PersonalizedHD: Hyperdimensional Online Learning with Scalable Personalization and Memory-Efficient Replay

Shinhyoung Jang, Junyoung Lee, Seohyun Kim, Hyunsei Lee, Ilhong Suh, Yeseong Kim

Abstract: Edge learning systems such as voice assistants and wearables must learn from few examples while adapting to diverse users under tight resource budgets. Conventional neural approaches rely on replay and gradient updates, leading to high latency and memory overhead. We present PersonalizedHD, a hyperdimensional computing framework for online FSCIL that unifies scalable personalization with compact memory replay. By binding task and user hypervectors and storing only binarized, bit-packed encodings, PersonalizedHD achieves accuracy comparable to deep neural networks while reducing parameters by up to $28 \times$, training time by $4-6 \times$, inference latency by $5.9-15.8 \times$, and memory usage by $58 \times$. These results demonstrate that hypervector-based representations provide a practical foundation for adaptive and efficient edge learning.

External IDs:dblp:conf/bigdataconf/JangLKLSK25