holovec 0.3.2

Hyperdimensional Computing / Vector Symbolic Architectures library

Vector Symbolic Architectures for compositional, high-dimensional computing.

Documentation • Installation • Quick Start • Examples

---

What is HoloVec?

HoloVec is a Python library for hyperdimensional computing (HDC) and Vector Symbolic Architectures (VSA). It represents data as high-dimensional vectors (~1,000-10,000 dimensions) that can be composed using algebraic operations—binding creates associations, bundling creates sets, and permutation encodes order.

This approach enables one-shot learning, noise-tolerant representations, and transparent symbolic reasoning without gradient descent.

---

Installation

pip install holovec

Optional extras:

pip install "holovec[torch]"   # PyTorch backend
pip install "holovec[jax]"     # JAX backend
pip install "holovec[all]"     # all optional extras

For source and development work:

git clone https://github.com/Twistient/HoloVec.git
cd HoloVec
uv sync --extra dev

For an editable install outside the project workflow, use uv pip install -e .[dev]. For GPU support add .[torch], for JIT compilation add .[jax], or use .[all] for everything.

---

Quick Start

from holovec import VSA

# Create a model
model = VSA.create('FHRR', dim=2048)

# Generate random hypervectors
a = model.random()
b = model.random()

# Bind them together (creates an association)
c = model.bind(a, b)

# Unbind to recover the original
a_recovered = model.unbind(c, b)
print(model.similarity(a, a_recovered))  # 1.0 (exact recovery)

---

Concepts

HoloVec has four layers: Backends provide numerical primitives, Spaces define vector types and similarity measures, Models implement algebraic operations, and Encoders transform data into hypervectors.

Operations

Every VSA model supports three core operations:

Operation	What it does	Use case
Bind	Associates two vectors → result dissimilar to both	Key-value pairs, role-filler structures
Bundle	Superposes vectors → result similar to all inputs	Sets, prototypes, averaging
Permute	Shifts coordinates → preserves similarity	Sequences, positions, temporal order

Models

Different models have different algebraic properties:

Model	Binding method	Inverse	Notes
FHRR	Complex multiply	Exact	Best capacity, continuous data
GHRR	Matrix product	Exact	Non-commutative, matrix-based extension
MAP	Element multiply	Self	Hardware-friendly, neuromorphic
HRR	Circular convolution	Approximate	Classic baseline
VTB	Matrix transform	Approximate	Non-commutative, directional
BSC	XOR	Self	Binary, FPGA-friendly
BSDC	Sparse XOR	Approximate	Memory efficient
BSDC-SEG	Segment XOR	Self	Segment-sparse, fast search

Choose FHRR for general use, GHRR when order matters, MAP for hardware-oriented use, or BSDC variants for memory constraints.

Spaces

Each model operates on a specific vector space that determines how random vectors are generated and how similarity is measured:

Space	Values	Similarity	Used by
Complex	Unit phasors (e^iθ)	Cosine	FHRR
Bipolar	{-1, +1}	Cosine	MAP, HRR
Binary	{0, 1}	Hamming	BSC
Sparse	Sparse binary	Overlap	BSDC
Matrix	Unitary matrices	Trace	GHRR, VTB

You typically don't interact with spaces directly—VSA.create() selects the appropriate space for each model.

Backends

HoloVec runs on NumPy (default), PyTorch (GPU), or JAX (JIT). The API is identical:

# NumPy (default)
model = VSA.create('FHRR', dim=2048)

# PyTorch with GPU
model = VSA.create('FHRR', dim=2048, backend='torch', device='cuda')

# JAX with JIT compilation
model = VSA.create('FHRR', dim=2048, backend='jax')

NumPy is the release-blocking backend. PyTorch and JAX remain optional backends and should be treated as conditional support paths unless their backend-specific tests are enabled in your environment.

Encoders

Encoders transform data into hypervectors:

Encoder	Input	Preserves
FractionalPowerEncoder	Scalars	Locality (similar values → similar vectors)
ThermometerEncoder	Ordinals	Order relationships
PositionBindingEncoder	Sequences	Position information
NGramEncoder	Text	Local context
ImageEncoder	2D grids	Spatial relationships
VectorEncoder	Feature vectors	Per-dimension encoding

---

Examples

Role-Filler Binding

Represent structured knowledge like "the ball is red":

model = VSA.create('FHRR', dim=2048)

# Create role and filler vectors
color = model.random(seed=1)
red = model.random(seed=2)
ball = model.random(seed=3)

# Bind role to filler, bundle into a structure
ball_repr = model.bundle([
    model.bind(color, red),
    model.bind(model.random(seed=4), ball)  # object role
])

# Query: what color?
query = model.unbind(ball_repr, color)
print(model.similarity(query, red))  # High similarity

Encoding Continuous Values

from holovec.encoders import FractionalPowerEncoder

model = VSA.create('FHRR', dim=2048)
encoder = FractionalPowerEncoder(model, min_val=0, max_val=100)

# Similar values produce similar vectors
v50 = encoder.encode(50)
v51 = encoder.encode(51)
v90 = encoder.encode(90)

print(model.similarity(v50, v51))  # ~0.99
print(model.similarity(v50, v90))  # ~0.70

Cleanup and Retrieval

from holovec.retrieval import Codebook, ItemStore

model = VSA.create('FHRR', dim=2048)

# Create a codebook of known items
cb = Codebook({
    "apple": model.random(seed=1),
    "banana": model.random(seed=2),
    "cherry": model.random(seed=3)
}, backend=model.backend)

# Query with a noisy vector
store = ItemStore(model).fit(cb)
noisy_apple = cb["apple"] + model.random() * 0.1
result = store.query(noisy_apple, k=1)
print(result)  # [('apple', ~0.99)]

See examples/ for more.

---

Project Structure

holovec/
├── backends/     NumPy, PyTorch, JAX implementations
├── spaces/       Vector spaces (Bipolar, Complex, Sparse, etc.)
├── models/       VSA models (MAP, FHRR, HRR, BSC, GHRR, VTB, etc.)
├── encoders/     Scalar, sequence, spatial, structured encoders
├── retrieval/    Codebook, ItemStore, cleanup strategies
└── utils/        Search, similarity, helper functions

---

Testing

uv run --extra dev pytest
uv run --extra dev pytest --cov=holovec --cov-report=html
uv run --extra dev ruff check holovec tests
uv run --extra dev mypy holovec

The engine CI enforces tests, Ruff, and source-only mypy on every push and pull request.

---

References

HoloVec implements ideas from:

• Kanerva (1988) — Sparse Distributed Memory
• Plate (2003) — Holographic Reduced Representations
• Gayler (2003) — Vector Symbolic Architectures
• Frady et al. (2021) — Fractional power encoding
• Schlegel et al. (2022) — VSA model comparison
• Kleyko et al. (2023) — HDC/VSA survey
• Kymn et al. (2024) — Resonator networks

---

Citation

@software{HoloVec2025,
  author       = {Brodie Schroeder},
  title        = {HoloVec: Vector Symbolic Architectures for Python},
  year         = {2025},
  version      = {0.3.2},
  url          = {https://github.com/Twistient/HoloVec},
  license      = {Apache-2.0}
}

---

Contributing

See CONTRIBUTING.md. Key areas: new models, encoders, documentation, and performance.

License

Apache 2.0. See LICENSE.

Contact

GitHub Issues • Discussions • [e-mail]