Extreme Learning Machine With Affine Transformation Inputs in an Activation Function.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Neural Networks

Background:

• Extreme Learning Machine (ELM) offers fast learning and good generalization.
• A key limitation of ELM is the non-uniform distribution of hidden node outputs due to random parameter generation, impacting performance.
• This non-uniformity can lead to suboptimal generalization capabilities.

Purpose of the Study:

• To introduce a novel activation function with affine transformation (AT) to address ELM's generalization issues.
• To develop an improved ELM algorithm, termed AT-ELM, that enhances hidden layer output distribution.
• To evaluate the performance and robustness of the proposed AT-ELM algorithm.

Main Methods:

• Incorporated a novel activation function with an affine transformation (AT) into the ELM architecture.
• Calculated AT scaling and translation parameters using the maximum entropy principle for uniform hidden layer outputs.
• Applied the AT-ELM to nonlinear function regression, real-world data classification, and benchmark image recognition tasks.

Main Results:

• The AT-ELM algorithm demonstrated robustness to input range scaling in nonlinear function regression.
• Experiments showed AT-ELM outperformed original ELM, regularized ELM, and kernel ELM in regression, classification, and image recognition.
• AT-ELM achieved superior performance compared to several state-of-the-art algorithms on benchmark image datasets.

Conclusions:

• The proposed AT-ELM effectively addresses the non-uniform hidden node output issue in ELM.
• AT-ELM offers improved generalization performance and robustness across diverse machine learning applications.
• The AT-ELM represents a significant advancement over existing ELM variants and other contemporary algorithms.