基於記憶體內運算平台以重定位以及點雲密集化提供即時三維感知

Abstract

隨著物聯網邊緣設備和智慧型感測器技術的發展，具備 3D 傳感器邊緣運運算裝置，相較於過去無需仰傳統賴集中式叢集計算機系統即可執行神經網絡推理任務。而可以用於加速邊緣運算裝置的神經網絡推理任務的記憶體內計算（CIM）架構近年在研究領域中受到許多關注，配合3D感測器可快速地感知周圍環境豐富的3D幾何資訊。基於高速揮發性 SRAM CIM 運算單元的記憶體內運算異質系統，可用以進行快速三維感知工作。此系統可用來進行即時三維物件的特徵粗定位工作（<1s），並且可有效的減少傳統馮紐曼系統中神經網絡乘加工作於定位任務中處理單元的系統負載。

對於廣泛部署於邊緣嵌入式系統上的三維感知工作，由於記憶體計算單元內的低精度（~8 bit）乘加運算操作與記憶體內運算中以類比為基礎電荷共享計算方式於ADC轉換過程所產生的非線性不確定性誤差，此兩因素降低基於結構光系統感知所生成的稀疏且具雜訊點雲對於系統神經網絡推理工作的精確度。

在此研究中，所提出的神經網絡訓練方法，在訓練過程中嵌入不確定性模型的神經網絡訓練，作為增強於推理期間記憶體內運算中容忍ADC非線性不確定性錯誤變化以及三維雜訊感知資料之方法。同時將具有乘法與累加 (MAC) 運算的 SRAM CIM 硬體用於計算 k 個最近的高維特徵向量之間的歐幾里得距離之推論加速。並配合以注意機制為基礎的三維特徵抽取與理解的卷積神經網路定位設計，在實際實驗驗證中降低了 27×模型大小、297×推論時間以及100×系統能耗。若增加 CIM 單元於記憶體內之數量，可以進一步提高大型智慧型感知系統定位以及緻密化任務之效能表現。

IoT edge devices with 3D sensors perform neural network inference tasks without relaying a centralized computing system. 3D sensors can precept and provide rich 3D geometric information about the environment quickly and accurately. The architecture of computing-in-memory (CIM) to accelerate neural network inference tasks on edge devices has recently gained attention in the research community. Combining a high-speed SRAM CIM unit and fast 3D sensing on the sensor system provides a coarse processing function to reduce neural network loading on sensor localization tasks.

To wildly deploy 3D sensors on edge devices, the low precision (~8 bits) computing scheme and uncertainty error caused by the analog charge-sharing computing scheme of the ADC converter in an in-memory computing unit are the two main issues for system design. Severely degrades the accuracy during neural network inference under the sparsity and noisy 3D sensing of structure light system.

In this work, the proposed neural network training scheme embedded an uncertainty model enhancing the tolerance of inquiry variation and noise during sensing. The SRAM CIM macro for multiplication-accumulation (MAC) operation accelerates the task of Euclidean distance between the k nearest high-dimension feature vector, with the designed convolution base attention layer for feature realization, which reduces up to 27× model size, 100× power consumption, and 297× inference time in our experiment. Adding augmented abundant instances of CIM units can further improve the performance of the neural network task on intelligent sensing systems.