RBM demo regenerating images

Regenerating handwritten digit images using neural network weights created by a Restricted Boltzmann Machine. The images are being compressed(encoded) into a 30 digit number, (the small top layer), then an attempt is made to reconstruct an image from this 30 digit number.

More Details:

This is a GUI for playing with some neural network weights trained by a Restricted Boltzmann Machine.

The training was done by a C program based on matlab code from Ruslan Salakhutdinov and Geoff Hinton,
available here: http://www.cs.toronto.edu/~hinton/Mat...

Specifically these weights represent an autoencoder (no labels for the input data), equivalent to the matlab program "mnistdeepauto.m" from the link provided above.

The inputs are 60,000 MNIST handwritten digit images (28x28 pixel, greyscale) from: http://yann.lecun.com/exdb/mnist/trai...

The network layer sizes are 1000, 500, 250, 30.

Training consisted of 50 epochs (aka iterations) of pretraining
with the images split into 600 batches of 100 images.
Followed by 200 epochs (aka iterations) of backpropigation
with 60 batches of 1000 images.

Category:

Science & Technology

Tags:

• Restricted
• Boltzmann
• Machine
• rbm
• ocr
• ai
• autoencoder
• MNIST
• handwritten

[davepamn]

Has the RBM been applied too navigating a trail or road?  Can the RBM make enough correct decisions combined with an expert system too travel from one coast to the other? Could the RBM separate what was a road from a nonroad material?

[dhillonv10]

@davepamn deep belief nets as shown in this case are unsupervised, then later on a supervised learning algorithm is used to fine tune the weights on the connections :)

[davepamn]

Are the glyphs being produced from the 100 dimensional Convolutional Network? Is the supervised learning or unsupervised?

[Kotesu]

Sweet! Do you happen to have the C code for this?