25 — Efficient codes and balanced networks

This paper begins by exploring the ways in which neuroscientists have understood neuron “spike trains” in the past. Conventional neuron signals can only convey information in the action potential timing; the amplitude does not vary between action-potentials (“spikes”) because the underlying biochemical process is the same for each spike.

In the past, electrophysiologists have modeled this information transfer as a Poisson rate code: In principle, take some time duration, count the number of spikes, and then extrapolate error-bars, assuming that this distribution of spikes adheres roughly to a normal distribution.

This paper opens by challenging this paradigm, which — from my data-privileged perspective from a computer science background, seems almost laughably intuitive. (If you’re receiving time-series signals and the amplitude is the same in each, then turn the radio knob to the “FM” setting instead!) …But. I appreciate that this understanding is based on my years of learning as a neuroscientist in the 21st century, and so what is intuitive to me might be a result of it becoming common-knowledge before my time.

Fundamentally, this means that neurons are encoding something significant in both the spikes and lack of spikes: Almost like a barcode, the whitespace is as important as the visible signal.

The authors suggest that this “balanced signal” comes from the strong role that inhibitory signals play in the brain. To test this, one simulates a basic “point model” neuron using the “integrate and fire” function (sum all excitatory inputs $e$ and inhibitory inputs $i$, and when it surpasses threshold $t$, that is, $(\sum_j^n{e_j} - \sum_k^n{i_k}) \geq t$), send one spike to all downstream neighbors).

The authors cite two main paradigms to enable this sort of E/I pairing: The first is a loosely-coupled network, where inhibitory and excitatory signals must self-assemble to “average out” to a biologically sustainable equilibrium. This is biologically cheaper, but the signals are, necessarily, only loosely related. The other paradigm is a tightly-balanced network, in which neurons are densely connected and this interconnectivity fuels the synchronicity of E/I signals. In this tightly balanced system, much more energy is expended to maintain a dense synaptic fabric of connectivity, but the inhibitory and excitatory signals can more quickly influence each other, leading to a more finely tuned network. Naturally, it is unclear which model the brain follows, and the authors propose that different systems may exist at different places on a continuum.

The review closes by pitching the idea that by better understanding the nature of E/I signal correlation, we will be able to closer model the inner workings of biological neural nets. The authors pose ideas like multi-channel recording or calcium imaging to accomplish this task, and I wonder if even high-fidelity, strictly-anatomical data (such as EM) might be able to lend insight into the ways in which different neuron types intercommunicate.