MvdM: a powerful technique in neuroscience is to record neural activity in behaving animals. Current techniques allow for the simultaneous recording of up to ~150 isolated neurons, potentially from multiple brain areas, along with field potentials. (Wilson & McNaughton?, Science 1993; Buzsaki, Nat Neurosci 2004)

Video example of an ensemble recording from the hippocampus, attached below.

This results in rich data sets that benefit from close interplay with quantitative and theoretical approaches. I highlight two ways in which this is apparent in my own experience as an experimentalist:

1) models in a very general sense (theories, detailed network implementations, etc.) are essential to highlight features of the data that may be important or interesting.

Example: a particular spike timing pattern known as theta phase precession (O'Keefe & Recce, Hippocampus 1993; Maurer & McNaughton?, TINS 2007). The feasibility of a functional role for this in rapid learning was demonstrated in simulations, and other theoretical work (Jensen & Lisman 2000; Lisman & Redish 2009) has suggested intriguing alternative or complementary functions that have inspired much experimental work.

Example 2: spike counts from individual passes through a place field tend to have higher variance than would be expected from a Poisson process (Fenton & Muller 1998). The significance of this was initially unclear until models demonstrated that this could arise from switching between hippocampal states or "maps"; a prediction subsequently confirmed experimentally (e.g. Jackson & Redish, Hippocampus 2007, Fenton et al. J Neurosci 2010).

2) models in a narrower, statistical sense are key to the interpretation of the data.

Example: ensemble decoding. We would like to estimate p(location | spikes) because this allows us access to fine-timescale representations of future or past spatial trajectories -- of interest given the hippocampus's known role in memory and navigation.

This requires us to know p(spikes | location) which requires a statistical model. In the context of the hippocampus, pioneering statistical work includes Brown et al. J Neurosci 1998; Zhang et al. J Neurophys 1998. For a review, see Brown et al., Nat Neurosci 2004.

For application to experimental questions in the hippocampus, examples are Johnson & Redish, J Neurosci 2007; Davidson et al. Neuron 2010; van der Meer et al. Neuron 2010.


ESB: Ensemble coding raises a suite of beautiful questions that arise across neuroscience. The fundamental object is p(spikes | stimulus), where the stimulus is the key variable that the ensemble is assumed (by the brain, or the experimenter, or in the best of words both) to represent. In the data above, stimulus = location.

Q1: Is simultaneous data from multiple neurons needed to characterize p(spikes | stimulus), or could we assemble this from sequenced observations from individual neurons?

NOTE -- below is a undersampled, incomplete ref list, as a starting point!

Key review paper with many references: Averbeck et al Nature Nsci. 06. See in particular Zohary et al; Nature 94, Gawne and Richmond J Nsci 93, Montani et al Proc Royal Soc 09.

Q2: If that "conditionally independent" approach doesn't work, is the whole p(spikes | stimulus) then the "sum of its pairs?" (cf. Schneidman et al Nature '06 and earlier works cited therein)

Experiments -- see key references (Schneidman, Shlens, Montani, Grun, Martignon, Amari, others): Reviewed in Beggs, "Maximum Entropy Approaches to Living Neural Networks," Entropy 2010

Theory -- why and when do we need to go beyond pairwise correlations? Macke et al Phys Rev Let 2011, Barreiro et al ( http://arxiv.org/abs/1011.2797) 2010, Roudi et al PLOS Comp Bio 09

Q3: Beyond just DESCRIBING p(spikes | stimulus) accurately ... do we need multiple-neuron description for neural CODING?

Again, see key references toward decoding and encoding perspective in Averbeck review paper above. Also, see Montani et al above and Still, Schneidman et al Phys Rev Lett 2003 for contributions of higher-order correlations to coding.

Q4: Beyond CODING, what's the role for multi-neuron activity in circuit dynamics?

See role of correlations in signal transmission: e.g. Salinas and Sejnowski, J Nsci 2000.