Acceleration of DNN Training Regularization: Dropout Accelerator

Abstract

The training time of a deep neural network has increased such that training process may take many days or even weeks using a single device. Further, conventional devices such as CPU and GPU pursuit generality on their use, it is inevitable that they have drawbacks on energy efficiency for a specific use case such as DNN training. So accelerating DNN training is becoming an important part of DNN accelerator to achieve a high energy efficiency. This paper proposes an idea to save both execution time and DRAM energy consumption during DNN training by implementing dropout hardware efficiently. Simulation results show that our idea can save execution time and DRAM energy consumption on backward propagation as much as the dropped activations.