Neural network model contraction device

摘要

PROBLEM TO BE SOLVED: To suppress an increase in a design period related to model contraction while combining a plurality of contraction methods. A model contraction device (1) of a neural network uses a first contraction parameter to perform a first contraction process for changing an element of a weight matrix, and a weight calculation processing unit (first) for updating the weight matrix. (Reduction processing unit) 30 and the second reduction parameter are used to perform the second reduction processing to reduce the size of the updated weight matrix, and the network deformation to deform the network shape corresponding to the reduced weight matrix. A resynthesis processing unit (second contraction processing unit) 50 is provided. [Selection diagram] Fig. 3