Learning Through Deterministic Assignment of Hidden Parameters.

Area of Science:

• Machine Learning
• Computational Mathematics

Background:

• Supervised learning involves estimating parameters in hypothesis spaces.
• Traditional one-stage learning (OSL) trains hidden and bright parameters simultaneously, leading to high computational costs.

Purpose of the Study:

• To propose a novel two-stage learning scheme, learning through deterministic assignment of hidden parameters (LtDaHPs).
• To reduce the computational burden associated with traditional supervised learning methods while maintaining generalization performance.

Main Methods:

• Deterministic assignment of hidden parameters using minimal Riesz energy points on a sphere and equally spaced points in an interval.
• Implementation of the LtDaHP scheme using neural networks.

Main Results:

• Theoretical analysis showing LtDaHP achieves generalization performance comparable to OSL.
• Simulations and application examples demonstrating the outperformance of LtDaHP over OSL in terms of computational efficiency.

Conclusions:

• LtDaHP offers an effective solution to mitigate the computational burden of OSL.
• The proposed method provides a practical approach for training complex models with improved efficiency.