Studies of self-adjusting non-linear polynomial neural network for machine learning

View at publisher Summarise with VIRAT
Authors: P. R. Nikhil, Jibrael Jos
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The traditional architectures of neural networks depend on fixed activation functions to introduce non-linearity, which applies uniform transformations across all input features. This could not able to capture distinct non-linear relationships that may exist in between features and the output, limiting the adaptability and representational ability of the network. In this paper, we are discussing an architecture for Polynomial Neural Networks that introduce dynamic powers and weight adjustment for overcoming these limitations. In this architecture, learnable power-law transformation for each feature is used with dynamically adjusting weights based on input feature and complexity of the data, allow the model for feature-specific non-linear mappings also better adaptability to our model. The approach is tested on synthetic datasets and a dataset of heart failure prediction for a real-world scenario by comparing the PPNN with the traditional architectures that leveraging fixed activations for non-linearity. The network outperformed the traditional ANN in terms of predictive accuracy and representation learning. Key challenges such as extreme value management were tackled using layer normalization, while stabilizing gradient calculation with epsilon adjustments. However, this flexibility of architecture also introduces drawbacks such as increased risks of overfitting and increased computational complexity. These results show the importance of dynamic powers and weights for advancing neural network architectures, majorly for tasks that needs a complex feature relations and non-linear representations.

Fields of Study & Subjects

Neural Networks and Applications Machine Learning and ELM Advanced Neural Network Applications

Publication details

Journal / source
CRC Press eBooks
Pages
174–184
Publication date
25 March 2026
Language
EN
Document type
Book chapter