Library of Dynamic Neural Networks for time series related tasks Main library consist of multithreaded simulator of recurrent spiking (impulse) neural networks dynamics written in C++. It contains variety components which can be connected with each other in different combinations

Basic integrate and fire neuron model. Has specification in constants:

{
    "tau_ref" : double, ms, //refractory period
    "rest_pot" : double, mV, // resting potential
    "tau_m" : double, ms, // time constant of leaky integrator
    "noise" : arbitrary units, // factor of white noise on mebrane
}

####SRMNeuron Spike-Response Model

####SpikeSequenceNeuron Mock neuron for input spike sequences from files ##Synapses ####StaticSynapse static synapse, can be described by simple exponential decay ####STDSynapse synapse with short-term memory (Tsodyks et al)

####Determ Determinate threshold, neuron is firing if membrane reached threshold ####ExpThreshold Exponential version of activation function, it has specific increase near threshold value (Hennequin et al) ##Learning rules: ####OptimalStdp Supposed to maximize pre-post information (Toyoizumi et al) ####Stdp Simple spike-timing dependent plasticity ####TripleStdp (under development) - Complicated version of Stdp with long range dynamics (Pfister et al) ##Reinforcements: ####InputClassifier Will modulate positive reward when layer-wise local id of neuron matches current class id, and negative in another case. Amounts of positive and negative reward can be pointed through constants, ltp and ltd repectively in arbitrary units:

To install dnn on UNIX like machine you need to satisfy dependencies:

• clang >=3.5 or gcc >=4.9.2
• cmake >=2.8
• libprotobuf-dev
• protobuf-compiler
• r-base (for plots and analytics)

To use R API and R scripts you need to install these R packages:

• Rcpp
• zoo
• lattice
• kernlab
• rjson

After satisfying of all prerequisites main installation must be done through installation script:

$ git clone https://github.com/alexeyche/dnn
$ cd dnn
$ ./install.sh

Script will ask where to install dnn, by default it will use /dnn, but you can change it. Script will build dnn and R package (Rdnn), then trying set through **/.profile** file environment variables:

export DNN_HOME=~/dnn # or wherever you'd chosen installation path
export LD_LIBRARY_PATH=$DNN_HOME/lib:$LD_LIBRARY_PATH

Last part can be different for different systems.

Lets generate some testing time series:

require(Rdnn)
proto.write(time.series(values=rnorm(1000)), ts.path("test_ts.pb"))

last command will create in $DNN_HOME/ts file named test_ts.pb

To run simulation you need to set up constants for neccessary neuron specification, e.g. LeakyIntegrateAndFire in $DNN_HOME/const.json file. After that you can set up simulation configuration in the same file for that neuron. You need to add spec into sim_configuration->layers":

{
    "size" : 1000,
    "neuron" : "LeakyIntegrateAndFire",
    "act_function" : "Determ",
    "input" : "InputTimeSeries"
}

size stands for size of the layer, all other entities can be found in specification above:

• neuron for neuron of that layer
• act_function function of activation of neuron, for Determ threshold can be set in Determ key above
• input is about providing membrane of neuron with some current specified in input time series. Entity with the name InputTimeSeries responsible for this.

In the field sim_configuration->files you may find information related to InputTimeSeries with macro @macro-name. Generaly speaking it awaits path to filename, but this macro can be replaced by file while running simulation with option --macro-name

Simulation can be run through python script:

cd $DNN_HOME
./scripts/run_sim.py --ts-input ./ts/test_ts.pb

Option --ts-input related to macro in user_const.json which allows to point this filename directly into InputTimeSeries object. This run will create in $DNN_HOME/runs/sim directory where you can find artefacts of running and png pictures of some introspection of simulation (raster plot)

Layer specification consists of:

{
    "size" : 100, // required
    "neuron" : "NameFromNeurons",  // required
    "act_function" : "NameFromActFunctions", // required for all neurons with activation
    "input" : "NameFromInputs", : // non-required
    "learning_rule" : "NameFromLearningRules", // non-required
    "weight_normalization" : "NameFromWeightNormalizations", // non-required
    "reinforcement" : "NameFromReinforcements", // non-required
    "axon_delay" : double, ms units, // non-required, 0 by default
}

All names can be found above from corresponding sections

{
    "synapse" : "NameFromSynapsesAbove", // required
    "type" : "NameFromConnectionsAbove", // required
    "start_weight" : double, arbitrary units, // required
    "dendrite_delay" : double, ms, // non-required, 0 by default
    "inh_synapse" : "NameFromSynapsesAbove", // required only by special type of connections, e.g. DifferenceOfGaussians
}

All names can be found above from corresponding sections, type can be found in connections section.