Recursive Deep Learning Quantization for Weight Bit Reduction

The patent applies parameter changes by systematically varying the bit resolution of network weights and data inputs across different layers of the deep-learning network. Instead of using a fixed precision level throughout, the method dynamically adjusts precision parameters to match the computational requirements of each specific layer, thereby reducing overall power consumption while maintaining necessary model precision where critical.

The patent segments the deep-learning network into multiple layers and applies different quantization strategies to different segments. Each layer is analyzed independently to determine its specific precision requirements, allowing high-precision computation only where necessary and lower-precision computation in other layers, thus optimizing the trade-off between model precision and power consumption across the entire network.

The patent changes the precision parameter from a fixed 64-bit representation to variable bit representations ranging from 8-bit down to lower precisions for different network components. This parameter adaptation allows the hardware implementation to achieve the necessary model precision with significantly reduced storage requirements, making the physical size practical for deployment.

The patent divides the network into segments with different precision requirements, storing high-precision weights only in layers where they are critical for model accuracy, while using lower-precision representations in other layers. This segmented approach reduces the total volume of storage required while maintaining the necessary model precision.

The patent applies local quality by assigning different quantization levels to different layers and components of the network based on their specific requirements. Critical layers that require high model precision maintain higher bit resolutions, while less sensitive layers use lower precision quantization. This localized approach ensures that model precision is maintained where necessary while achieving power consumption reductions through quantization in other areas.

The patent dynamically adjusts the precision parameter for each layer based on its computational characteristics and importance to the overall model accuracy. By changing the bit resolution parameter locally rather than applying a uniform quantization level, the system maintains model precision in critical areas while achieving energy efficiency through quantization in less sensitive areas.

The patent applies local quality by implementing non-uniform quantization strategies tailored to each layer's characteristics. Instead of applying a coarse quantization level uniformly across all layers, the method uses finer quantization steps in layers where high accuracy is critical and coarser steps where hardware constraints are more pressing. This localized adaptation achieves both hardware feasibility and adequate quantization accuracy.

The patent introduces dynamics by making the quantization parameters adaptive rather than static. The quantization levels and precision requirements are determined dynamically based on the specific layer's computational needs and its contribution to overall model accuracy. This dynamic approach allows the system to optimize between hardware implementation feasibility and quantization accuracy for each layer individually.