Neural Network Subgroup Scaling for Accuracy

The patent applies local quality by introducing subgroup-specific scaling coefficients that operate locally on different regions of the weight space. Instead of using a single global scaling factor, the weight space is divided into subgroups (e.g., along different dimensions of the kernel filter or input feature map), and each subgroup receives its own scaling coefficient. This local adaptation allows the system to maintain higher prediction accuracy while still using binarized or ternary weights, as each local region can be optimized independently to compensate for the quantization effects.

The patent changes the parameters of the neural network representation by introducing scaling coefficients that modify the effective weight values. By applying these scaling coefficients to subgrouped weights, the system can achieve better approximation of full-precision weights without increasing the actual number of stored parameters. The scaling coefficients are learned during training and allow the binarized/ternary network to adapt its parameter distribution to maintain accuracy while using fewer bits per weight.

The patent applies segmentation by dividing the weight space into multiple subgroups based on spatial or functional criteria. This segmentation allows the computational operations to be organized in a structured manner where scaling coefficients are applied to specific subgroups rather than all weights uniformly. The segmentation enables efficient parallel computation and reduces the computational overhead by localizing the scaling operations to relevant weight subsets, thereby improving accuracy without proportionally increasing the overall computational footprint.