| 摘要 |
PROBLEM TO BE SOLVED: To enable improvement in recognition performance while reducing an operation quantity.SOLUTION: A model learning device according to an embodiment comprises a conversion part, an allocation part, an update part and a projection part. The conversion part converts each of N (N≥1) covariance matrices and obtains N logarithm covariance vectors. The allocation part allocates each of the N logarithm covariance vectors to a proximate rotation matrix of K (1≤K≤N) rotation matrices obtained from the N covariance matrices. The update part, with respect to each of K' (1≤K'≤K) rotation matrices, identifies the logarithm covariance vector allocated to the rotation matrix, and updates the rotation matrix on the basis of the identified logarithm covariance vector. The projection part projects each of the N logarithm covariance vectors to each proximate rotation matrix in the K' updated rotation matrices and in the K-K' non-updated rotation matrices. |
