[Paper Review] Convolutional Neural Networks Demystified: A Matched Filtering Perspective Based Tutorial

Ljubiša Stanković, Danilo P. Mandic|arXiv (Cornell University)|Aug 26, 2021

Neural Networks and Applications4 citations

TL;DR

This paper reinterprets Convolutional Neural Networks (CNNs) through the lens of matched filtering, framing convolution as a feature detection mechanism that identifies predefined patterns in signals and images. By establishing a direct link between convolution operations and matched filtering theory, the authors provide a physically intuitive, mathematically grounded tutorial that demystifies CNNs' operation, back-propagation, and design choices like pooling and zero-padding, offering a theoretically sound and visually supported framework for understanding deep learning in signal and image processing.

ABSTRACT

Deep Neural Networks (DNN) and especially Convolutional Neural Networks (CNN) are a de-facto standard for the analysis of large volumes of signals and images. Yet, their development and underlying principles have been largely performed in an ad-hoc and black box fashion. To help demystify CNNs, we revisit their operation from first principles and a matched filtering perspective. We establish that the convolution operation within CNNs, their very backbone, represents a matched filter which examines the input signal/image for the presence of pre-defined features. This perspective is shown to be physically meaningful, and serves as a basis for a step-by-step tutorial on the operation of CNNs, including pooling, zero padding, various ways of dimensionality reduction. Starting from first principles, both the feed-forward pass and the learning stage (via back-propagation) are illuminated in detail, both through a worked-out numerical example and the corresponding visualizations. It is our hope that this tutorial will help shed new light and physical intuition into the understanding and further development of deep neural networks.

Motivation & Objective

• To demystify the inner workings of Convolutional Neural Networks (CNNs) by grounding them in first principles of signal processing.
• To establish a direct and physically meaningful connection between convolution operations in CNNs and matched filtering theory.
• To provide a step-by-step tutorial that explains CNNs' feed-forward and back-propagation processes using numerical examples and visualizations.
• To improve physical intuition and theoretical clarity in CNN design, especially for researchers and students in signal processing and machine learning.
• To offer a pedagogical framework that avoids black-box thinking and supports deeper understanding and future development of deep neural networks.

Proposed method

• Reinterprets the convolution layer in CNNs as a matched filter that detects predefined features by correlating the input signal with a reversed filter kernel.
• Uses the matched filter output equation $ y(n) = x(n) * w(-n) $ to formalize feature detection, where $ w(n) $ is the feature template and $ * $ denotes convolution.
• Applies the matched filtering principle to both 1D signals and 2D images, showing that convolution slides a filter across the input to detect feature presence regardless of position.
• Introduces a numerical example with visualizations to demonstrate the feed-forward pass, including convolution, activation, pooling, and fully connected layers.
• Explains back-propagation and weight updates using the matched filtering framework, linking gradient descent to optimal feature matching.
• Introduces practical CNN components such as zero-padding, strided convolutions, max-pooling, and 1×1 filters as natural extensions of matched filtering principles.

Figure 1: Example of matched filtering, with two input signals, $\mathbf{x}_{1}$ and $\mathbf{x}_{2}$ , (containing different features) observed in two different scenarios. (a) The input signal with the first feature, ${\bf x}_{1}$ . (b) The second input signal. The impulse responses of the correspo

Experimental results

Research questions

• RQ1 How can the convolution operation in CNNs be interpreted as a matched filtering process for feature detection?
• RQ2 What is the physical and mathematical justification for using convolution in CNNs rather than treating it as a black-box operation?
• RQ3 How do common CNN components like pooling, zero-padding, and 1×1 convolutions align with the matched filtering paradigm?
• RQ4 Can the back-propagation process in CNNs be understood through the lens of matched filtering and optimal feature matching?
• RQ5 How does this matched filtering perspective improve intuition and reduce the 'black box' perception of CNNs in signal and image processing?

Key findings

• The convolution operation in CNNs is mathematically equivalent to matched filtering, where each filter detects a specific feature by correlating with the input signal.
• The output of a convolution layer corresponds to the matched filter response, which peaks when the input signal contains a feature matching the filter kernel.
• The framework explains why CNNs are invariant to feature position: the convolution process slides the filter across the input, detecting features regardless of location.
• The use of 1×1 convolutions reduces parameter count by reducing channel dimensions, and this reduction is naturally explained by the matched filtering perspective.
• Max-pooling and dropout are shown to be consistent with the matched filtering framework, as they preserve or regularize feature detection in a robust manner.
• The tutorial’s numerical example achieved high training accuracy (5 epochs, 1000 iterations) and demonstrated clear visual and numerical progression of feature detection and network learning.

Figure 2: Illustration of the matched filter operation. Top panels: Two noisy input signals of the length $N=8$ (noise is denoted by $\varepsilon(n)$ ). Middle panels: Two features that we are searching for in the input signals of the length $M=3$ . The reversed versions of these features serve as t