Fisher information matrix of unipolar activation function-based multilayer perceptrons

• Main Authors: Guo, Weili, Ong, Yew-Soon, Zhou, Yingjiang, Hervas, Jaime Rubio, Song, Aiguo, Wei, Haikun
• Other Authors: School of Computer Science and Engineering
• Format: Article
• Language: English
• Published: 2020
• Subjects: Engineering::Computer science and engineering; Analytical Form; Fisher Information Matrix (FIM)
• Online Access: https://hdl.handle.net/10356/139916
• Institution: Nanyang Technological University

The multilayer perceptrons (MLPs) are widely used in many fields, however, singularities in the parameter space may seriously influence the learning dynamics of MLPs and cause strange learning behaviors. Given that the singularities are the subspaces of the parameter space where the Fisher informati...