Extreme Kernel Machine
Viktor Karlsson, Erik Rosvall · KTH Publication Database DiVA (KTH Royal Institute of Technology) · 2017
The purpose of this report is to examine the combinationof an Extreme Learning Machine (ELM) with the KernelMethod. Kernels lies at the core of Support Vector Machines successin classifying non-linearly separable datasets. The hypothesisis that by combining ELM with a kernel we will utilize featuresin the ELM-space otherwise unused. The report is intended asa proof of concept for the idea of using kernel methods in anELM setting. This will be done by running the new algorithmagainst five image datasets for a classification accuracy and timecomplexity analysis.Results show that our extended ELM algorithm, which we havenamed Extreme Kernel Machine (EKM), improve classificationaccuracy for some datasets compared to the regularised ELM,in the best scenarios around three percentage points. We foundthat the choice of kernel type and parameter values had greateffect on the classification performance. The implementation ofthe kernel does however add computational complexity, but wherethat is not a concern EKM does have an advantage. This tradeoffmight give EKM a place between other neural networks andregular ELMs.