Restricted Boltzmann Machine fantasizes MNIST digits

This three-layer, all-binary RBM was trained for 20 minutes using CUDA-accelerated Persistent Divergence training.

Category:

Science & Technology

Tags:

• RBM
• machine learning
• CUDA
• MNIST
• vision

[latanius]

cool! How many neurons do you have in the layers? and... how much faster was it with CUDA than without? (if you have an estimate :))

[deeplearning]

@latanius the limiting factor in densely connected RBMs is the multiplication of activations in one layer with the weights. Matrix multiplication is quite optimized on CPU as well, so you get an overall speedup of about 20-25. Dense RBMs do not scale to real images, however.

[davepamn]

Todd Jochem, a student of Pomerleau, built the next generation of code called RALPH. RALPH used 32x32 pixel low resolution picture of the road. The land ahead appeared as a wedge in the distance. If the road angles left or right, it estimated the blur in brightness changes, one cell from the next and the sharpest vector was kept. Learning was instantaneous. Could a RBM be used to drive a car?

[bonsalty]

Nice