A two-layer spiking neural network consisting of discretely tuned, disjoint inputs and a layer of recurrently connected neurons containing assemblies that reinforce the input tuning.
Layer 1 network activity has two-dimensional shared variance by design, where the two dimensions correspond to 2 different tuned inputs. Recurrent layer network activity has high (>2) dimensional shared variance when recurrent clustering is strong.
All codes are for the manuscript: Rager, D. M., Khanna, S., Smith, M., & Doiron, B. "Assembly structure expands the dimension of shared variability in cortical networks" (2020, pending) More information about this research can be found at cortical.network
Julia (v1.5) codes simulate the spiking neural network. Matlab (R2019b) codes analyze the network activity simulated by the Julia codes.
runsim.jl:
- initalizes multiple Julia processes (one per physical computer core)
- initializes network parameters
simParams(simulation length in ms),sysSize(number of input, recurrent layer E, and recurrent layer I neurons),connProbs(network connection probabilities for each celltype pair),taus(membrane and synaptic time constants in ms),v4OU(strength and timescale of OU process that correlates input layer spiking activity) - calls all network simulation functions.
Call
times,ns,times0,ns0,weights,bias,connStrength = simTwoPopHemiInputUnpack_NoCoupleInit(simParams,sysSize,connProbs,taus,v4OU)to simulate 1 trial (default 12 seconds) of activity from a network with tuned, disjoint inputs and no recurrent connections.
times,ns,times0,ns0,weights,voltageOverTime,bias,connStrength = simTwoPopHemiInputUnpack_WeakCoupleInit(simParams,sysSize,connProbs,taus,v4OU)simulates 1 trial of activity from a network with tuned, disjoint inputs and weak recurrent connections.
times,ns,times0,ns0,weights,synInputPerNeuronOverTime,bias,connStrength = simTwoPopHemiInputUnpack_StrongRecSymmClusters(simParams,sysSize,connProbs,taus,v4OU)simulates 1 trial of activity from a network with tuned, disjoint inputs and strong (uniform or clustered) recurrent connections. Within the simTwoPopHemiInputUnpack_StrongRecSymmClusters method, you can change R, the clustering strength of the recurrent architecture. R = 1.25 corresponds to a network in which recurrent layer neurons receiving the same tuned input connect with 1.25x the strength as recurrent layer neurons receiving opposite tuned inputs. R is the 2nd argument in the ConnStrength constructor. Upcoming versions of the code base will make the R parameter accessible as an argument to simTwoPopHemiInputUnpack_StrongRecSymmClusters.
runSimsFrozenGraph(numSims,simParams,sysSize,connProbs,connStrength,taus,v4OU,bias,weights)simulates numSims trials of activity for a frozen network architecture that was initialized with one of the above functions.
runSimsFrozenGraphLinResp(numSims,simParams,sysSize,connProbs,connStrength,taus,v4OU,bias,weights)mirrors the behavior of runSimsFrozenGraph(numSims,simParams,sysSize,connProbs,connStrength,taus,v4OU,bias,weights) plus runs an additional 20 trials of activity for a network with the same, frozen input structure and no recurrent connectivity. This is useful for the paper's linear response calculations.
- Julia (v 1.2+)
- Hwloc
- PyPlot
- JLD2
- HDF5
- CSV
- Parameters
- BenchmarkTools
- MATLAB (R2019b)
- Statistics and Machine Learning Toolbox
- Shape Language Modeling
For bugs, comments, concerns: use the Github issue tracker.
Author: Danielle Rager, danielle [at] cortical.network