iKAN: Global Incremental Learning with KAN for Human Activity Recognition Across Heterogeneous Datasets

TL;DR

The paper tackles incremental learning for human activity recognition across heterogeneous sensor datasets, addressing catastrophic forgetting and non-uniform inputs. It introduces iKAN, which substitutes the standard MLP classifier with Kolmogorov-Arnold Networks and employs task-specific feature branches plus a feature redistribution layer to align diverse modalities while keeping a fixed classifier size. The approach is evaluated across six public HAR datasets, achieving a last weighted F1 score of 84.9% and an average incremental weighted F1 of 81.34%, outperforming EWC and experience replay by notable margins, and demonstrating reduced forgetting and intransigence. This cross-dataset IL capability advances scalable HAR models and highlights KAN-based classifiers as a viable path for continual learning in multimodal sensor domains.

Abstract

This work proposes an incremental learning (IL) framework for wearable sensor human activity recognition (HAR) that tackles two challenges simultaneously: catastrophic forgetting and non-uniform inputs. The scalable framework, iKAN, pioneers IL with Kolmogorov-Arnold Networks (KAN) to replace multi-layer perceptrons as the classifier that leverages the local plasticity and global stability of splines. To adapt KAN for HAR, iKAN uses task-specific feature branches and a feature redistribution layer. Unlike existing IL methods that primarily adjust the output dimension or the number of classifier nodes to adapt to new tasks, iKAN focuses on expanding the feature extraction branches to accommodate new inputs from different sensor modalities while maintaining consistent dimensions and the number of classifier outputs. Continual learning across six public HAR datasets demonstrated the iKAN framework's incremental learning performance, with a last performance of 84.9\% (weighted F1 score) and an average incremental performance of 81.34\%, which significantly outperforms the two existing incremental learning methods, such as EWC (51.42\%) and experience replay (59.92\%).

Figures (6)

• Figure 1: Comparison between the existing framework for task incremental learning lu2024revisiting and our proposed iKAN framework. (The existing framework has one encoder and increased classifiers to incrementally learn the new tasks, which can only process the task with the same input, while the iKAN framework has multiple encoders aiming to receive non-uniform inputs and one KAN-based classifier. Thus iKAN can incrementally learn the tasks across heterogeneous datasets. The shape of the datasets(tasks) represents the different sensor modalities in the datasets(tasks))
• Figure 2: The Architecture of Encoder (The input shape is (1, W, C), W: window size, C: sensor channel number, output shape is (1, 2*F), F: the number of out channels in CNN2D layers, where four CNN2D layers keep the same out channels)
• Figure 3: Feature distribution result after the feature redistribution layer
• Figure 4: The performance of incremental learning by different methods (The grid number was set was 30 in iKAN, the lambda was configured as 80 in EWC, and store 200 samples of each class in the experience replay methods)
• Figure 5: Comparasion of the performance of the iKAN framework and existing methods in terms of forgetting and intransigence measure, the smallest result achieved by iKAN with 30 grid, whose forgetting measure is 0.0017, intransigence measure is 0.0012. (both the smaller the better)