Wavelet-Based Prototype Networks

We develop a novel self-explainable neural network model by combining the prototype-based network architectures with wavelet representation of the inputs. The wavelet decomposition provides an invertible representation, and this representation is sent via the computation pipeline to localized prototype network. Due to the invertibility of the wavelet-representation, there is a possibility to “understand” the actual output, making the decisions interpretable. We provide experimental evidence to support that the wavelet-representation improves network interpretability, whilst keeping performance metrics.