Binary Computing for Edge Intelligence

This PhD project will investigate the combination of two promising binary machine learning paradigms for future intelligent edge systems:

Binary Neural Networks (BNN) use 1-bit for data representation for both weights and activations, reducing the memory footprint, while also benefiting from using binary XNOR operations and pop-count as alternatives to the dense matrix multiplication operations. Consequently, BNN are prime candidates for including machine learning capabilities in resource-constrained devices, as they save both area and power consumption. [Sayed23, Yuan23]

Binary Hyperdimensional Computing (HDC) represents data as high-dimensional binary vectors (e.g. of dimension 10,000) called hypervectors, and uses binary XOR, majority sum, and Hamming distance as operations. HDC is used both for pattern recognition and for reasoning tasks., and its binary variants have hardware-level approaches that result in power efficient processing for edge devices. [Chang23]

This PhD track will combine BNN and HDC to realize neuro-symbolic AI systems with rich internal representations, aiming for more capable and flexible AI systems. 

[Sayed23] Sayed, et al, A Systematic Literature Review on Binary Neural Networks. IEEE Access. PP. 1-1. 10.1109/ACCESS.2023.3258360 (2023)
[Yuan23] Yuan, Chunyu, et al. "A comprehensive review of binary neural network." Artificial Intelligence Review (2023): 1-65
[Chang23] C. -Y. Chang, et al, "Recent Progress and Development of Hyperdimensional Computing (HDC) for Edge Intelligence," in IEEE Journal on Emerging and Selected Topics in Circuits and Systems, vol. 13, no. 1, pp. 119-136 (2023)
[Sheth23] Sheth, A., et al, Neurosymbolic ai-why, what, and how. arXiv preprint arXiv:2305.00813 (2023)