Skip to main content
Figure 4 | Fixed Point Theory and Algorithms for Sciences and Engineering

Figure 4

From: Learning without loss

Figure 4

The three types of network featured in this tutorial: (a) single layer network for nonnegative matrix factorization (Sect. 4), (b) multi-layer classifier network (Sect. 5), (c) autoencoder network (Sect. 6). Networks are not required to be layered, except for special sets of nodes: data layer (blue), code layer (green), class layer (red). Classifier and autoencoder networks may be arbitrarily deep, with layers of “hidden” nodes (black). In the autoencoder network the single data layer is rendered twice (top and bottom) and the network is cyclic. There are unknown weight parameters on all the edges that training seeks to reconstruct from data. Data constrains only the blue and red nodes, where the red nodes in the classifier are special in that they are indicator variables for the classes. Neuron (post-activation) values at all the black and green nodes also get reconstructed, where the green nodes are interpreted as “codes” for the representation of data

Back to article page