Choosing a VSA Model¶

HoloVec provides 7 different VSA models. This guide helps you choose the right model for your use case.

Quick Selection Guide¶

Quick Recommendations:

Most versatile: FHRRModel - Works everywhere, smooth encodings
Fastest: MAPModel - Binary, ultra-fast, lowest memory
Best for complex reasoning: FHRRModel or HRRModel
Best for sparse data: BSCModel or BSDCModel
Best for hardware: MAPModel or BSCModel (binary operations)

Decision Tree¶

Do you need exact similarity reversal (unbinding)?
├─ YES → Consider FHRR, HRR, GHRR, VTB
│   │
│   ├─ Need smooth encodings (continuous data)?
│   │   └─ YES → FHRR (best choice)
│   │
│   ├─ Working with symbolic data only?
│   │   └─ YES → HRR (classic, well-studied)
│   │
│   └─ Need advanced features?
│       ├─ Exact encoding/decoding → GHRR
│       └─ Variable binding strength → VTB
│
└─ NO → Consider MAP, BSC, BSDC
    │
    ├─ Need maximum speed?
    │   └─ YES → MAP (binary, ultra-fast)
    │
    ├─ Working with sparse data?
    │   └─ YES → BSC or BSDC
    │
    └─ Need hardware efficiency?
        └─ YES → MAP or BSC (binary operations)

Model Comparison Table¶

Feature	MAP	FHRR	HRR	BSC	BSDC	GHRR	VTB
Data Type	Binary	Complex	Real	Binary	Binary	Complex	Real
Unbinding	Approx	Exact	Exact	Approx	Approx	Exact	Exact
Speed	⭐⭐⭐⭐⭐	⭐⭐⭐	⭐⭐⭐	⭐⭐⭐⭐	⭐⭐⭐⭐	⭐⭐	⭐⭐
Memory	Lowest	Medium	Medium	Low	Low	Medium	Medium
Hardware	Best	Good	Good	Best	Best	Good	Good
FPE Support	No	Yes	Yes	No	No	Yes	Yes
Best For	Speed	General	Classic	Sparse	Sparse	Advanced	Research

Detailed Model Descriptions¶

MAP (Multiply-Add-Permute)¶

MAPModel

Type: Binary hypervectors

Key Properties:

Ultra-fast operations (binary XOR, shifts)
Lowest memory footprint (1 bit per dimension)
Hardware-friendly (FPGA, ASIC implementations)
Approximate unbinding

Best For:

Real-time applications
Edge devices with limited resources
Maximum speed requirements
Hardware implementations

Limitations:

Cannot use FractionalPowerEncoder
Less precise similarity measurements
Approximate rather than exact unbinding

Example Use Cases:

Gesture recognition on wearables
Real-time sensor fusion
Embedded systems
Mobile applications

See Also:

Demo: Basic VSA Operations - MAP examples
Models - Full API reference

FHRR (Fourier Holographic Reduced Representations)¶

FHRRModel

Type: Complex-valued hypervectors

Key Properties:

Exact unbinding via complex conjugate
Smooth similarity for continuous values
Supports FractionalPowerEncoder
Well-suited for numerical data

Best For:

Continuous data (temperature, time, coordinates)
Applications requiring exact reversal
General-purpose VSA tasks
Symbolic + numeric fusion

Limitations:

Higher memory than binary models
Complex arithmetic overhead
May need normalization

Example Use Cases:

Time series analysis
Sensor data processing
Multimodal data fusion
Scientific computing

See Also:

Fractional Power Encoder Deep Dive - FPE with FHRR
Gesture Recognition from Motion Trajectories - Trajectory encoding
Models - Full API reference

HRR (Holographic Reduced Representations)¶

HRRModel

Type: Real-valued hypervectors

Key Properties:

Classic model (well-studied since 1990s)
Circular convolution for binding
Exact unbinding via correlation
Supports FractionalPowerEncoder

Best For:

Classic VSA applications
Research with established baselines
Symbolic reasoning
Applications with existing HRR code

Limitations:

Slower than FHRR for some operations
Higher memory than binary models

Example Use Cases:

Natural language processing
Knowledge representation
Analogical reasoning
Cognitive modeling

See Also:

../examples/03_binding_and_unbinding - Binding examples
Models - Full API reference

BSC (Binary Spatter Codes)¶

BSCModel

Type: Binary hypervectors

Key Properties:

Very sparse representations
Fast binary operations
Low memory footprint
Hardware-friendly

Best For:

Sparse data
Fast similarity search
Large-scale storage
Hardware implementations

Limitations:

Cannot use FractionalPowerEncoder
Approximate unbinding
Requires careful capacity management

Example Use Cases:

Document retrieval
Large vocabulary applications
Recommendation systems
Content-based search

See Also:

Models - Full API reference

BSDC (Binary Spatter Codes with Discrete Cleanup)¶

BSDCModel

Type: Binary hypervectors with cleanup

Key Properties:

Enhanced BSC with cleanup mechanism
Better error correction
Improved similarity discrimination

Best For:

BSC applications requiring better accuracy
Noisy environments
Complex retrieval tasks

See Also:

Models - Full API reference

GHRR (Generalized Holographic Reduced Representations)¶

GHRRModel

Type: Complex-valued with advanced features

Key Properties:

Generalization of HRR/FHRR
Flexible encoding/decoding
Research-oriented features

Best For:

Advanced VSA research
Custom encoding schemes
Specialized applications

See Also:

Models - Full API reference

VTB (Variable Binding)¶

VTBModel

Type: Real-valued with variable binding strength

Key Properties:

Tunable binding strength
Research model for variable binding
Flexible unbinding

Best For:

Research on binding mechanisms
Applications requiring variable binding strength

See Also:

Models - Full API reference

Use Case Recommendations¶

By Application Domain¶

Text Classification / NLP

Primary: FHRRModel or HRRModel
Alternative: MAPModel for speed
Reason: Symbolic data with optional continuous features (word embeddings)
Example: Document Classification with N-grams

Image Recognition

Primary: FHRRModel
Alternative: MAPModel for edge devices
Reason: Spatial encoding benefits from smooth similarity
Example: Image Pattern Recognition

Gesture Recognition / Motion

Primary: FHRRModel
Alternative: MAPModel for wearables
Reason: Continuous trajectories, real-time requirements
Example: Gesture Recognition from Motion Trajectories

Time Series / Sensor Fusion

Primary: FHRRModel
Reason: Continuous temporal data, exact unbinding helpful
Example: Demonstration of Trajectory Encoder for continuous sequence encoding.

