Robust noise-aware algorithm for randomized neural network and its convergence properties.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Randomized neural networks (RNNs), including random vector functional link networks (RVFL) and extreme learning machines (ELM), are efficient for single-hidden layer feedforward networks (SLFNs).
• RNNs offer strong approximation capabilities but can be unreliable in the presence of weight noise and data outliers.
• Existing RNN algorithms require enhancement for robustness in imperfect operational conditions.

Purpose of the Study:

• To develop a robust RNN algorithm that addresses the combined effects of weight noise and training data outliers.
• To introduce a novel objective function and optimization method for enhanced RVFL network performance.
• To extend the proposed algorithm for ensemble deep RVFL (edRVFL) networks and fault tolerance.

Main Methods:

• A new objective function is proposed to mitigate the impact of weight noise and outliers in RVFL networks.
• The half-quadratic optimization method is employed to develop the noise-aware RNN (NARNN) algorithm.
• Theoretical validation of NARNN convergence and its extension to ensemble and fault-tolerant configurations are discussed.

Main Results:

• The proposed NARNN algorithm effectively optimizes the objective function designed for noisy and outlier-prone data.
• The convergence properties of the NARNN algorithm are theoretically established.
• Experimental results show that NARNN outperforms existing state-of-the-art robust RNN algorithms.

Conclusions:

• The NARNN algorithm provides a robust solution for RVFL networks facing weight noise and data outliers.
• The NARNN framework is adaptable for ensemble deep RVFL networks and can be extended for fault tolerance.
• The proposed approach offers significant improvements in reliability and performance over current robust RNN methods.