visualize.hidden.units: Visualize features learned by a sparse autoencoder

In autoencoder: Sparse Autoencoder for Automatic Learning of Representative Features from Unlabeled Data

Description

Visualizes features learned by a sparse autoencoder, by plotting (norm bounded) input images that maximally activate each of the hidden units of the trained autoencoder. Here it is assumed that the autoencoder is trained on a set of images of size Nx.patch by Ny.patch (in pixels).

Usage

visualize.hidden.units(object, Nx.patch, Ny.patch)

Arguments

object	an object of class autoencoder produced by the autoencode function, and containing information (architecture, weights, biases, unit type, etc.) about the autoencoder network.
Nx.patch	width (in pixels) of images in data set used for training the autoencoder. See 'Examples'.
Ny.patch	height (in pixels) of images in data set used for training the autoencoder. See 'Examples'.

Value

A figure in which each square shows the (norm bounded) input image of size Nx.patch by Ny.patch (in pixels) that maximally activates each of the hidden units. These squares represent the features learned by the autoencoder from the unlabeled data used for its training.

Author(s)

Yuriy Tyshetskiy

Examples

## Load a pre-trained autoencoder object with N.input=100 and N.hidden=10*10,
## trained on unlabeled set of 5000 image patches of size 
## Nx.patch=10 by Ny.patch=10 pixels, 
## randomly cropped from 10 nature photos, and visualize the features 
## learned by its hidden units:

data('autoencoder_Ninput=100_Nhidden=100_rho=1e-2')          

## Visualize hidden units' learned features:
visualize.hidden.units(autoencoder.object,Nx.patch=10,Ny.patch=10)

Example output

autoencoder documentation built on May 2, 2019, 5:52 a.m.