Abstract
This paper proposes a full frequency-domain convolutional neural network (FCNN) architecture designed to explore deep learning theory within the frequency domain. By introducing a Frequency-Domain Randomized Offset Rectified Linear Unit (FRReLU) and a Frequency-Domain Chunk Max Pooling (Fcmp) method, the architecture overcomes the complexity of forward and backward pipelines in frequency-domain training. FRReLU mitigates gradient issues, while Fcmp preserves spatial context. Results using ResNet-50 show a significant reduction in error rates (Top-1 error from 24.90% to 17.95%) and a 13.03x acceleration ratio compared to cuDNN, demonstrating improved accuracy without sacrificing time performance.
Introduction
Convolutional Neural Networks (CNNs) are the backbone of modern computer vision, yet their computational demands are substantial, primarily due to time-domain convolutions. While the Fast Fourier Transform (FFT) can accelerate these operations by converting convolutions into point-wise multiplications, most existing "FFT-based" models are hybrid. They frequently toggle between time and frequency domains to apply non-linear activations and pooling, creating a significant memory bandwidth bottleneck. This paper introduces a "full" frequency-domain architecture that eliminates these transitions, maintaining data in the Fourier space from input to the final classification layers.
Methodology
The proposed FCNN architecture relies on two novel representation strategies to maintain stability and accuracy within the frequency domain:
1. FRReLU (Frequency-Domain Randomized Offset Rectified Linear Unit)
Standard activation functions like ReLU do not translate directly to the complex-valued frequency domain. FRReLU is an unsaturated activation function specifically designed for the complex plane. It processes positive and negative semi-axis functions independently and utilizes randomized offsets to introduce non-linearity. This design is crucial for preventing the vanishing and exploding gradient phenomena that typically plague frequency-domain training.
2. Fcmp (Frequency-Domain Chunk Max Pooling)
Traditional frequency-domain pooling often involves simple truncation of high-frequency components, which discards vital spatial information. Fcmp addresses this by dividing the frequency matrix into "chunks" and performing max pooling within each. This "chunk decomposition mechanism" allows the network to capture coarse-grained location information and feature strength, significantly improving classification performance.
Experimental Results
The FCNN was evaluated using the ResNet-50 backbone on the ImageNet and MetData-1 datasets.
- Accuracy Boost: The proposed architecture significantly outperformed existing frequency-domain models. On ImageNet, it reduced the top-1 error from 24.90% to 17.95% and the top-5 error from 12.85% to 9.23%.
- Massive Speedup: In high-bandwidth scenarios (batch size 128), the FCNN achieved an acceleration ratio of 13.03x compared to the NVIDIA cuDNN reference model. This speedup stems from avoiding repeated FFT/IFFT cycles in every layer.
- Stability: Testing on the MetData-1 dataset confirmed that the FCNN achieves high classification accuracy with minimal variance (average difference of 0.18), proving its robustness compared to competitors like koCNN and fbFFT.
Conclusion
By integrating FRReLU and Fcmp, this research demonstrates that it is possible to train deep learning models entirely in the frequency domain without sacrificing accuracy. The FCNN architecture offers a high-performance alternative to traditional time-domain models, particularly for deep networks where the overhead of domain switching is most pronounced. Future work will focus on optimizing chunk sensitivity and scaling the architecture to even larger datasets.
Summary updated