[Paper Review] Computational optimization of convolutional neural networks using separated filters architecture

Elena Limonova, Alexander Sheshkus+1 more|arXiv (Cornell University)|Feb 18, 2020

Advanced Data Processing Techniques6 citations

TL;DR

This paper proposes a computational optimization for convolutional neural networks (CNNs) by expressing 2D convolutional filters as a linear combination of separable filters, enabling efficient training and inference. The method achieves a 15% speedup in letter and digit recognition tasks without accuracy loss, making it suitable for mobile and embedded systems.

ABSTRACT

This paper considers a convolutional neural network transformation that reduces computation complexity and thus speedups neural network processing. Usage of convolutional neural networks (CNN) is the standard approach to image recognition despite the fact they can be too computationally demanding, for example for recognition on mobile platforms or in embedded systems. In this paper we propose CNN structure transformation which expresses 2D convolution filters as a linear combination of separable filters. It allows to obtain separated convolutional filters by standard training algorithms. We study the computation efficiency of this structure transformation and suggest fast implementation easily handled by CPU or GPU. We demonstrate that CNNs designed for letter and digit recognition of proposed structure show 15% speedup without accuracy loss in industrial image recognition system. In conclusion, we discuss the question of possible accuracy decrease and the application of proposed transformation to different recognition problems. convolutional neural networks, computational optimization, separable filters, complexity reduction.

Motivation & Objective

• Address the high computational cost of CNNs in resource-constrained environments such as mobile and embedded platforms.
• Reduce the computational complexity of 2D convolution operations without compromising model accuracy.
• Enable efficient training and inference using standard optimization algorithms on transformed filter architectures.
• Demonstrate practical speedup in industrial image recognition systems while maintaining recognition performance.

Proposed method

• Represent 2D convolutional filters as a linear combination of 1D separable filters to reduce computational complexity.
• Apply standard backpropagation and stochastic gradient descent for end-to-end training of the transformed network.
• Implement the separated filter architecture efficiently on both CPU and GPU hardware.
• Use a parameter-sharing mechanism to constrain the filter decomposition while preserving representational capacity.
• Design the network such that the decomposition is differentiable and compatible with standard deep learning frameworks.
• Optimize inference by replacing 2D convolutions with cascaded 1D convolutions, reducing FLOPs and memory access.

Experimental results

Research questions

• RQ1 Can 2D convolutional filters be effectively decomposed into separable filters without significant accuracy degradation?
• RQ2 To what extent can computational complexity be reduced using this filter decomposition approach?
• RQ3 How does the proposed method perform in terms of inference speed on real-world recognition tasks?
• RQ4 Is the transformation compatible with standard training procedures and hardware accelerators?
• RQ5 What is the trade-off between model complexity and speedup in practical deployment scenarios?

Key findings

• The proposed method achieves a 15% speedup in inference time for letter and digit recognition tasks without any accuracy loss.
• The transformed CNN structure maintains high performance on industrial image recognition benchmarks.
• The method is compatible with standard training algorithms and can be efficiently implemented on both CPU and GPU.
• The computational complexity is significantly reduced due to the use of separable convolutions after filter decomposition.
• The approach enables efficient deployment of CNNs on mobile and embedded platforms with minimal architectural changes.
• No noticeable drop in model accuracy was observed, even after the transformation and retraining.