Sparse BD-Net: A multiplication-less DNN with sparse binarized depth-wise separable convolution

Keyphrases

• Binarization 100%
• Deep Convolutional Neural Network (deep CNN) 100%
• Depthwise Separable Convolution 100%
• Subtraction by Addition 66%
• Energy Efficient 33%
• Hardware Efficiency 33%
• Design Guidance 33%
• Parallelization 33%
• Hardware-software Co-design 33%
• Intuitive Design 33%
• Subtraction Technique 33%
• Experiment Results 33%
• In-memory 33%
• Group Size 33%
• Model Size 33%
• Sparsity Pattern 33%
• Convolution Operation 33%
• Inference Accuracy 33%
• Convolution Block 33%
• SOT-MRAM 33%
• Neural Network Acceleration 33%
• Multiplication-and-accumulation 33%
• Drop-in Replacement 33%
• Spatial Convolution 33%
• Weight Penalty 33%
• Journal Submission 33%
• Binarization Method 33%
• Weight Binarization 33%
• Disharmony 33%
• Mobile Embedded Systems 33%
• Modified Groups 33%
• Group Lasso Regularization 33%
• Structured Sparsity 33%
• Channel Expansion 33%
• Clipping Technique 33%
• Lasso Regularization 33%
• Separable Convolution Neural Network 33%

Computer Science

• Convolutional Neural Network 100%
• Regularization 50%
• binarization 50%
• Depthwise Separable Convolution 50%
• Energy Efficient 25%
• Parallelism 25%
• Embedded Systems 25%
• Convolutional Neural Network 25%
• Extended Version 25%
• Design Guidance 25%
• Sparsity Pattern 25%
• Convolution Operation 25%
• Accelerator Memory 25%
• Hardware Co-Design 25%