Open terkkila opened 10 years ago
This requires that we can turn any Computational Graph into an Autoencoder. One solution is to take the CG, discard the output layer, mirror the other layers with transposed weights, learn the autoencoder, and revert the CG back to original one but with the weight values kept as they are.
Instead of randomly initializing the weight matrices, we run autoencoder pre-training before learning the classifier/regressor: http://jmlr.org/papers/volume11/erhan10a/erhan10a.pdf