This is an R implementation of Adam Coates' paper on unsupervised learning [1]. The name Aha is derived from Adam-Honglak-Andrew, the initial of the three authors, signifying sort of an epiphany that a single layer network could learn as well as a deep neural net, achieving state-of-the-art result on certain datasets.
#Installation To install directly from github, open a terminal, type R, then
devtools::install_github('htso/AHA')
#Dependencies You need the following packages. To install from a terminal, type
install.packages("Rcpp", "LiblineaR", "pixmap", "R.matlab", "R.utils")
#Datasets In their paper, the algorithm is tested on three datasets : Cifar10, NORB, and STL. I provide the functions to download these datasets directly from the web. But beware that it may take a long time to download certain files, e.g. CIFAR10.
I include a toy dataset, which is small subset of CIFAR10. To load it, just type
data(tinyCifar10)
#Run I provide a demo in the /demo subfolder. To run it,
setwd(system.file(package="AHA"))
demo(runscript)
#WARNING It may take a long time to finish a run, even on the toy dataset.
#Platforms Tested it on Linux (ubuntu 14.04) and Windows 7, should work on OS X.
[1] Coates, Adam, Andrew Y. Ng, and Honglak Lee. "An analysis of single-layer networks in unsupervised feature learning." International conference on artificial intelligence and statistics. 2011.