Convolutional Neural Networks Demystified: A Matched Filtering Perspective-Based Tutorial

Deep neural networks (DNNs) and especially convolutional (CNNs) have revolutionized the way we approach analysis of large quantities data. However, largely ad hoc fashion their development, albeit one reason for rapid success, has also brought to light intrinsic limitations CNNs—in particular, those related black box nature. In addition, ability “explain” both such systems behave results they produce is increasingly becoming an imperative in many practical applications. Therefore, it would be particularly useful establish physically meaningful mechanisms underpinning operation CNNs, thus helping resolve issue interpretability processing steps explain input-output relationship. To this end, revisit CNNs from first principles show that very backbone—the convolution operation—represents a matched filter which examines input presence characteristic patterns Our treatment based on temporal signals, naturally generated by physical sensors, admit rigorous through science. This serves as vehicle unifying account overall functionality whereby convolution-activation-pooling chain learning strategies are shown compact elegant interpretation under umbrella filtering. addition reveal providing intuitive understanding operation, filtering perspective offer platform support further developments area.