Skip to content

🧬 Hyperdimensional Computing (HDC)

The spector-hdc module is an experimental module providing the first purpose-built SIMD-native HDC library for the JVM. Module: spector-hdc, Issue: #359, Status: Experimental.

Experimental

This module is under active development. APIs may change without notice.

What is HDC?

Hyperdimensional Computing (HDC) uses high-dimensional binary vectors (typically 10,000+ bits). In this space, randomly chosen vectors are nearly orthogonal. The core operations include: - Bind: XOR operation to combine vectors. - Bundle: Majority vote to aggregate vectors. - Permute: Cyclic shift to encode sequences.

These operations are based on the principles introduced by Pentti Kanerva, allowing robust representation and fast similarity computation.

Quick Start

import com.spectrayan.spector.hdc.*;

// Encode text to hypervectors
var encoder = new TextEncoder(10_000, 3);
Hypervector a = encoder.encode("the quick brown fox");
Hypervector b = encoder.encode("the fast brown fox");

// Compute similarity
double sim = HammingDistance.similarity(a, b);
System.out.println("Similarity: " + sim);

// High-level API
var hdc = new HdcSimilarity();
double score = hdc.similarity("hello world", "hello there");

API Reference

Class Description
Hypervector Represents a high-dimensional binary vector.
TextEncoder Encodes text strings into Hypervector objects.
HammingDistance Computes Hamming distance and similarity between vectors.
HdcSimilarity High-level API for comparing sequences.
VectorOperations Low-level operations like bind, bundle, permute.
MemoryPool Manages off-heap allocation for hypervectors.
BundleBuilder Assists in efficiently computing majority vote.
ShiftRegister Helps in generating n-grams via permute operations.

SIMD Architecture

The module utilizes Java's Vector API for maximum throughput:

  • LongVector.SPECIES_PREFERRED: Automatically selects the optimal vector shape for your hardware (e.g., AVX-512).
  • VectorOperators.BIT_COUNT: Maps directly to hardware popcount instructions like VPOPCNTDQ.
  • Masked tail pattern: Safely handles vector lengths that are not a multiple of the SIMD lane width without performance drops.
  • Off-heap Panama FFM: Integrates with the Foreign Function & Memory API for efficient native memory management, bypassing JVM GC overhead.
graph TD
    A[Input Text] --> B[Tokenizer]
    B --> C[Trigram Generation]
    C --> D[Permute & Bind]
    D --> E[Bundle Vectors]
    E --> F[Majority Threshold]
    F --> G[Final Hypervector]

Core Operations

Operation Symbol Implementation Purpose
Bind ⊗ Bitwise XOR Associates two hypervectors (e.g., Key-Value pair).
Bundle ⊕ Majority Vote (Add + Threshold) Aggregates multiple hypervectors into a set.
Permute ρ Cyclic Bit Shift Encodes order/sequence information.

Limitations

Limitations

  • Lexical, not semantic: HDC currently measures lexical overlap (like character n-grams) rather than deep semantic meaning.
  • Experimental: This is a labs feature and is not yet integrated into the core engine pipeline.