Skip to main content

Glossary

Quick reference for terms used throughout Trinity documentation. If a term is missing, check the API Reference for module-specific definitions.

Balanced ternary​

Number system using {-1, 0, +1} instead of {0, 1, 2}. Negation is a simple sign flip. Truncation rounds to the nearest value automatically. See Ternary Computing Concepts.

Bind​

VSA operation that links two vectors via element-wise multiplication. The result is dissimilar to both inputs. Binding is its own inverse: bind(bind(a, b), b) = a. See VSA API.

BitNet b1.58​

Neural network architecture using ternary weights {-1, 0, +1}. Each weight uses approximately 1.58 bits of storage. Eliminates multiplication in matrix operations. See BitNet.

Bundle​

VSA operation that combines vectors via majority vote. The result is similar to all inputs. Used to store multiple items in a single vector. See VSA API.

Codebook​

A mapping from symbols (characters, words, labels) to hypervectors. Also called ItemMemory. Each symbol gets a unique random vector. See Sequence HDC.

Cosine similarity​

Measure of the angle between two vectors. Range: [-1, 1]. A value of +1 means identical, 0 means unrelated, and -1 means opposite. The primary similarity metric in Trinity's VSA.

Dense vector​

Vector that stores all elements explicitly, including zeros. Contrast with sparse vector. See HybridBigInt.

Dimension​

The number of elements (trits) in a vector. Typical range: 1000 to 10000. Higher dimensions give better noise tolerance and cleaner separation between unrelated vectors.

Dot product​

Sum of element-wise products of two vectors. Related to cosine similarity but without normalization by vector magnitudes.

Hamming distance​

Count of positions where two vectors differ. A distance of zero means the vectors are identical. Maximum distance equals the dimension.

HDC​

Hyperdimensional Computing. A computing framework that uses high-dimensional vectors for representation and reasoning. Equivalent to VSA. The two terms are used interchangeably.

HybridBigInt​

Trinity's main vector type. Supports dual-mode storage: packed mode for memory efficiency and unpacked mode for fast computation. Switches between modes automatically. See Hybrid API.

Hypervector​

A vector with thousands of dimensions, typically 1000 to 10000 trits. The high dimensionality ensures that random vectors are quasi-orthogonal.

JIT​

Just-In-Time compilation. Trinity's JIT engine generates native SIMD instructions at runtime for faster VSA operations. See JIT API.

Majority vote​

Decision rule used in bundling. For each position, take the most common value among the input vectors. Ties are broken randomly. This preserves the signal from each input.

N-gram​

A contiguous subsequence of N items. For text, these are usually characters. The word "hello" contains the trigrams (3-grams): "hel", "ell", "llo". Used in Sequence HDC for text encoding.

Packed mode​

Storage mode using approximately 1.58 bits per trit. Highly memory-efficient. Element access requires bit manipulation, making it slower than unpacked mode. Ideal for storage and transfer. See Hybrid API.

Permute​

Cyclic shift of vector elements. Shifts all elements by a given count, wrapping around at the boundary. The result is dissimilar to the original vector. Used to encode position or sequence order. See VSA API.

Quasi-orthogonal​

Two vectors with cosine similarity near zero. In high-dimensional spaces, random vectors are almost always quasi-orthogonal. This is the mathematical foundation that allows random vectors to represent distinct concepts.

Sparse vector​

Vector that stores only non-zero elements using coordinate (COO) format. Efficient when most elements are zero. Uses less memory than dense vectors for high-sparsity data. See Sparse API.

Trit​

A ternary digit. Takes the value -1, 0, or +1. Carries log2(3) = 1.58 bits of information. The fundamental unit of data in Trinity.

Unbind​

Reverse of bind. Recovers one vector from a binding given the other. Because binding uses element-wise multiplication and trits are self-inverse, unbinding is mathematically identical to binding. See VSA API.

Unpacked mode​

Storage mode using 8 bits (one byte) per trit. Allows fast element access with no bit manipulation. Uses 5x more memory than packed mode. Ideal for computation. See Hybrid API.

VSA​

Vector Symbolic Architecture. A framework for symbolic AI that uses high-dimensional vectors. Core operations are bind, bundle, and permute. Trinity implements VSA using balanced ternary vectors. See VSA API.