SparseTrain: Leveraging dynamic sparsity in software for training DNNs on general-purpose SIMD processors

Keyphrases

• Sparsity 100%
• SIMD Processors 100%
• Dynamic Sparsity 100%
• Backward Propagation 66%
• Major Components 33%
• Data Representation 33%
• Specialized Hardware 33%
• Vectorized 33%
• Dense Datasets 33%
• ImageNet 33%
• Location of Zeros 33%
• Forward Propagation 33%
• Leverage Dynamics 33%
• Convolutional Layer 33%
• Deep Learning Applications 33%
• Weight Sparsity 33%
• VGG16 33%
• Intel Skylake 33%
• ResNet50 33%
• Direct Convolution 33%
• Squeeze-and-excitation Network (SENet) 33%
• ReLU Activation Function 33%

Computer Science

• Sparsity 100%
• SIMD 100%
• Deep Neural Network 100%
• Residual Neural Network 33%
• Data Representation 16%
• Major Component 16%
• Deep Learning 16%
• Forward Propagation 16%
• Convolutional Layer 16%
• Activation Function 16%