Part of AGI experiments
VAE Convolution Kernels
What if a variational autoencoder were used as a convolution kernel?
This line asks what happens if a VAE is used as the convolution kernel itself. A four-layer network builds up a small top-level latent representation of MNIST digits, with each layer trained two ways at once: independently, like a deep belief network, and with top-down feedback from the layers above. It is an exploration of how local generative learning and global predictive feedback can be combined in a single stacked architecture.