- The paper introduces hdlib 2.0, a unified toolkit that integrates supervised, unsupervised, regression, graph-based, and quantum ML methods within the VSA framework.
- It refines feature selection and hyperparameter optimization with a novel regression module and a VSA-based k-means clustering approach, improving interpretability and performance.
- The framework pioneers quantum hyperdimensional computing by integrating IBM’s Qiskit, enabling hybrid quantum-classical operations to explore quantum advantages in ML.
Comprehensive Expansion of Machine Learning in Vector-Symbolic Architectures with hdlib 2.0
Introduction
The presented work introduces hdlib 2.0, a major update to the Python-based Vector-Symbolic Architecture (VSA) library, with the aim of providing a unified and extensible toolkit for advanced ML within the VSA paradigm. Previous efforts established a foundation for high-dimensional information encoding, yet lacked integrated support for regression, unsupervised learning, graph-based learning, and quantum extensions. hdlib 2.0 directly addresses these gaps, augmenting the utility of VSA for data-driven modeling across scientific and engineering domains.
Enhanced Classification and Feature Selection
The update brings critical advancements to classification by overhauling the core model architecture. Notably, the new feature selection framework refines the stepwise regression method, offering enhanced user control, increased selection precision, and detailed importance rankings for input features. This enables VSA-based classifiers to support model interpretability at a level previously unattainable in this context. Additionally, the expanded hyperparameter optimization via an updated auto-tune function supports robust parameter sweeps and more exhaustive exploration of the hyperparameter space, ultimately resulting in improved model generalizability and empirical performance.
Regression and Clustering Models
hdlib 2.0 incorporates a regression module grounded in recent VSA-based regression algorithms. The regression encoder nonlinearly embeds input data into high-dimensional space using random hypervectors and biasing, mapping similarity structure from the original domain. The regression model utilizes an ensemble of cluster-specific regression submodels, whose outputs are combined through softmax-derived confidence scores. Model parameters are iteratively refined based on prediction error feedback, enabling robust learning even in presence of highly nonlinear dependencies. The module’s support for binary and quantized prediction accelerates inference and reduces resource consumption, broadening the range of deployable VSA regression applications.
For unsupervised learning, a clustering model is introduced that closely follows the VSA-based k-means formalism. Both cluster centers and datapoints are encoded as hypervectors, and similarity is assessed using cosine distance in high-dimensional space. Centroid updates leverage VSA’s bundling operator, aligning the method with the underlying mathematical properties of high-dimensional vector addition. The procedure guarantees convergence within the VSA framework, extending traditional k-means to the domain of brain-inspired computing.
Graph-Based Machine Learning
A major extension is the inclusion of a graph model capable of encoding complete weighted graphs—directed or undirected—into single hyperdimensional representations. Each node is assigned a distinct hypervector, and neighborhood structure is aggregated by binding neighbor node hypervectors with corresponding edge-weight hypervectors, followed by VSA-native bundling. This schema supports both efficient existence queries (edge discovery) and edge weight prediction through direct vector operations. The model is complemented by an error mitigation routine for iterative refinement, closing the gap between memory-based encoding and robust predictive performance. The atomic representation of an entire graph within a single high-dimensional vector supports scalability and efficient manipulation of large structured datasets, as demonstrated in recent domain applications.
Quantum Hyperdimensional Computing (QHDC)
hdlib 2.0 pioneers Quantum Hyperdimensional Computing (QHDC), integrating IBM’s Qiskit for quantum-native arithmetic. The framework leverages phase encoding for mapping hypervectors, quantum phase oracles for binding, quantum-native bundling via Linear Combination of Unitaries and Oblivious Amplitude Amplification, and permutation via Quantum Fourier Transform. Similarity is assessed using the Hadamard test to estimate inner products between quantum states. The inclusion of a quantum classification model introduces a novel approach for Quantum Machine Learning, with all core VSA operations now available in both classical and quantum computational modalities. This dual implementation provides a foundation for the exploration of quantum advantage in similarity-based ML and neuromorphic quantum architectures.
Implications and Future Prospects
The integration of regression, clustering, and graph encodings within a unified VSA framework constitutes a significant methodological advancement, consolidating previously disparate approaches into a single, easily accessible library. The quantum extensions position hdlib as the first open-source platform supporting both classical and quantum VSA computation. This fosters rapid experimentation at the intersection of ML, hardware-aware computation, and quantum information processing.
Practically, hdlib 2.0 enables rapid prototyping, rigorous feature selection, and deployment of VSA-based ML in life sciences, physical sciences, and embedded systems. The quantum module unlocks exploration of hybrid quantum-classical workflows, making the architecture relevant for quantum machine learning research and potential NISQ-era devices. The formalism and implementation adopted in hdlib 2.0 pave the way for scalable, interpretable, and hardware-efficient ML, and lay groundwork for future research in neuro-inspired and quantum-accelerated learning systems.
Conclusion
hdlib 2.0 represents a comprehensive evolution of VSA-based machine learning, through the integration of robust supervised, unsupervised, regression, and graph-based models, alongside the inception of quantum-native computation. The modular architecture, improved automated methods, and support for quantum backends establish hdlib as a practical and theoretically grounded toolkit, with direct implications for research spanning from classical ML to quantum neuromorphic computing. The framework stands to facilitate further advances in interpretable, scalable, and hardware-compatible machine learning pipelines.