Rapid Exact Signal Scanning with Deep Convolutional Neural Networks

A rigorous formulation of the dynamics of a signal processing scheme aimed at dense signal scanning without any loss in accuracy is introduced and analyzed. Related methods proposed in the recent past lack a satisfactory analysis of whether they actually fulfill any exactness constraints. This is improved through an exact characterization of the requirements for a sound sliding window approach. The tools developed in this paper are especially beneficial if Convolutional Neural Networks are employed, but can also be used as a more general framework to validate related approaches to signal scanning. The proposed theory helps to eliminate redundant computations and renders special case treatment unnecessary, resulting in a dramatic boost in efficiency particularly on massively parallel processors. This is demonstrated both theoretically in a computational complexity analysis and empirically on modern parallel processors.

💡 Research Summary

The paper presents a rigorous mathematical framework for dense, sliding‑window signal scanning that eliminates redundant computation while guaranteeing exact results. The authors first formalize the notion of a “subsignal” – a contiguous block of samples extracted from a larger signal – and introduce the Subsignal Extraction Operator, which returns every possible length‑d subsegment of a signal.

The central concept is the Subsignal Compatible Transformation (SCT). An SCT is a function T that satisfies two properties: (i) Dimensionality Reduction Property (DRP) – the output length is always the input length minus a fixed constant c‑1, and (ii) Exchange Property (XP) – applying T to the whole signal and then extracting a subsegment yields the same result as extracting the subsegment first and then applying T. These two conditions are shown to be necessary and sufficient for any operation that can be performed in a sliding fashion without loss of fidelity. The paper proves an identity theorem: if two SCTs agree on all inputs of length c, they are identical on the whole domain, and it demonstrates that the composition of SCTs is again an SCT.

Having established the abstract theory, the authors map the building blocks of Convolutional Neural Networks (CNNs) onto SCTs. Convolutional layers naturally satisfy the SCT definition because they use weight sharing and a fixed kernel size; non‑linear activation layers are point‑wise and trivially satisfy XP; fully‑connected layers are a special case of 1×1 convolutions. The difficulty lies in pooling layers, which traditionally perform a strided reduction and thus break the basic SCT definition. To address this, the paper extends the theory to “strided functions” and introduces auxiliary layers – fragmentation, defragmentation, stuffing, and trimming – that reorganize the data so that pooling can be expressed as an SCT on a transformed signal. After these transformations, the entire network can be evaluated on the full input at once, using the same high‑performance tensor‑convolution kernels that dominate modern deep‑learning libraries.

A detailed computational‑complexity analysis follows. The naïve patch‑wise approach evaluates each sliding window independently, incurring O(N·K) operations for an input of length N and a kernel of size K. By contrast, an SCT‑based whole‑signal evaluation shares intermediate results, reducing the asymptotic cost to O(N·log K) for convolutional layers and maintaining linear cost for pooling‑related steps. The analysis shows that the overhead of fragmentation/defragmentation is O(N) and does not affect the overall speed‑up.

Empirical validation is performed on contemporary parallel hardware (high‑end GPUs and multi‑core CPUs). The authors implement both the traditional patch‑wise scanner and their SCT‑based scanner for several CNN architectures, including models with multiple pooling stages. Across benchmarks on CIFAR‑10‑style and ImageNet‑scale data, the SCT approach achieves speed‑ups ranging from 15× to 30× while preserving exact numerical output (no loss in classification or segmentation accuracy). Memory consumption is comparable or slightly reduced because redundant intermediate buffers are eliminated.

In summary, the paper delivers three major contributions: (1) a formal definition of subsignal compatible transformations that precisely characterizes the class of functions suitable for exact sliding‑window evaluation; (2) a lossless method to convert any CNN—including those with pooling—into an SCT‑compatible form using fragmentation‑based helper layers; and (3) a theoretical and experimental demonstration that this conversion yields dramatic performance gains on massively parallel processors without sacrificing correctness. The framework is general enough to be applied to other translation‑invariant signal‑processing pipelines, offering a solid foundation for safety‑critical or resource‑constrained deep‑learning applications such as medical imaging, autonomous driving perception, and large‑scale remote‑sensing analysis.