Updated nnlib2Rcpp package

Interested in publishing a one-time post on R-bloggers.com? Press here to learn how.

---

For anyone interested, ‘nnlib2Rcpp’ package has been updated with several added features. Among other changes, the latest v.0.2.1 of the nnlib2Rcpp package allows users to define the behavior of custom Neural Network (NN) components they create using only R code. This includes custom layers and sets of connections.

Package nnlib2Rcpp is based on the nnlib2 C++ NN library. It interfaces compiled C++ NN components with R. New types of NN components can be created using the provided C++ classes and templates. However, as of v.0.2.1, user-defined NN components can also be created without any need for C++. Of course, NN components defined using nnlib2 C++ classes and templates (as described in an older post), or components already included in the package can still be used. All such components can be added to neural networks defined in R via the nnlib2Rcpp package’s “NN” module and cooperate with each other.

Defining custom NN component behavior in R does have a cost in terms of runtime performance and, to a certain degree, defies many of the reasons for using the provided C++ classes. However, it may be useful for some purposes.

The goal of the simple example listed below is to implement, using only R, a NN with functionality similar to that described in the aforementioned post and required some steps done in C++. In the example, component (connection set and output layer) functions required for a -simplified- perceptron-like NN are defined and the NN is set up. This is essentially a single layer perceptron as the first (“generic” layer just accepts the data and transfers it to the connections without performing any computations.

library(nnlib2Rcpp)

# Function for connections, when recalling/mapping:

CSmap <- function(WEIGHTS, SOURCE_OUTPUT,...)
	SOURCE_OUTPUT %*% WEIGHTS

# Function for connections, when encoding data:

learning_rate <- 0.3

CSenc <- function( WEIGHTS, SOURCE_OUTPUT,
				   DESTINATION_MISC, DESTINATION_OUTPUT, ...)
{
  # desired output should have been placed in misc registers:
  a <- learning_rate *
          (DESTINATION_MISC - DESTINATION_OUTPUT)
  # compute connection weight adjustments:
  a <- outer( SOURCE_OUTPUT, a , "*" )
  # compute adjusted weights:
  w <- WEIGHTS + a
  # return new (adjusted) weights:
  return(list(WEIGHTS=w))
}

# Function for layer, when recalling/mapping:
# (no encode function is needed for the layer in this example)

LAmap <- function(INPUT_Q,...)
{
	x <- colSums(INPUT_Q)		# input function is summation.
	x 0,1,0)		# threshold function is step.
	return(x)
}

# prepare some data based on iris data set:

data_in <- as.matrix(iris[1:4])
iris_cases <- nrow((data_in))

# make a "one-hot" encoding matrix for iris species
desired_data_out <- matrix(data=0, nrow=iris_cases, ncol=3)
desired_data_out[cbind(1:iris_cases,unclass(iris[,5]))]=1

# create the NN and define its components:
# (first generic layer simply accepts input and transfers it to the connections)

p <- new("NN")
p$add_layer("generic",4)
p$add_connection_set(list(name="R-connections",
                          encode_FUN="CSenc",
                          recall_FUN="CSmap"))
p$add_layer(list(name="R-layer",
                 size=3,
                 encode_FUN="",
                 recall_FUN="LAmap"))
p$create_connections_in_sets(0,0)

# encode data and desired output (for 50 training epochs):

for(i in 1:50)
	for(c in 1:iris_cases)
	{
		p$input_at(1,data_in[c,])
		p$set_misc_values_at(3,desired_data_out[c,])  # put desired output in misc registers
		p$recall_all_fwd();
		p$encode_at(2)
	}

# Recall the data and show NN's output:

for(c in 1:iris_cases)
{
	p$input_at(1,data_in[c,])
	p$recall_all_fwd()
	cat("iris case ",c,", desired = ", desired_data_out[c,],
		" returned = ", p$get_output_from(3),"\n")
}

More information can be found in the package’s documentation by typing:

help(NN_R_components)

A complete list of other changes done and features added to the package can be found here.