一种用于加速神经视觉识别的硬件架构

A Hardware Architecture for Accelerating Neuromorphic Visual Recognition

深度学习的广泛应用带来了视觉分析中许多类似人类认知任务的实现。HMAX是基于视觉皮层的生物启发模型,已在多类物体识别中被证明优于标准计算机视觉方法。但是,由于神经形态算法的高复杂性,在边缘设备上实现HMAX模型仍然面临巨大挑战。已有研究表明,HMAX的S2阶段是运行最耗时的阶段。该文提出了一种基于脉动阵列的新架构来加速HAMX模型的S2阶段。仿真结果表明,与基准模型相比,HMAX模型最耗时的S2阶段执行时间平均减少了14.65%、内存所需的带宽减少了3.34倍。

The widespread application of deep learning has led to the realization of many human-like cognitive tasks in visual analysis.HMAX is a visual cortex-based bio-inspired model that has proven superior to standard computer vision methods in multi-class object recognition.However,due to the high complexity of neural morphology algorithms,implementing HMAX models on edge devices still faces significant challenges.Previous experimental results show that the S2 phase of HMAX is the most time-consuming stage.In this paper,we propose a novel systolic array-based architecture to accelerate the S2 phase of the HAMX model.The simulation results show that compared with the baseline model,the execution time of the most time-consuming S2 phase of the HMAX model is reduced by 14.65%on average,and the required memory bandwidth is reduced by a factor of 3.34 X.