PeakRNN and StatsRNN: Dynamic Pruning in Recurrent Neural Networks

This paper introduces two dynamic real-time pruning techniques PeakRNN and StatsRNN for reducing costly multiplications and memory accesses in recurrent neural networks. The methods are demonstrated on a gated recurrent unit in a multi-layer network, solving a single-channel speech enhancement task with a wide variety of real-world acoustic environments and speakers. The performance is compared against the baseline gated recurrent unit and the DeltaRNN method. Compared to the unprocessed speech, the SNR and Perceptual Evaluation of Speech Quality were on average improved by 8.11 dB and 0.43 MOS-LQO, respectively. Additionally, the two proposed methods outperformed DeltaRNN by 0.7 dB and 0.11 MOS-LQO in the two objective measures, while using the same computational budget per timestep and reducing the original operations by 88%. Furthermore, PeakRNN is fully deterministic, i.e. it is always known in advance how many computations will be executed. Such worst-case guarantees are crucial for real-time acoustics applications.