Code for the paper "Coordinated drift of receptive fields in Hebbian/anti-Hebbian network models during noisy representation learning" S Qin, S Farashahi, D Lipshutz, AM Sengupta, DB Chklovskii, C Pehlevan Nature Neuroscience 26 (2), 339-349. See also the priprint https://www.biorxiv.org/content/10.1101/2021.08.30.458264v1.abstract

Dependence

Some of the scripts depend on ITE (information theoretical estimators) which can be downloaded from https://bitbucket.org/szzoli/ite/src/master/

To plot figures without running the simulations which could take upto a few hours for some simulations, download the data from https://www.dropbox.com/sh/hgcxraa6kvv7jm3/AACVKWIomrHdyU46iMiQdQl4a?dl=0 and put the data folder in the main directory. You might need to modify the dictory of the plotting scripts.

Simulations

To run the simulation using the scripts in the folder simulation. The resulted data should be stored in the folder data

• drift_PSP.m Drift in Principal Subspace Projection (PSP) task, related to Figure 2
• PSP_D_Dependency.m Change the input statistics in PSP and study its effect on diffusion constant, related to Figure 2
• ringPlaceModel.m The nonlinear Hebbian/anti-Hebbian with "ring-shaped" data manifold, related to Figure 3 and Figure 4
• ring_model_three_phases.m Simulation of the "ring" model with different noise sources: noise only in forward matrix, noise in recurrent weight and noise in both weight matrices. Related to Figure 4F
• placeCell1D_slice.m 1D place cell model, input is draw from 1D grid fields which are slices through 2D grid fields. Related to Figure 5
• placeCell1D_ExciInhi.m 1 D place cell model with both excitatory and inhibitory neurons
• place_cell_learn_forget.m 1D place cell model with alternating learning and forgetting sessions.
• placeCell1D_slice_three_phases.m Different noise source in the 1D place cell model
• place1D_compare_model_experiment.m Comparison of experimental data of hippocampal CA1 place cells.
• placeCell1D_slice_multi_timescale.m Show that learned representation can be quite stable if there are both fast and slow timescale in the synaptic dynamics
• placeCells.m Simulation of the 2D place cell model, related to Figure 5
• Tmaze.m Simulation of representational drift of parietal cortex neurons during T-maze task, related to Figure 6
• comparePSP_PCA_SangerNoise.m Show that "degeneracy" of objective function is important for observing representational drift in linear networks, we compare PSP with PCA and Sanger's learning rule.

Plot the figures

To plot the figures, run the code in the folder plot. The script name contains the information about which figure it plots.

MATLAB version tested: 20220(b)