Approximating the Hotelling observer with autoencoder-learned efficient channels for binary signal detection tasks.