Thread by @dlevenstein

We just pushed a big revision of our preprint, so it seemed like a good time for a thread re: what we found and the bigger picture of why I think it's interesting. (1/?)
biorxiv.org/content/early/…

It's 2018, so at this point we all know the brain is just a big deep neural network with some behavior attached.

Problem: In order to form long lasting memories, neural networks take forever to train... many training examples. But brains can learn things v quicky... few training examples, and still remember a long time.

One way we think the brain can do this is by using a short term store of the "training data" to train itself... during sleep! This is called the two stage model of memory consolation.

TL;DR A part of the brain called the hippocampus keeps track of important things during the day, and then re-plays them during the nonREM stage of sleep to burn them into the much slower-learning neocortex.

Theoretical work has shown that this strategy is a solution to the “catastrophic forgetting problem”, and it was recently used to reach human-level performance at Atari… DeepMind’s DQN agent replayed it’s best training games during offline periods to teach itself to play better.

But DQN had the luxury of some smart humans to hard-code the PROCESS of offline training... The brain has to do all this work using internally-generated, self-organized dynamics.

Useful to think of this like a phase diagram (you know - solid, liquid, etc). In one phase, the system navigates the world. In another it does stuff it needs to do so that it can navigate the world in the future... Subcortical neuromodulators bring the system between phases.

And finally we get to the main motivation for this work:
What is the state of hippocampal/neocortical populations during the NREM sleep?

So what are neuronal populations in the hippocampus and neocortex doing during NREM sleep anyway?

The hippocampus does sharp wave ripples (SWRs): occasional bursts of activity. It's during SWRs that you see replay of recent-experience-related activity patterns. (Data collected by Andres Grosmark)

The neocortex alternates between UP and DOWN states (UP: activity. DOWN: silence). DOWN states correspond to slow waves in the EEG: the main signature of NREM. (Data collected by @brendon_watson)

As a theorist, my approach is to start by making the most idealized (read: simple 😉) model that gives some intuition of how a system might work. So our first task was to make a toy model that does UP/DOWN alternations like the neocortex.

Fortunately, we didn't have to start from scratch. UP/DOWN is a kind of "default mode" of cortical tissue: also seen under anesthesia, when an animal isn't engaged, or even if you cut tissue out of the brain. But the alternation properties are different in different situations.

As a result, there’s a large body of work on UP/DOWN models. At the end of the day they all come down to the same basic mechanism: local excitatory connections keep the system in an UP state, slow negative feedback (adaptation) brings the system to a DOWN state.

These systems can show 4 different "regimes" (or phases...) of alternations: 1) regular oscillations, 2) noise-induced switches between UP and DOWN, 3,4) two excitable regimes where the UP or DOWN state is stable but fluctuations can evoke a brief excursion to the opposite state.

We were able to show how the interplay between the strength of adaptation, recurrence, and the level of drive determines which of these regimes a population is in: a phase diagram for an adapting recurrent neuronal population.

So where in our phase diagram do we find the UP/DOWN alternations from the NREM sleep data? The key was to use the distributions of UP and DOWN state durations.

When we matched the data to the model, we found that the neocortical dynamics reflect an excitable UP state, from which internal noise or a kick from another brain region can induce a brief stereotyped DOWN state - a slow wave!

We then looked at the durations of hippocampal SWRs and inter-SWR periods... And found that the SWR dynamics reflect the opposite excitable regime: an excitable DOWN state that can get kicked into a brief stereotyped UP state (a SWR).

So what does this have to do with sleep functions? In the excitable regimes each region can generate its stereotyped event spontaneously (due to noisy internal fluctuations) or in response to the complementary event in the other structure.

So the hippocampus can spontaneously produce SWRs, the content of which will be determined by the local synaptic network (replay).

And the neocortex can respond with a slow wave. This coupling has been shown to be important for memory consolidation.

But some of the time, the cortex spontaneously produces slow waves... We think that these "unperturbed" slow waves maintain the statistics of local synaptic weights.
sciencedirect.com/science/articl…

And finally, a cortical slow wave can induce a SWR... a lot less is known about this, but it's observed in experimental data and maybe it allows for two-way dialogue between HPC and CTX during sleep.

TL;DR we think excitable dynamics allow the hippocampal-neocortical system to produce and coordinate the neural population events that support NREM function.

Fin. Super happy to have this out there (it's been a long time coming…:)