DEFESA DE DISSERTAÇÃO DE MESTRADO
Aluno: Rafael Fão de Moura
Orientador: Prof. Dr. Luigi Carro
Título: Software-only Computation Reuse Techniques for Energy Efficient CNNs
Linha de Pesquisa: Sistemas Embarcados
Local: Prédio 43412, sala 215 (Sala de videoconferência) do Instituto de Informática da UFRGS.
Prof. Dr. Álvaro Freitas Moreira (UFRGS)
Prof. Dr. Gabriel Luca Nazar (UFRGS)
Prof. Dr. Ricardo dos Santos Ferreira (UFV – por videoconferência)
Presidente da Banca: Prof. Dr. Luigi Carro
Abstract: In the past years, several efforts in algorithm and architectural research were put together to enable large-scale use of CNNs as we know today. Thus far, most of these achievements have been based on improving convolutions by chasing the parallel execution of MAC operations through the replication of floating-point units. However, these solutions fall far short of what is allowed from the energy budget when it comes to embedded systems running these NN models. Given specific image characteristics, such as recurrent input patterns, we propose an algorithmic changing for performing CNN inferences by employing a computation reuse technique instead of the original implementation. Based on statistical analysis, we address computation reuse at three granularity levels: convolution kernel-level and grid-level through employing lookup tables in place of the original convolutions, and frame-level by replacing entire frame computations with a movement prediction algorithm. Experimental results show that it is possible to achieve energy savings up to 27.5×, while reducing the inference time to 116× of the baseline, with an accuracy loss of 13%.
Keywords: Convolutional neural networks. computation reuse.