基于分组量化的高效超维计算分类方法

Efficient Hyperdimensional Computing Classification Method Based on Grouping Quantization

针对当前超维计算(hyperdimensional computing,HD)中大多数方法计算量大、效率低的问题,提出了一种基于分组量化的高效超维计算分类方法,在保证准确性的情况下提高HD模型的计算效率.该方法首先使用点积操作替代余弦相似度运算来降低HD计算推理阶段的计算量;其次,考虑到查询超向量的相似度计算随着类数的增加而增加,设计了一个分组查询方案,通过检查类的子集来减少相似度计算;最后,使用双值2次幂的量化方式来消除推理阶段的乘法运算,进一步提高计算速度.实验结果表明,与其他HD计算模型相比,所提方法性能优良,在相同的精度水平下,明显降低了能耗和执行时间.

Aiming at the problem of large amount of calculation and low efficiency of most methods used in current hyperdimensional computing(HD),an efficient hyperdimensional computing classification method based on grouping quantization was proposed,which could improve the computational efficiency of the HD model while ensuring the accuracy of calculation.Firstly,the method used dot product operation instead of cosine similarity operation to reduce the amount of calculation in HD computing reasoning stage.Secondly,considering the similarity calculation of query super vector increasing with the increase of the number of classes,a grouping query scheme was designed to reduce the similarity calculation by checking the subset of classes.Finally,the double value 2-power quantization method was used to eliminate the multiplication in reasoning stage,and further improved the calculation speed.Experimental results show that compared with other HD computing models,the proposed method achieved the best performance and significant reduction in the energy consumption and execution time at the same accuracy level.