Accelerating Very Deep Convolutional Networks for Classification and Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are crucial for computer vision.
• Very deep CNNs offer high performance but are computationally expensive.
• Existing acceleration methods often approximate linear components, limiting effectiveness.

Purpose of the Study:

• To accelerate the test-time computation of very deep CNNs.
• To develop a method that accounts for nonlinear units in CNNs.
• To reduce accumulated errors in multi-layer approximations.

Main Methods:

• Developed a nonlinear optimization solution, avoiding Stochastic Gradient Descent (SGD).
• Implemented an asymmetric reconstruction technique for multi-layer approximation.
• Applied the method to the VGG-16 model.

Main Results:

• Achieved a 4x whole-model speedup for the VGG-16 model.
• Introduced a minimal 0.3% increase in top-5 error on ImageNet classification.
• Demonstrated graceful accuracy degradation in object detection with Fast R-CNN.

Conclusions:

• The proposed nonlinear optimization method effectively accelerates deep CNNs.
• The approach significantly reduces computational cost with negligible impact on accuracy.
• This method offers a promising solution for deploying deep learning models in resource-constrained environments.