Hyperdimensional Computing (HDC), as a novel neurally-inspired computing methodology, uses lightweight and high-dimensional operations to realize major brain functionalities. Recent HDC works mainly focus on two aspects: brain-like learning and cognitive computation. However, it lacks differentiation between these functions and their requirements for HDC algorithms. We address this gap by proposing an adaptable hyperdimensional kernel-based encoding method. We explore how encoding settings impact HDC performance for both tasks, highlighting the distinction between learning patterns and retrieving information. We provide detailed guidance on kernel design, optimizing data points for accurate decoding or correlated learning. Experimental results with our proposed encoder significantly boost image classification accuracy from 65% to 95% by considering pixel correlations and increase decoding accuracy from 85% to 100% by maximizing pixel vector separation. Factorization tasks are shown to require highly exclusive representation to enable accurate convergence.
Keywords: hyperdimensional computing; vector symbolic architecture; decoding;
Abstract:
Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12307
Loading