A practical introduction
Today's AI systems are powerful, but their learned representations can be difficult to inspect and expensive to adapt. Hyperdimensional computing (HDC) gives us another approach: encode data as very high-dimensional hypervectors that we can compose, compare and update using simple operations. This lets us search by association, preserve explicit structure and learn online from a small number of examples.
[−0.20, +1.00, +0.20, +0.20, −0.60, +0.60, +0.20, −0.60, …]
high-dimensional vector HVX ∈ ℝ10,000 typically 10,000 dimensions music + hiking → similar people prototype ← prototype ⊕ sample What is HDC?
Instead of starting with a learned embedding for each whole item, HDC lets us build a representation from the structure already present in the data. Those representations share one hypervector space, so we can compose them, compare them and update them using a simple algebra.
Strings, numbers, lists and relationships can all be encoded as fixed-width hypervectors. We can then combine heterogeneous data without forcing every field into the same raw format.
E(string | number | list | edge) → hv A query can act as a cue rather than an exact identifier. Results rise when more of their encoded roles and values agree with the structure in that cue.
rank by cos(query_hv, item_hv) We can bundle a small number of labeled examples into prototypes and update them online. The encoder stays fixed while each new sample changes what the prototype "remembers".
prototype ← prototype ⊕ sample_hv The core algebra uses simple, composable operations over thousands of values. Information is distributed across the hypervector, so partial noise usually changes similarity gradually instead of erasing the entire representation.
binding · bundling · permutation How it works
Understand HDC fundamentals by learning how to encode entity metadata and the relationships between entities. The same hypervectors can be used for associative search, and a form of online learning that resembles building memory from experience.
Explore the conceptsOur approach
At HDC Labs, we start with the representation. A lot of the work in HDC lies in building the right encoder for the data at hand, so we test that representation on realistic downstream tasks and compare it with the baselines that already exist.
We're especially interested in problems involving connected or heterogeneous data, associative search, agent memory and systems that need to learn as new examples arrive.
We map the structure already present in the data: its roles, values, relationships and useful notions of similarity. That gives us a concrete encoder to test against an existing baseline, rather than a generic HDC layer.
We compare the HDC representation with conventional retrieval and machine learning approaches on real data. The goal is to understand where HDC helps, where it doesn't and how it compares.
We evaluate the representation through the application or task it needs to support, such as associative search, connected-data queries, agent memory or online learning.
We actively write about HDC on our blog, grounding our claims in evidence and explaining the general methods behind our approach. We aim to make results inspectable and interpretable, rather than presenting them as a black box.
Blog
At HDC Labs, we publish blog posts as we discover new insights. If you want to know how HDC works in practice, start here.
How Hyperdimensional Computing distributes information across an entire hypervector, helping representations degrade gradually under noise.
ReadA closer look at the high-dimensional, distributed mathematical objects that give hyperdimensional computing its unusual geometry.
ReadA gentle introduction to how hyperdimensional computing represents, combines, and retrieves information using high-dimensional hypervectors.
ReadContact
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.
We usually reply within two business days.
