Latent code-based fusion: A Volterra neural network approach

We propose a deep structure encoder using Volterra Neural Networks (VNNs) to seek latent representation of multi-modal data whose features are jointly captured by union subspaces. The so-called self-representation embedding the codes leads simplified fusion which is driven similarly constructed decoding. Filter architecture achieved reduction in parameter complexity primarily due controlled non-linearities being introduced higher-order convolutions lieu generalized activation functions. Experimental results on two different datasets have shown significant improvement clustering performance for VNNs auto-encoder over conventional Convolutional (CNNs) auto-encoder. In addition, we also show that proposed approach demonstrates much-improved sample CNN-based with robust classification performance.