msd.md

Introduction

This is a demonstration of the mean spike distance task by Gers et al. (2002). The task is about learning a representation of time, which involves counting discrete time steps, reading out the counter and also learning to reset the timer when an external stimulus is present.

First, we load the library and some plotting libraries and set a random seed for all bnnlib internals.

library(bnnlib)

## Loading required package: ggplot2

## Loading required package: gridExtra

library(ggplot2)
library(gridExtra)
library(tictoc)
setRandomSeed(92357)

## NULL

We create the network from Gers02a with LSTM hidden layer (including peepholes) with 1 output and 1 input.

#network = NetworkFactory_create_gers02a()
network = LSTMNetwork(1,1,1, TANH_NODE, LINEAR_NODE)

Generate data

set.seed(4598234)
n <- 100
interval.lengths <- 1:5
#interval.lengths <- 1:10
sequence_set <- SequenceSet()

for (k in 1:10) {
intervals <- sample(interval.lengths, size=n, replace=TRUE)
input_seq <- c(1,rep(0, sum(intervals)))
target_seq <- rep(0, sum(intervals)+1)
input_seq[cumsum(intervals)+1] <- 1
target_seq[cumsum(intervals)+1] <- intervals
sequence <- toSequence(data.frame(input_seq, target_seq),1,2)
SequenceSet_add_copy_of_sequence(sequence_set, sequence)
}

Training

#trainer <- ADAMTrainer(network)
trainer <- ImprovedRPropTrainer(network)
#trainer <- ARPropTrainer(network)


tic()
Trainer_add_abort_criterion__SWIG_0(trainer, ConvergenceCriterion(1e-3),10)

## NULL

Trainer_train2(trainer, sequence_set, 500)

## NULL

toc()

## 1.645 sec elapsed

plotTrainingerror(trainer) + theme_light()

## `geom_smooth()` using method = 'loess' and formula 'y ~ x'

Predictions

Plot the network predictions zoomed in to the first 25 steps of the sequence.

plotPredictions(network, sequence) + theme_light()

## NULL

Investigate activations

Plot all activations for the entire sequence.

plotActivations(network, sequence) + theme_light() +   coord_cartesian(xlim=c(0,25))

plotActivations(network, sequence,"CEC") + theme_light() +   coord_cartesian(xlim=c(0,25))

plotActivations(network, sequence) + theme_light() +   coord_cartesian(xlim=c(0,25))