In this paper, we exploit the unique properties of a deterministic projected
belief network (D-PBN) to take full advantage of trainable compound activation
functions (TCAs). A D-PBN is a type of auto-encoder that operates by "backing
up" through a feed-forward neural network. TCAs are