bainro / MSTd_AE_EA

MSTd SNN model trained using autoencoder like loss function & an evolutionary search of parameter space.

Sparse, reduced representations in model of MSTd

This is the code accompanying the following publication:

Chen, K., Beyeler, M., Krichmar, J. L. (2022) Cortical Motion Perception Emerges from Dimensionality Reduction with Evolved Spike-Timing Dependent Plasticity Rules. The Journal of Neuroscience (in Press).

Requiremments

• CARLsim5: https://github.com/UCI-CARL/CARLsim5 (this branch)
• ECJ: https://github.com/GMUEClab/ecj
• MATLAB 2017a

Generate input stimuli

Stimuli in the training, validation, and test datasets can be generated with the MATLAB script matlab_analysis_scripts/GenerateInputStim/generateInputStim.m.

Network Simulation

evolve_MST_SNN_model_CARLsim

• Project code to evolve network parameters.
• Compile with make, and use ./launchCARLsimECJ.sh to launch evolutionary runs. Network fitness values and evolved hyper-parameters are saved to out.stat.

test_MST_SNN_model_CARLsim

• Project code that includes additional test trials.
• Pass in evolved parameters to simulate individual networks. Network activity and trial indices are saved in the results folder.

Analysis scripts

Folder matlab_analysis_scripts contains MATLAB scripts for data analysis and visualization. These scripts were partially based on code developed for Beyeler et al. (2016).

References

Beyeler, M., Dutt, N., and Krichmar, J.L. (2016). 3D Visual Response Properties of MSTd Emerge from an Efficient, Sparse Population Code. The Journal of Neuroscience 36, 8399-8415.