Learning neural networks with noisy inputs using the errors-in-variables approach.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Current neural network (NN) learning predominantly uses output error with least squares, effective for known inputs and noisy outputs.
• Training NNs with noisy input data or when both inputs and outputs are noisy presents significant challenges for standard methods.

Purpose of the Study:

• To propose a novel cost function for neural network learning specifically designed to address noisy input data.
• To introduce a new learning scheme based on the errors-in-variables stochastic framework for improved NN training.

Main Methods:

• Development of a new cost function rooted in the errors-in-variables stochastic framework.
• Implementation of a corresponding learning scheme for neural network model training.
• Application and evaluation of the proposed method in neural network curve fitting scenarios.

Main Results:

• Demonstrated improved performance in neural network curve fitting when using the novel cost function with noisy inputs.
• The proposed method shows enhanced robustness compared to traditional output error approaches under noisy input conditions.

Conclusions:

• The novel cost function offers a viable solution for training neural networks with noisy input data.
• The errors-in-variables approach enhances NN performance in curve fitting, albeit with increased computational demands.