TY - GEN
T1 - ProdNet
T2 - 5th International Conference on Intelligent Systems and Pattern Recognition, ISPR 2025
AU - Ouissam Lakas, Badia
AU - Berdjouh, Chemousse
AU - Bounane, Khadra
AU - Lamine Kherfi, Mohammed
AU - Aiadi, Oussama
AU - Brahim Belhouari, Samir
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
PY - 2026/6/3
Y1 - 2026/6/3
N2 - Matrix multiplication is a fundamental operation with significant applications across diverse fields, such as physics, electronics, and artificial intelligence. Traditional implementations of this operation exhibit cubic time complexity, which presents computational challenges, particularly in deep learning scenarios that necessitate large-scale matrix computations. In this study, we introduce ProdNet, a lightweight neural network model designed to autonomously discover efficient matrix multiplication algorithms without the need for extensive computational resources or prior knowledge of existing methods. Our approach seeks to alleviate complexity by minimizing the number of multiplicative operations involved. To achieve this, we utilize a combination of Mean Squared Error (MSE) and a regularization function, targeting weight values to be constrained to 0, 1, or -1. We emphasize the effectiveness of our regularization techniques in accelerating the algorithm discovery process. Furthermore, we demonstrate that ProdNet successfully identifies multiplication algorithms for 2x2 and 3x3 matrices, achieving this in a limited number of iterations (starting from 500 iterations for 2 x 2 and 8800 iterations for 3 x 3)
AB - Matrix multiplication is a fundamental operation with significant applications across diverse fields, such as physics, electronics, and artificial intelligence. Traditional implementations of this operation exhibit cubic time complexity, which presents computational challenges, particularly in deep learning scenarios that necessitate large-scale matrix computations. In this study, we introduce ProdNet, a lightweight neural network model designed to autonomously discover efficient matrix multiplication algorithms without the need for extensive computational resources or prior knowledge of existing methods. Our approach seeks to alleviate complexity by minimizing the number of multiplicative operations involved. To achieve this, we utilize a combination of Mean Squared Error (MSE) and a regularization function, targeting weight values to be constrained to 0, 1, or -1. We emphasize the effectiveness of our regularization techniques in accelerating the algorithm discovery process. Furthermore, we demonstrate that ProdNet successfully identifies multiplication algorithms for 2x2 and 3x3 matrices, achieving this in a limited number of iterations (starting from 500 iterations for 2 x 2 and 8800 iterations for 3 x 3)
KW - Matrix product algorithm
KW - Neural network
KW - ProdNet
KW - Regularization
UR - https://www.scopus.com/pages/publications/105039847313
U2 - 10.1007/978-3-032-21585-7_7
DO - 10.1007/978-3-032-21585-7_7
M3 - Conference contribution
AN - SCOPUS:105039847313
SN - 9783032215840
VL - 2859
T3 - Communications In Computer And Information Science
SP - 95
EP - 108
BT - Intelligent Systems And Pattern Recognition, Ispr 2025, Pt Ii
A2 - Ensari, T
A2 - Bennour, A
A2 - Bouaziz, B
A2 - Rida, I
A2 - Mahdi, W
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 25 September 2025 through 27 September 2025
ER -