Optimal hyperdimensional representation for learning and cognitive computation.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Hyperdimensional Computing (HDC) emulates brain functions using high-dimensional operations.
• Current HDC methods excel in learning (classification) and cognitive computation (reasoning) but lack unified encoding strategies.
• Existing approaches offer limited guidance on optimizing hyperdimensional representations for distinct learning and cognitive task requirements.

Purpose of the Study:

• To propose the first universal hyperdimensional encoding method adaptable to both learning and cognitive computation.
• To theoretically and empirically investigate the impact of representation correlation on learning and cognition.
• To provide a systematic framework for designing hyperdimensional encoders that unify these two domains.

Main Methods:

• Developed a neural-symbolic approach using complex hypervectors and algebraic operations in hyperspace.
• Controlled the correlation structure of encoded data points.
• Derived a separation metric to quantify the trade-off between correlation and orthogonality.
• Validated the method on image classification and decoding tasks.

Main Results:

• Learning tasks benefit from correlated representations for enhanced memorization and generalization.
• Cognitive tasks require orthogonal, highly separable representations for accurate decoding and reasoning.
• Tuning encoder correlation improved classification accuracy from 65% to 95%.
• Maximizing representation separation enhanced decoding accuracy from 85% to 100%.

Conclusions:

• The proposed universal hyperdimensional encoding method dynamically adapts to distinct task requirements.
• Representation correlation is key for learning, while separation is crucial for cognition.
• This work establishes a theoretically grounded framework for unified HDC encoders.