| Title |
A Greedy Neural Network Partitioning Algorithmfor Crossbar-based Neuromorphic Computing Syste |
| Authors |
김충민(Choongmin Kim) ; 정재용(Jaeyong Chung) |
| DOI |
https://doi.org/10.5573/ieie.2019.56.11.22 |
| Keywords |
Deep learning ; Deep neural network ; Efficient deep learning ; Neuromorphic system ; Neuromorphic computing system ; |
| Abstract |
Neuromorphic systems are attracting attention as a energy-efficient platform for neural network through many studies. However, the hardware neurons of the neuromorphic system have a fixed number of synapses, and it is impossible to map a perceptron of neural network with a large number of connections one-to-one. Therefore, in order to map a perceptron to a hardware neuron, the perceptron must be decomposed. Percetron decomposition affects the predictive performance of neral network and the neuron usage of the neuromorphic system. In this paper, we propose a greedy neural network partition algorithm for crossbar-based neuromorphic systems using dynamic-fixed point. The proposed algorithm is applied to each layer of the network. It is a greedy algorithm that finds the suitable connections that can be mapped to the crossbar as much as possible. To reduce the quantization error, the scaling factors of the mapped connections are restricted to be the same. we perform experiments on four network, Alexnet, VGG11, VGG13 and VGG16. we reduce the accuracy loss caused by weight quanzation by using a very small number of extra neurons. |