Symbolic Reasoning

Primary: HRRModel or FHRRModel
Reason: Classic applications, exact unbinding essential
Example: ../examples/16_compositional_structures

Large-Scale Retrieval

Primary: BSCModel or MAPModel
Reason: Fast similarity search, low memory
Example: ../examples/05_similarity_and_distance

By Hardware Platform¶

Desktop / Server (No Constraints)

Recommended: FHRRModel
Reason: Best general-purpose capabilities

Mobile / Edge Devices

Recommended: MAPModel
Reason: Lowest memory and compute requirements

FPGA / ASIC Implementation

Recommended: MAPModel or BSCModel
Reason: Binary operations map directly to hardware

GPU Acceleration

Recommended: FHRRModel with PyTorch backend
Reason: Complex operations parallelize well
See: Backends and Performance

By Performance Requirements¶

Maximum Speed

MAPModel - Fastest
BSCModel - Very fast
FHRRModel - Fast enough for most applications

Minimum Memory

MAPModel - 1 bit/dimension
BSCModel - 1 bit/dimension
FHRRModel - 16 bytes/dimension (complex float64)

Best Accuracy

FHRRModel - Smooth similarity, exact unbinding
HRRModel - Exact unbinding, well-studied
BSDCModel - Enhanced cleanup

Performance Trade-offs¶

Speed vs Accuracy¶

Model	Speed Characteristics	Accuracy Characteristics
MAP	Ultra-fast (binary XOR)	Approximate (discrete similarity)
BSC/BSDC	Very fast (binary ops)	Good (sparse codes)
FHRR	Fast (optimized FFT)	Excellent (smooth similarity)
HRR	Moderate (convolution)	Excellent (exact unbinding)
GHRR/VTB	Moderate to slow	Research models

Memory vs Capacity¶

Memory per dimension:

Binary (MAP, BSC, BSDC): 1 bit
Real (HRR, VTB): 8 bytes (float64)
Complex (FHRR, GHRR): 16 bytes (complex128)

Capacity (number of distinct items):

All models: ~dimension / 100 for good orthogonality
Example: 10,000 dimensions → ~100 items

Trade-off:

Higher dimensions → More capacity but more memory
Binary models → Less memory per dimension
Choose dimension based on application needs

Switching Models¶

Changing models is easy - just modify the create() call:

from holovec import VSA

# Try different models with same code
model = VSA.create('FHRR', dim=10000)  # Complex-valued
# model = VSA.create('MAP', dim=10000)   # Binary
# model = VSA.create('HRR', dim=10000)   # Real-valued

# Rest of code stays the same
a = model.random(seed=1)
b = model.random(seed=2)
c = model.bind(a, b)

Important: Some encoders only work with specific models:

FractionalPowerEncoder requires FHRR, HRR, GHRR, or VTB
ThermometerEncoder and LevelEncoder work with all models

See Encoding Data for encoder selection guidance.

Frequently Asked Questions¶

Q: Which model should I start with?

A: FHRRModel - It’s the most versatile and works for most applications.

Q: When should I use MAP instead of FHRR?

A: When speed and memory are critical (embedded systems, real-time applications, hardware implementations).

Q: Can I mix models in one application?

A: No - hypervectors from different models are incompatible. Choose one model for your entire application.

Q: How do I know if my model has enough capacity?

A: Rule of thumb: dimension / 100 items. Use ../examples/06_cleanup_and_retrieval to test retrieval accuracy.

Q: Does model choice affect encoding?

A: Yes - some encoders (like FractionalPowerEncoder) only work with certain models. See Encoding Data.

Q: Can I change models after development?

A: Switching is usually straightforward, but you’ll need to re-encode all data. Test thoroughly.

Q: Which model has the best theoretical foundations?

A: HRRModel (classic, well-studied) and FHRRModel (modern, mathematically elegant).

Next Steps¶

Encoding Data - Learn about encoding different data types
Backends and Performance - Choose the right backend for performance
HoloVec Examples - See models in action
Models - Complete model API reference
Theory Guide: Hyperdimensional Computing & Vector Symbolic Architectures - Theoretical foundations

Choosing a VSA Model¶

Quick Selection Guide¶

Decision Tree¶

Model Comparison Table¶

Detailed Model Descriptions¶

MAP (Multiply-Add-Permute)¶

FHRR (Fourier Holographic Reduced Representations)¶

HRR (Holographic Reduced Representations)¶

BSC (Binary Spatter Codes)¶

BSDC (Binary Spatter Codes with Discrete Cleanup)¶

GHRR (Generalized Holographic Reduced Representations)¶

VTB (Variable Binding)¶

Use Case Recommendations¶

By Application Domain¶

By Hardware Platform¶

By Performance Requirements¶

Performance Trade-offs¶

Speed vs Accuracy¶

Memory vs Capacity¶

Switching Models¶

Frequently Asked Questions¶

Next Steps¶

See Also¶