Fully-Binarized Distance Computation Based On-Device Few-Shot Learning for XR Applications

Vivek Parmar, Sandeep Kaur Kingra, Syed Shakib Sarwar, Ziyun Li, Barbara De Salvo, Manan Suri; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 4502-4508

Abstract


Low-Power Edge-AI capabilities are essential for on-device extended reality (XR) applications to support the vision of Metaverse. A critical requirement for emerging AI applications is personalization and adaptability without requiring retraining. Few-shot learning using embedding-based computations present an attractive method for the same. However, quantization-based optimizations to map such computations are yet to be explored. In this work, we present a fully binarized distance computing (BinDC) framework to perform distance computations for few-shot learning using only accumulation and logic operations (XOR/XNOR). The proposed method leads to marginal loss in accuracy of 4% (for 4-bits). This leads to savings in memory ( 8x), energy ( 2.5-3x), power ( 2x) and latency ( 1.1-1.5x) compared to a floating-point cosine distance computation when using CPU-based computations performed on an embedded platform. We further demonstrate realizations utilizing RRAM (resistive random access memory) based IMC (in-memory computing) to further improve EDP (energy delay product) ( 1000x) in comparison to the embedded CPU-based realization.

Related Material


[pdf]
[bibtex]
@InProceedings{Parmar_2023_CVPR, author = {Parmar, Vivek and Kingra, Sandeep Kaur and Sarwar, Syed Shakib and Li, Ziyun and De Salvo, Barbara and Suri, Manan}, title = {Fully-Binarized Distance Computation Based On-Device Few-Shot Learning for XR Applications}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2023}, pages = {4502-4508} }