Tcl-based Simulation Platform for Light-weight ResNet Implementation

초록

The growing computational cost and size of artificial intelligence have led to a need for hardware accelerators. However, training times for neural networks remain a significant obstacle, leading to increased simulation times and decreased productivity. In this paper, we propose a runtime layer replaceable simulation platform using a depth-reduction algorithm. The proposed platform generates weights optimized for various ResNet depths using the first trained high-depth weights. This platform can reduce simulation time by reducing the number of trainings without significantly degrading the inference accuracy. As a result of the depth reduction simulation using the MNIST data set, the accuracy was over 97% when the number of layers was reduced to less than 71%.

키워드