Fig 1.
Network architecture and modeling of the study by Xu et al. [1] in the LIF-SORN.
(A) Distribution of neurons on the 2D grid of one network instance. Blue lines show all connections projecting from the excitatory population to an example excitatory neuron after 500 s of simulation time. Connections spanning a larger distance are unlikely to exist, due to the distance dependent connection probability. (B) Cross-section of the rate rspot of the Poissonian input spike trains with Δx being the distance to the center of the spot. (C) Distribution of the excitatory neurons including the start point , mid point
and end point
of the trajectory of the moving spot used in most experiments. The colored circles define the neurons that are pooled together for the analysis of the sequence learning ability.
Fig 2.
Characterization of spiking activity (based on activity in the time window 400s–500s).
(A) Spike trains of excitatory neurons. (C) Firing rate distribution of excitatory neurons. (B,D) Same as (A,C) for inhibitory neurons. (E) Distribution of pairwise correlation coefficients of the excitatory population, computed for time bins of 20 ms duration. (F) Interspike interval distribution of all excitatory neurons (Exponential function is fitted to the data points with tISI > 50 ms). (G) Distribution of coefficients of variation of the interspike interval distribution of each excitatory neuron. Data from a single network instance.
Fig 3.
Characterization of network connectivity.
(A) Connection fraction of recurrent excitatory connections. (B) Time course of the weights of 10 randomly selected recurrent excitatory connections after the connection fraction has stabilized. (C) Distribution of the weights of existing recurrent excitatory connections, determined after 500 s of simulation time. Data from a single network instance.
Fig 4.
Cue-triggered sequence replay.
(A) Example spike trains of neurons part of clusters in response to a brief flash of the light spot at before and after training and in response to the full motion sequence during training. Top and bottom row show spike trains for two different network instances. Neurons are ordered according to the projection of their location on the
axis. (B) Normalized pairwise cross-correlation between spikes of neurons part of clusters in response to a brief flash of the light spot at
before (top) and after (middle) training and normalized difference between the cross-correlograms after and before training (bottom). Panel B shows results from 20 instances of the LIF-SORN.
Fig 5.
Analysis of cue-triggered sequence replay.
(A) Top: Cumulative distribution of Spearman correlation coefficients when testing with a cue presented at . The distribution is shifted to the right after (solid) compared to before (dotted) training. Bottom: Same as top but for results in rats [1]. (B,C) Same as A but for
-evoked (B) and
-evoked (C) responses. Top plots of all panels show results from 20 instances of the LIF-SORN. Bottom plots show approximate experimental data that were obtained from [1] using WebPlotDigitizer [30].
Fig 6.
Effect of training with a moving spot along the axis on connectivity.
(A,B,C) Connection weights between neurons part of the clusters before (A) and after (B) training and their difference (C) for a single network instance. Neurons are ordered according to the projection of their position on the axis. (D) Change of the mean weight of all possible connections between neurons part of the clusters. Averaged over 20 network instances.
Fig 7.
Effect of training with a moving spot along the axis on spontaneous activity.
(A,B,C) Transition probabilities between clusters before training (A), after training (B) and their difference (C). Colored frames indicate transitions in the forward (green) and backward (purple) direction. (D) Mean of the changes of transition probabilities in the forward (green) and backward (purple) direction corresponding to the framed transitions in (C). Errorbars show sem over network instances. Data from 30 network instances. Stars in panel D show significance (⋆⋆⋆ P < 0.001; Wilcoxon signed rank test).
Fig 8.
Mean recall speed as a function of the training speed.
Mean recall speed of test trials with cue presentation at and a Spearman correlation coefficient greater than 0.9. Data from 20 network instances. Errorbars show sem.
Fig 9.
Persistence of training effect and time course of connection weights during and after training.
(A) Change in percentage of matches as a function of time after training in the LIF-SORN and in rats [1]. (B) Time course of the connection weights between adjacent clusters in the forward (green) and backward (purple) direction before, during and after training. Shaded area marks the training phase. Data from 30 network instances. Dotted line in panel A shows approximate experimental data that were obtained from [1] using WebPlotDigitizer [30]. Errorbars show sem in all plots. Stars in panels A show significance (⋆ P < 0.05; ⋆⋆ P < 0.01; ⋆⋆⋆ P < 0.001; Wilcoxon signed rank test).