以GHA實現快速主成分分析之硬體設計

Abstract

本論文為實現快速主成分分析之硬體，提出三種GHA的硬體架構，分別為達成高速計算及最少的硬體資源消耗為目的。在高速計算的架構，所有主成分值計算與突觸權重值之更新，皆使用專屬的電路作並行之運算。對於高維度訓練資料之適用架構，以所有主成分值計算共用一個主成分計算電路輪流完成計算，並將訓練資料區塊化方式逐步更新每個神經元的突觸權重值。所有實現的硬體架構訓練取得之權重向量，應用在紋理的分類。This dissertation proposes three types of GHA architectures, achieving the aims of high speed computation and low area costs, for fast principal component analysis. In the architecture with high speed computation, all principal components computation and synaptic weight vectors updating are operated concurrently in individual dedicated circuits. In the architectures with high-dimensional training vector, all principal components can be sequentially computed by a single circuit. Meanwhile, each synaptic weight vector is updated block by block in fixed number of synaptic weight updating modules. To demonstrate the effectiveness of the proposed architectures, texture classification will be adopted.