Reference #: 1747
The University of South Carolina is offering licensing opportunities for Online Training of KAN Autoencoder for Energy Efficient Channel Coding.
Background:
The AI-native radios using autoencoders along with traditional neural network structures (e.g., multi-layer perceptron (MLP)) to encode and decode information bits face limitations in computational power, memory, and battery life, as traditional neural networks often require a large number of parameters. Kolmogorov-Arnold Networks (KANs) with symbolic regression offer a solution by simplifying learned models into simple math expressions. This reduces computational demands and lowers energy consumption during operation. KANs can help preserve battery life and maintain performance.
Invention Description:
This innovation introduces an energy-efficient KAN autoencoder with symbolic regression for communication systems. Specifically, the transmitter and receiver convey information using a KAN encoder and a KAN decoder designed based on a non-linearity score, along with a pruning threshold and symbolic regressions. The resulting encoder and decoder introduce low-complexity structures and improve energy efficiency while preserving performance.
Potential Applications:
Telecommunications
Advantages and Benefits:
This invention improves energy efficiency compared to traditional MLP-based autoencoder implementations for channel coding or modulation. Our implementation enables new AEs for AI-native radios to preserve performance while consuming less energy.