BottleNet++: An end-to-end approach for feature compression in device-edge co-inference systems

Keyphrases

• Feature Compression 100%
• Inference System 100%
• Device-Edge Co-Inference 100%
• Mobile Devices 75%
• Encoder 50%
• Decoder 50%
• Intermediate Features 50%
• Neural Network 25%
• Resource Consumption 25%
• Classification Accuracy 25%
• Trainable 25%
• Sparsity 25%
• Communication Overhead 25%
• Compression Ratio 25%
• Fault-tolerant 25%
• High Compression Ratio 25%
• Bandwidth Enhancement 25%
• Promising Solutions 25%
• Resource-constrained 25%
• Additive White Gaussian Noise Channel 25%
• Channel Noise 25%
• Intermediate Data 25%
• Efficient Features 25%
• Deep Neural Network 25%
• Binary Erasure Channel 25%
• Deep Learning Model 25%
• Edge Computing 25%
• Intelligent Mobile Applications 25%
• Computing Servers 25%
• Feature Transmission 25%
• Online Computation 25%
• Channel Layer 25%
• Splitting Point 25%
• Model Splitting 25%
• End-to-end Architecture 25%
• Lightweight Neural Network 25%
• Application Demand 25%
• Inference Framework 25%
• Joint Source-channel Coding 25%
• Bit Compression 25%

Computer Science

• Mobile Device 100%
• Inference System 100%
• Compression Ratio 66%
• Deep Learning Model 33%
• Fault Tolerant 33%
• Deep Neural Network 33%
• Convolutional Neural Network 33%
• Sparsity 33%
• Communication Overhead 33%
• Additive White Gaussian Noise 33%
• Intermediate Data 33%
• Edge Computing 33%
• joint source channel coding 33%
• Neural Network 33%
• Mobile Application 33%
• Resource Consumption 33%