APPARATUS AND METHOD FOR EXECUTING RECURRENT NEURAL NETWORK AND LSTM COMPUTATIONS

摘要

The present disclosure provides an apparatus for performing a recurrent neural network and an LSTM operation, comprising an instruction storage unit, a controller unit, an interconnection module, a main operation module, and a plurality of slave operation modules. The slave operation module is configured to multiply and add input data to obtain a partial sum and save the partial sum until neuron data is all inputted, and a result is returned to the main operation module. The main operation module is configured to perform interpolation activation on the sum returned from the slave operation module during a forward process, and perform interpolation to get an activation derivative during a reverse process and multiplied by a gradient. The present disclosure can solve the problem that the operational performance of the CPU and the GPU is insufficient, and the power consumption of previous decoding is large, and effectively improved the support for the forward operation of the multiple layer artificial neural network.