A 2.47 μJ/sample QR-Decomposition-based Extreme Learning Machine Engine Supporting Online Class Incremental Learning for ECG-based User Identification

In the Industrial Internet of Things (IIoT) applications, smart devices generate massive data during monitoring machine status. Meanwhile, authorized access to these IIoT data and machines becomes crucial where biometrics comes in handy [1]. Among all bio-signals, Electrocardiography (ECG) signal has two key advantages: 1) ECG signals indicates electrical activity of the heart, providing intrinsic proof of liveness. 2) ECG signals cannot be easily eavesdropped without physical contact [2]. These anti-spoofing features make ECG an attractive biometric modality in the IIoT scenario. One challenge in such user identification system is how to identify a “newly enrolled user” efficiently without retraining the entire system as shown in Fig. 1. That is, the classifier of the system should “incrementally” learn the new authorized users without “forgetting” the previously learned/trained ones, called “online class incremental learning (O-CIL).” It involves both “online learning (OL)” and “class incremental learning (CIL).” However, most back-propagation-based neural networks (BP-NN) suffers from the notorious “catastrophic forgetting,” which results in severe accuracy degradation due to the recency learning bias towards new classes [3]. To deal with the issue of catastrophic forgetting, the scalable extreme learning machine (S-ELM) [4] provides an analytical solution by online updating the classifier with recursive least squares (RLS). However, the matrix-inversion operations of conventional RLS are very computation-intensive, which is not suitable for direct VLSI implementation. Moreover, most existing ELM chips only support inferencing [5], lacking the O-CIL capability.