Speaker
Description
Analysis of signals by means of mathematical transformations proved to be an effective method in various aspects, such as filtering, system identification, feature extraction, classification etc. The most widely used method in transform-domain techniques operates with fixed basic functions like the trigonometric functions in the Fourier transform, Walsh functions in the Walsh–Fourier transform, mother wavelet function for the wavelet transform, etc. In these cases, the flexibility of the method is in the proper choice of the function system. Once the system is set it is used regardless the difference between the individual signals. In other words, the system can be adjusted to the problem level, but not to the individual signal level. This limitation turned to be significant in such cases, especially in dynamically changing environments, where we need to adjust the system to the signal. One way to surpass this limitation is to use adaptive orthogonal transformations. In recent years, we generalized this concept, developed various adaptive mathematical models, and successfully applied them in a range of applications, including ECG, EEG signal processing, telecommunication, CT-, photoacoustic-, and thermographic imaging. In these applications the transformation step was followed by various machine learning techniques. Even though the interaction of these two phases were studied and considered but they were not integrated into a uniform method. In our recent project, by utilizing our former work, we incorporate the representation abilities of adaptive orthogonal transformations and the prediction abilities of neural networks (NNs) in form of a joint model called VPNet. This is a novel model-driven NN architecture for 1D signal-processing problems which utilize variable projection (VP). Applying VP operators to neural networks has the advantage of learnable features, interpretable parameters, and compact network structures. We show that, compared to fully connected and one-dimensional convolutional networks, VPNet offers fast learning ability and good accuracy at a low computational cost of both training and inference. Based on these advantages and the promising results obtained, we anticipate a profound impact on the broader field of signal processing, in particular on classification, regression and clustering problems. In order to demonstrate its efficiency, we evaluated the performance of the VPNet approach in three tasks including classification of normal and abnormal heartbeats in real electrocardiogram (ECG) signals, color classification of visually evoked potentials in EEG signals, and road surface detection based on wheel sensor data.
