SIM-ELM: Connecting the ELM model with similarity-function learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Science

Background:

• Extreme Learning Machines (ELMs) and similarity function learning are prominent randomization-based training schemes.
• Both paradigms utilize data remapping and linear separators but differ in randomization's role.

Purpose of the Study:

• To explore the affinities between ELMs and similarity function learning.
• To introduce an integrated approach that unifies these models.
• To propose a novel ELM variant with an analytical relationship between remapped space dimensionality and learning ability.

Main Methods:

• Comparative analysis of ELMs and similarity function learning frameworks.
• Development of an integrated learning scheme by connecting the two models.
• Analytical derivation of the relationship between remapped space dimensionality and predictor performance.

Main Results:

• A new variant of the Extreme Learning Machine (ELM) is presented.
• The proposed scheme establishes an analytical link between remapped space dimensionality and learning capabilities.
• Experimental validation demonstrates improved accuracy-complexity trade-offs compared to conventional ELMs.

Conclusions:

• The integrated approach yields a superior ELM variant.
• The novel learning scheme offers a better balance between classification accuracy and predictor complexity.
• This work advances the understanding and application of randomization in machine learning.