公开(公告)号：EP4528592A1

公开(公告)日：2025-03-26

申请号：EP24764117.8

申请日：2024-02-20

Applicant: Korea Advanced Institute of Science and Technology

Inventor： YOO, Hoi Jun ,  KIM, Sang Yeob

IPC: G06N3/063 ,  G06N3/04 ,  G06N3/0464 ,  G06N3/045 ,  G06N3/08 ,  G06F17/15

Abstract: A complementary deep neural network accelerator having a heterogeneous convolutional neural network and spiking neural network core architecture according to the present invention comprises: an accumulator array spiking neural network processing module that accumulates weights of synapses during spike generation and generates voltages of neurons; a multiplier-accumulator array convolutional neural network processing module that accumulates the products of inputs and weights of a neural network and generates output values of neurons; a highest RISC controller responsible for controlling the spiking neural network processing module and the convolutional neural network processing module, and processing an activation function and batch normalization; an attention module that performs channel-by-channel pooling of the inputs and then performs convolution using pre-trained weights to generate an attention map; and a neural network operation allocator that divides the inputs into multiple tiles, calculates the frequency of spikes generated in each tile to predict a neural network processing module with low energy consumption, and transfers the tiles to the corresponding neural network processing modules to enable operations to be performed. Therefore, the complementary deep neural network accelerator has the effect that energy efficiency can be increased while maintaining the inference accuracy and training accuracy of the deep neural network accelerator through the complementation of the spiking neural network and the convolutional neural network.