Content area
AI has made remarkable strides with the availability of large datasets and the growth in computational power brought by large-scale cloud servers. Despite this, significant challenges remain in democratizing AI: (1) sky-rocketing computational cost, (2) more significant gaps at the edge between algorithm demands and hardware capabilities, and (3) the lack of human-level reasoning and explainability. Meanwhile, the widening usage of AI and machine learning technologies in critical areas raises attention in model efficiency, transparency, and advanced reasoning. This dissertation tackles such challenges with a focus on designing brain-inspired machine learning and reasoning algorithms and their efficient acceleration solutions.
Specifically, we propose a spectrum of neuro-symbolic learning and reasoning frameworks that are capable of human-like adaptability, transparency, and efficient decision-making. Our algorithms center around the Vector Symbolic Architecture (VSA), which integrates flexible distributed representation with composable and structured symbolic representation while providing opportunities for edge-optimized implementations through co-designing algorithms and hardware platforms. We redesign deep learning algorithms with a unique set of high-dimensional vector-based algebra, achieving benefits including: (I) wall-clock runtime saving in both learning and inference, (II) orders of magnitude improvement in energy consumption especially on edge platforms, and (III) error-robust and transparent symbolic representation.