ACCURATE NEURAL TRAINING WITH 4-BIT MATRIX MULTIPLICATIONS AT STANDARD FORMATS

Keyphrases

• 4-bit 100%
• 4-bit Quantization 66%
• Bit Matrix 100%
• Deep Neural Network Training 33%
• ImageNet 33%
• Log Scale 33%
• Logarithmic Quantization 33%
• Matrix multiplication 100%
• Neural Gradient 66%
• Neural Layer 33%
• Neural Network Training 33%
• Neural Training 100%
• Quantization Method 33%
• Quantized Neural Networks 33%
• Reference Implementation 33%
• ResNet50 33%
• Training Process 66%
• Variance Reduction 33%

Computer Science

• Deep Neural Network 50%
• Matrix Multiplication 100%
• Neural Network Training 100%
• Reference Implementation 50%
• ResNet50 50%
• Supplementary Material 50%
• Training Process 100%
• Variance Reduction 50%