons), and connected the neurons together according to their data – so that only 1% of all possible pairs of neurons were connected. And it failed. It failed to do pattern completion because there were not enough connections to turn a small input to a few neurons into a lot of active neurons. The neurons that are turned on by that small input just aren’t connected to enough of the other neurons that are part of the same memory.
But in their recordings they’d noticed something really odd about how the neurons were connected together, and this oddness saved the day.
What they’d noticed was that the neurons were not connected together at random. They found that particular patterns of connections between two or three neurons occurred far more than expected by chance. Patterns like these:
As ever in science, a nagging oddity like this turns a scientists’ thoughts to: what is that for?
What if, they wondered, these “motifs” were there in the real CA3 precisely to solve the problem that there were not enough connections? So they returned to their model and, without changing the number of connections, added in these motifs, these patterns of connections between neurons. And, lo! The model of CA3 could now do pattern completion. It could recall a memory given just a fragment of that memory as input.
So the answer was: yes, the real CA3 can do pattern completion. The ancient theory was correct. But there are all sorts of things we still don’t know. How, for instance, do these motifs actually help? The paper offers one clue: if they left the “chain” motif out of the model, it failed. So, somehow, the existence of many sets of three neurons linked in a chain allows pattern completion to happen.
We can take guesses as to why. And my guess would be that we do not specifically need chains of three neurons; rather these chains are a signature of loops in the real CA3. That is, if we take a chains of chains that end up linking each neuron back to itself, we get a loop: a set of neurons that connect to each other in turn, and end up connecting back to the start neuron. That way, if only a small number of these neurons gets an input, they can each help make the next in the loop fire, which will make the next fire, and so on. Completing the pattern. But this is just a guess, an example of how any bit of good science inspires ideas, hypotheses, creative offshoots that could be tested, probed, added or discarded.
Your memories are you. You sniff a glass of whiskey and are transported back to a bar in Edinburgh where you’re trying to avoid eye contact with an itinerant bag-pipe player. You hear a snatch of Tame Impala’s Elephant, and your mind floods with the sights, sounds, and smells of that festival where you met the love of your life (though, worryingly, the fact they hadn’t washed for three days turned out to have nothing to do with being at a festival). So the next time, thank your CA3. It completes you.
Want more? Follow us at The Spike
Twitter: @markdhumphries
(1) Technically, Beethoven’s Fifth is not quite instant – the very first note is silence: the “dun dun dun duh” starts after the first beat. Just think how much more useful stuff I could do if I didn’t have this useless trivia floating around in here.
(2) Storing a new memory is also simple, in principle. When a new set of inputs happens – when you glimpse the Eiffel Tower for first time – these will activate a set of neurons in CA3. Doesn’t matter for now which ones – any old neurons will do. Because the neurons are active at the same time, so the connections between them will likely get stronger. And the more active they are together, the stronger those connections will get. Voila: we have a memory stored as a set of neurons strongly connected together, which can now be recalled by, you guessed it, input to just a few of those neurons. Cool, huh?
