November 29, 2024
On Thursday 28 November 2024, IIT Researchers Tatiana Boura and Dr Stasinos Konstantopoulos organised a talk at the NCSRD Library, titled “Kolmogorov-Arnold Networks: Theoretical Background, current approaches, and potential next steps” attended by more than 40 participants.
During the talk, they presented the Kolmogorov-Arnold Network (KAN), a new and promising approach for machine learning. More specifically, KAN shifts the objective of training from parameterising how fixed activation functions contribute to the next layer, to parameterising activation functions that are then fed to the next layer under simple summation.
They discussed the mathematical background on multivariate functions and in particular the Kolmogorov-Arnold Representation Theorem that underlies KANs. Furthermore, they discussed works that use the Representation Theorem and subsequent mathematical results as a tool for understanding deep learning and how it differs from the perceptron.
Finally, the talk focused on recent ML articles and the original KAN article was presented as well as relevant literature on its extensions and applications, before presenting and discussing future plans in this field, exploring opportunities to collaborate.
