Computation And Memory Optimized Spectral Domain Convolutional Neural

Conventional convolutional neural networks (CNNs) present a high computational workload and memory access cost (CMC). Spectral domain CNNs (SpCNNs) offer a computationally efficient approach to compute CNN training and inference. This paper investigates CMC of SpCNNs and its contributing components analytically and then proposes a methodology to optimize CMC, under three strategies, to enhance inference performance. In this methodology, output feature map (OFM) size, OFM depth or both are progressively reduced under an accuracy constraint to compute performance-optimized CNN inference. Before conducting training or testing, it can provide designers guidelines and preliminary insights regarding techniques for optimum performance, least degradation in accuracy and a balanced performance–accuracy trade-off. This methodology was evaluated on MNIST and Fashion MNIST datasets using LeNet-5 and AlexNet architectures.

When compared to state-of-the-art SpCNN models, LeNet-5 achieves up to 4.2× (batch inference) and 4.1× (single-image inference) higher throughputs and 10.5× (batch inference) and 4.2× (single-image inference) greater energy efficiency at a maximum loss... When compared to the baseline model used in this study, AlexNet delivers 11.6× (batch inference) and 5× (single-image inference) increased throughput and 25× (batch inference) and 8.8× (single-image inference) more energy-efficient inference with just... This is a preview of subscription content, log in via an institution to check access. Price excludes VAT (USA) Tax calculation will be finalised during checkout. The datasets employed to evaluate the CNN models in the current study are available in [62] and https://github.com/zalandoresearch/fashion-mnist. Alzubaidi L, Zhang J, Humaidi A J, Al-Dujaili A, Duan Y, Al-Shamma O, Santamaría J, Fadhel M A, Al-Amidie M, Farhan L (2021) Review of deep learning-concepts, CNN architectures, challenges, applications, future directions.

J Big Data 8(1):1–74 A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity.© Copyright 2025 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions. This article is made available via the PMC Open Access Subset for unrestricted research re-use and secondary analysis in any form or by any means with acknowledgement of the original source. These permissions are granted for the duration of the World Health Organization (WHO) declaration of COVID-19 as a global pandemic. Conventional convolutional neural networks (CNNs) present a high computational workload and memory access cost (CMC).

Spectral domain CNNs (SpCNNs) offer a computationally efficient approach to compute CNN training and inference. This paper investigates CMC of SpCNNs and its contributing components analytically and then proposes a methodology to optimize CMC, under three strategies, to enhance inference performance. In this methodology, output feature map (OFM) size, OFM depth or both are progressively reduced under an accuracy constraint to compute performance-optimized CNN inference. Before conducting training or testing, it can provide designers guidelines and preliminary insights regarding techniques for optimum performance, least degradation in accuracy and a balanced performance–accuracy trade-off. This methodology was evaluated on MNIST and Fashion MNIST datasets using LeNet-5 and AlexNet architectures. When compared to state-of-the-art SpCNN models, LeNet-5 achieves up to 4.2× (batch inference) and 4.1× (single-image inference) higher throughputs and 10.5× (batch inference) and 4.2× (single-image inference) greater energy efficiency at a maximum loss...

When compared to the baseline model used in this study, AlexNet delivers 11.6× (batch inference) and 5× (single-image inference) increased throughput and 25× (batch inference) and 8.8× (single-image inference) more energy-efficient inference with just... Keywords: Convolutional neural network, Spectral domain CNN, Energy efficiency, Computational workload, Memory access cost Deep neural networks (DNNs) have recently evolved as the prevalent solution for a range of challenging problems in computer vision, such as image recognition [1], image segmentation [2–4], set based image classification [5], and... Convolutional neural networks (CNNs) [9, 10], a class of DNNs, have achieved unprecedented success in various fields of computer vision, audio analysis, and text processing including–inter alia–object classification [11, 12], object detection [13, 14],... In the last few years, CNNs have been deployed in diverse applications such as autonomous driving [23], navigation systems [24] and flight safety [25] for drones, skin cancer detection [26], COVID-19 prognosis [27], and... In CNNs, convolution layers play a central role in feature-extraction [9–11] but demand high computational resources [29].

They account for 90% of CNN operations [30]. Moreover, deeper CNNs (with more convolution layers), which tend to produce higher accuracy, possess a larger number of parameters. This results in increased memory requirements [29]. However, run-time memory during inference is dominated by storage of intermediate output feature maps [31], even with a batch size of 1 [32]. Memory consumed by these output feature maps (OFMs) can exceed parameter memory by 10 to a 100 times [33]. Storing intermediate OFMs requires off-chip dynamic random access memory (DRAM) accesses, which consume more power and energy than computations [29, 34].

In fact, for devices limited by memory bandwidth, the memory access cost can be the main bottleneck for power consumption and inference latency including in GPU-based platforms [15, 35, 36]. Therefore, the computational, memory and power budgets of even classical CNNs (with only a few convolutional layers) are such that deploying them in embedded systems or mobile platforms is extremely challenging [29, 30, 32,... Conventional convolutional neural networks (CNNs) present a high computational workload and memory access cost (CMC). Spectral domain CNNs (SpCNNs) offer a computationally efficient approach to compute CNN training and inference. This paper investigates CMC of SpCNNs and its contributing components analytically and then proposes a methodology to optimize CMC, under three strategies, to enhance inference performance. In this methodology, output feature map (OFM) size, OFM depth or both are progressively reduced under an accuracy constraint to compute performance-optimized CNN inference.

Before conducting training or testing, it can provide designers guidelines and preliminary insights regarding techniques for optimum performance, least degradation in accuracy and a balanced performance–accuracy trade-off. This methodology was evaluated on MNIST and Fashion MNIST datasets using LeNet-5 and AlexNet architectures. When compared to state-of-the-art SpCNN models, LeNet-5 achieves up to 4.2× (batch inference) and 4.1× (single-image inference) higher throughputs and 10.5× (batch inference) and 4.2× (single-image inference) greater energy efficiency at a maximum loss... When compared to the baseline model used in this study, AlexNet delivers 11.6× (batch inference) and 5× (single-image inference) increased throughput and 25× (batch inference) and 8.8× (single-image inference) more energy-efficient inference with just... This is a preview of subscription content, log in via an institution to check access. Price excludes VAT (USA) Tax calculation will be finalised during checkout.

The datasets employed to evaluate the CNN models in the current study are available in [62] and https://github.com/zalandoresearch/fashion-mnist. Alzubaidi L, Zhang J, Humaidi A J, Al-Dujaili A, Duan Y, Al-Shamma O, Santamaría J, Fadhel M A, Al-Amidie M, Farhan L (2021) Review of deep learning-concepts, CNN architectures, challenges, applications, future directions. J Big Data 8(1):1–74

People Also Search

• Computation and memory optimized spectral domain convolutional neural ...
• PDF arXiv:1905.10915v6 [cs.CV] 8 Feb 2021
• Computation and memory optimized spectral domain - ProQuest
• Spectral Domain Convolutional Neural Network - IEEE Xplore
• Spectral domain convolutional neural network optimized for ...
• Spectral-based convolutional neural network without multiple spatial ...
• Spectral Domain Convolutional Neural Network and Its Applications ...
• Optimizing Fully Spectral Convolutional Neural Networks on FPGA
• PDF Computation and memory optimized spectral domain convolutional neural ...