Neural network weight compression with nnw-bdi
Published in Unknown, 2020
Citation: Andrei Bersatti, Nima Shoghi Ghalehshahi, Hyesoon Kim, Proceedings of the International Symposium on Memory Systems, 335-340, 2020 https://dl.acm.org/doi/abs/10.1145/3422575.3422805
This paper proposes NNW-BDI, a memory compression scheme for neural network weights that leverages techniques such as quantization, downscaling, randomized base selection, and base-delta-configuration adjustment. By compressing the weights of an MNIST classification network, NNW-BDI achieves up to 85% memory usage reduction without compromising inference accuracy, making it a promising solution for memory-constrained deep learning applications.