The cerebral cortex contains hundreds of millions of neurons. The activity of one neuron is lost in this morass, so it is thought that the co-ordinated activity of groups of neurons – “neural ensembles” – are the basic element of cortical computation, underpinning sensation, cognition, and action. Understanding such neural ensembles is then fundamental to our understanding of the basic science of learning and decision-making, and to our understanding of the factors on which the normal operation of cortex depends.
However, there are competing, high-profile theories that differ in their accounts of how cortical ensembles form during learning, how they encode information, and how they are used during behaviour. Each thus makes differing predictions about how they would appear over learning or appear during subsequent use in performance. We set out to test these hypotheses using experimental data.
A recent hypothesis for population coding is that the activity of the population at some time t represents a sample from an underlying probability distribution (Fiser et al, 2010, TICs; Berkes et al 2011, Science). In this hypothesis, the basic computational operation of cortical networks is inference-by-sampling. The underlying probability distribution can be reconstructed by integrating over samples.
One key prediction of this hypothesis is that the distribution will change over experience due to inferential updating (via Bayes rule): namely, the distribution during “spontaneous” activity and during evoked activity differ before experience, and converge over repeated experience. Experimental evidence for this hypothesis has come from Berkes et al (2011, Science), where they showed just this convergence in small populations from V1 over development.
Recent models by Maass and colleagues have shown how cortical circuits can indeed produce these samples, and compute with them (Buessing et al 2011; Habenschuss et al, 2013, PLoS Comp Biol).
Missing is any empirical evidence that such stochastic computation is a general computational principle for cortex: whether it can be observed during learning, or in higher-order cortices, or during ongoing behaviour. We are currently testing this hypothesis using population recordings from the prefrontal cortex of rats learning a maze task.
Hebb's "cell assembly" hypothesis has been highly influential in modern neuroscience. The basic hypothesis is that neurons which fire together will become more strongly coupled. As that firing will be elicited by some common factor (stimulus, action, etc), so the formation of a group of coupled neurons will represent the learning of a new thing. This group is a "cell assembly".
Despite its influence, definitive tests for the existence of cell assemblies have been lacking. Numerous studies have reported neural ensembles, groups of correlated neurons (e.g. Laubach et al 2000 Nature; Harris et al 2003, Nature; Fujisawa et al 2008 Nat. Neurosci; Nicolelis & Lebedev 2009, Nat Rev Neurosci; Peyrache et al 2009 Nat. Neurosci; Benchenane et al, 2010 Neuron; Humphries, 2011). A few studies have looked at the changes in pairwise correlation over learning as a proxy for the formation of cell assemblies (Baeg et al 2007 J Neurosci; Mohdi et al 2014 Elife).
To our knowledge, no studies have explicitly tested whether or not ensembles of neurons form during learning. It follows that also unknown is what the cell assemblies are learning - what do they encode?