This is the code for the paper "Spontaneous emergence of fast attractor dynamics in a model of developing primary visual cortex", by T. Miconi, J. McKinstry and G. M. Edelman (published in Nature Communications, vol. 7, n. 13208, 2016).
This code simulates a small patch of cortex, containing 100 principal neurons and 20 inhibitory neurons. Excitatory neurons have voltage-based spike-timing dependent plasticity (Clopath-Gerstner algorithm) at both feedforward and recurrent synapses. Connections to and from inhibitory neurons are fixed and random. The patch is repeatedly exposed to natural images, preprocessed in a way that emulates retinal and thelamic filtering. After many iterations, the network develops recurrent connectivity that supports attractor dynamics (groups of mutually-connected, similarly-tuned neurons, as observed in cortex), as well as realistic feedforward receptive fields.
The entire code is contained in one program: stdp.cpp. It requires the Eigen v3 algebra library, which is very easy to install since it's source-only. After dowloading and unzipping the Eigen library into the $EIGEN_DIR directory, you can compile stdp.cpp as follows (be sure to adapt to your own compiler options and directory structure):
g++ -I $EIGEN_DIR/ -O3 -std=c++11 stdp.cpp -o stdp
The present code base includes pre-trained weights (w.* and wff.*). If you want to run the full training yourself, just run ./stdp learn to train a network from scratch. The code is configured to run 500K presentations, which could take about a day. However, good results are evident by ~300K evaluations.
If you want to evaluate the trained network and generate the figures from the paper, you will need to run ./stdp test, and also ./stdp mix and ./stdp pulse. Consult makefigures.m for more details.
For convenience, we include a Jupyter notebook (V1Clusters.ipynb) that downloads, compiles and runs the code (it also installs the Eigen library). By default it uses the included pre-trained weights, and generates the main figure from the paper. By uncommenting the relevant cell, you can run the training yourself. This notebook should run fine on Google Colab.
We also include additional Matlab programs that generate the image stimuli (makepatches.m - note that pre-generated image patches are included in the present code base) and produce figures (showw.m, analyzewlat.m and makefigures.m)
- showw.m produces a picture of the receptive fields (subtracting feedforward weights from the ON and OFF retinal cells).
- analyzewlat.m produces various graphs and preovides information about cluster formation
- makefigures.m makes the figures in the paper. You need to read it and manually activate the code for each figure, after having produced the necessary data as instructed in the code for this figure. Note that figures will look slightly different from those in the paper, but qualitatively similar.