The present invention performs calculations of gradient descent in secret calculations, at high speed while maintaining accuracy. This secret gradient descent calculation method calculates gradient descent while hiding gradients and parameters. An initialization unit initializes (S11) hidden values [M], [V] for matrices M, V. A gradient calculation unit finds (S12) a hidden value [G] for matrix G having gradient g. A parameter updating unit: calculates [M] ← β 1 [M] + (1- β1) [G] (S13-1); calculates [V] ← β2 [V] + 1-β2)[G] ○[G] (S13-2); calculates [M^] ← β^1, t [M] (S13-3); calculates [V^] ← β^2, t [V] (S13-4); calculates [G^] ← Adam ([V^]) (S13-5); calculates [G^] ← [G^] ○ [M^] (13-6); and calculates [W] ←[W]-[G^] (S13-7).
展开▼