is this based on hinton's RBM? The performance is OK..but it does make some mistakes that a human would not... is that a limitation of the NN or was the training set small?
Yes, this is based on Hinton's RBM, stacking RBMs on top of each other in what he calls Deep Belief Nets. I think the performance is not so good because I didn't tune the parameters well enough. Also, in this example I just did local layer by layer training, without backprop refining.
Yes, it is. Given the image of a handwritten digit on the left the neural network tries to produce it's best guess on the right, based on prior experience from a few different handwritten digits it had seen before.
You should train using 30000 digits instead of 300. I think that's the bottleneck right now.
anukefromrussia 2 months ago
Would it be too selfish to ask you to share your code?
hanooter 2 years ago
is this based on hinton's RBM? The performance is OK..but it does make some mistakes that a human would not... is that a limitation of the NN or was the training set small?
scorpius420 2 years ago
Yes, this is based on Hinton's RBM, stacking RBMs on top of each other in what he calls Deep Belief Nets. I think the performance is not so good because I didn't tune the parameters well enough. Also, in this example I just did local layer by layer training, without backprop refining.
tioguerra 2 years ago
Yes, it is. Given the image of a handwritten digit on the left the neural network tries to produce it's best guess on the right, based on prior experience from a few different handwritten digits it had seen before.
tioguerra 2 years ago
is this like AI?
Terabytekit 2 years ago