Reviewing TrainableHD: What changes when the encoder learns?
How TrainableHD learns its encoder as well as its classes. We follow a mistake through training, work through an incident example, and examine the costs of deployment.
Posts by David Hughes
Showing 1–10 of 14 posts
Durable memory, transient prototypes
An HDC experiment that separates stored examples from temporary class summaries, and compares rebuilding with caching.
Find the case that fits now
The most similar past incident may not be the most useful. What if search could take today's equipment, conditions, and resources into account?
How context can change a retrieval ranking
Why shared bipolar binding preserves a ranking, how weighted similarity can change it, and what remains to be tested for useful retrieval.
Semantic compression inside wide hypervectors
How a compact abstraction can use thousands of coordinates, and what the extra width buys an HDC system.
When AI remembers what happened after the answer
A warehouse robot's detour shows how connected evidence and HDC memory could help AI reuse experience while checking what is valid now.
A category is something a system does
What a new neuroscience framework proposes about context, compression, and action, and how HDC could turn those ideas into a computational architecture.
Does HDC fit the physics of light? Reviewing PhotoHDC
A critical review of PhotoHDC's modeled electro-photonic accelerator, what its evidence supports, and how OLIX and recent optical HDC work sharpen the deployment question.
Can a compiler spend HDC's accuracy budget? Reviewing ApproxHDC
How ApproxHDC searches application, compiler, and simulated-hardware choices to trade HDC task quality for speed and energy efficiency, and what the reported evidence establishes.
Can hypervectors learn the hardware they run on?
How learnable HDC encoders can adapt similarity geometry to modeled compute-in-memory nonlinearities, and what the current evidence establishes.
Contact
Have a representation problem in mind?
If you're working with connected data, retrieval, agent memory or online learning, we'd love to hear what you're building and where the current representation is falling short.
Get in touch
We usually reply within two business days.