Buffer Sizes Reduction for Memory-efficient CNN Inference on Mobile and Embedded Devices

Nowadays, convolutional neural networks (CNNs) are the core of many intelligent systems, including those that run on mobile and embedded devices. However, the execution of computationally demanding and memory-hungry CNNs on resource-limited mobile and embedded devices is quite challenging. One of the main problems, when running CNNs on such devices, is the limited amount of memory available. Thus, reduction of the CNN memory footprint is crucial for the CNN inference on mobile and embedded devices. The CNN memory footprint is determined by the amount of memory required to store CNN parameters (weights and biases) and intermediate data, exchanged between CNN operators. The most common approaches, utilized to reduce the CNN memory footprint, such as pruning and quantization, reduce the memory required to store the CNN parameters. However, these approaches decrease the CNN accuracy. Moreover, with the increasing depth of the state-of-the-art CNNs, the intermediate data exchanged between CNN operators takes even more space than the CNN parameters. Therefore, in this paper, we propose a novel approach, which allows to reduce the memory, required to store intermediate data, exchanged between CNN operators. Unlike pruning and quantization approaches, our proposed approach preserves the CNN accuracy and reduces the CNN memory footprint at the cost of decreasing the CNN throughput. Rus, our approach is orthogonal to the pruning and quantization approaches, and can be combined with these approaches for further CNN memory footprint reduction.