Circuit design method for inaccurately calculating neural network

• Main Authors: ZHI TIAN, CHEN YUNJI, GUO QI, REN YIXUAN, DU ZIDONG, CHEN TIANSHI
• Format: Patent
• Language: Chinese; English
• Published: 06.07.2018
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

The invention provides a circuit design method for inaccurately calculating a neural network. The method comprises the steps that bit width is determined to determine the bit width selection range ofan inaccurate multiplier; a circuit structure is generated; a variety of specific circuits are generated for each kind of candidate bit width through a heuristic random algorithm driven by the importance ranking of nodes in the circuit, and multiple specific circuits are all optional inaccurate multiplier structures; independent replacement is carried out, and each accurate multiplier in the neural network is independently replaced with an inaccurate multiplier of a certain structure; all acquired possible solutions form a solution space; traversal comparison is carried out; a part of solutions are randomly selected within the solution space; and the training and testing effects of the inaccurate neural network under each solution are traversed to acquire the optimal solution. According tothe invention, under the