Fig 1.
Neural correlations contribute to information encoding of behaviour or stimulus.
(a) We reanalysed data from mouse posterior parietal cortex (PPC) during maze navigation [1] and mouse primary visual cortex (V1) during presentation of artificial (PDG) and naturalistic (MOV) stimuli [4]. (b) Histograms of tuning stability (Eq 4) in the two datasets. Note that the stability is computed on different time scales between the two datasets. (c) Schematic of potential sources of redundant and synergistic neuron pairs: functional connections between the pair, and tuning to a second unobserved input variable V. (d) Example PPC neuron pair with high redundancy from mouse 4 on session 5, from [1]. ΔF/F data points are coloured according to the position in the maze. (e) Contours of Gaussian distributions fit to the data in (d). We binned the position into 7 bins and fit covariance matrices to bins 1 and 4. The contours in the top plot show these fits, and those in the bottom plot show covariance matrices with the covariance set to 0, corresponding to the neurons acting independent of each other. (f) Cumulative distribution of the redundant information over all neuron pairs for each behavior or stimulus. (g) Same as (f) for redundancy index over all neurons. (h)–(k) Same as (d)–(g) for synergy.
Fig 2.
Redundancy index correlates with tuning stability.
(a) Constructing a multivariate linear model for stability with three regressors: mutual information (MI), redundancy index (Red), and synergy index (Syn). Pair plots showing the stability plotted against each of the predictors for one mouse on a single session (Driscoll et al. [1], 130 neurons, day 5, right turns). (b) The three regressors are all strongly correlated with each other. Each data point represents a neuron from one mouse on a given day (Driscoll et al. [1], 130 neurons, 17 days). The redundancy index is plotted against the synergy index, with the size and colour showing the mutual information and stability, respectively. The session-averaged correlation coefficient ρ between each pair of regressors is given in the bottom right. (c) Regularised regression coefficients for the three predictors with position and heading as the target external variable [1]. Each data point represents a given session and trial type (left or right turns) for a given mouse. (n = 617 PPC neurons across 4 mice.) (d) Same as (c) but with gratings and movies as the external variable [4]. (n = 1053 V1 neurons across 4 mice.)
Fig 3.
High-stability neurons form redundant cliques.
(a) Graph plots of pairwise redundant and synergistic information. Nodes represent neurons and are coloured and ordered according to position tuning stability. Edges represent the redundant or synergistic component of the pairwise information, thresholded for the purposes of visualisation. Inset graphs show the top 40 most stable neurons in the population. (b) Clique detection in thresholded graphs R and S retaining the top 0.35% of edges. Redundant and synergistic cliques CR and CS are highlighted in red and green, respectively. (c) Mean stability of redundant and synergistic cliques CR and CS. Each data point is an average over 10–50 size-matched graph pairs for a single session and mouse (see Methods). (d) Nearest neighbors and
of redundant and synergistic cliques CR and CS in their opposing graphs. Nearest neighbors are highlighted in yellow. (e) Mean stability of opposing nearest neighbors
and
. Each data point is an average over 10–50 size-matched graph pairs for a single session and mouse (see Methods). Illustrative graph plots in (a), (b), and (d) are from a single session for one mouse (Driscoll et al. [1], 130 neurons, day 4, left turns).