Włodzisław Duch

Interpretability is a central requirement in EEG-based brain-computer interfaces (BCIs), where decoding models are expected to provide insight into the neural representations underlying their predictions. However, it remains unclear how different interpretability strategies perform under realistic motor-imagery EEG conditions, particularly when post-hoc attribution methods are compared with architectural mechanisms intended to enforce sparsity. In this study, we present a systematic multi-dataset evaluation of channel-level interpretability in motor-imagery (MI) EEG decoding using the EEGNet architecture. We compare three gradient-based attribution methods, Saliency, SmoothGrad, and Integrated Gradients, with two architectural approaches designed to promote channel sparsity: Group LASSO regularization and differentiable Concrete Masking. We assess the stability and structure of channel-level relevance estimates across four widely used MI datasets. The gradient-based attribution methods produce highly consistent relevance patterns across subjects, folds, and datasets, with reproducible spatial structure. Group LASSO regularization largely preserves these patterns but produces only modest changes in channel relevance and limited additional channel selectivity. In contrast, Concrete Masking consistently converges to nearly uniform gate values across datasets and tested hyperparameter configurations. Although classification performance is preserved, the learned gates do not reveal consistent channel-selection structure. These results indicate that, in EEGNet-based MI decoding, post-hoc gradient attribution provides more stable channel-level explanations than the examined sparsity-inducing architectural mechanisms, while also emphasizing that attribution stability should not be equated with neurophysiological validity.

A brain-computer interface (BCI) establishes a direct communication pathway between the human brain and a computer. It has been widely used in medical diagnosis, rehabilitation, education, entertainment, etc. Most research so far focuses on making BCIs more accurate and reliable, but much less attention has been paid to their privacy. Developing a commercial BCI system usually requires close collaborations among multiple organizations, e.g., hospitals, universities, and/or companies. Input data in BCIs, e.g., electroencephalogram (EEG), contain rich privacy information, and the developed machine learning model is usually proprietary. Data and model transmission among different parties may incur significant privacy threats, and hence privacy protection in BCIs must be considered. Unfortunately, there does not exist any contemporary and comprehensive review on privacy-preserving BCIs. This paper fills this gap, by describing potential privacy threats and protection strategies in BCIs. It also points out several challenges and future research directions in developing privacy-preserving BCIs.

The complexity of cultures in the modern world is now beyond human comprehension. Cognitive sciences cast doubts on the traditional explanations based on mental models. The core subjects in humanities may lose their importance, but new, interdisciplinary branches of humanities emerge. Humanities have to adapt to the digital age. Instant access to information will be replaced by instant access to knowledge. Artificial intelligence is going to change humanities in a radical way, from art to political sciences and philosophy. Understanding the cognitive limitations of humans and the opportunities opened by the development of artificial intelligence and interdisciplinary research necessary to address global challenges is the key to the revitalization of humanities.

One important application of machine learning is the analysis of electroencephalographic recordings (EEG). Such oscillatory signals are noisy, non-stationary, full of artifacts, contain transients, and chaotic transitions between meta-stable states. EEG data rarely consists of recordings obtained for more than 100 patients with some brain disorder. This makes objective, reliable clinical diagnosis of many mental disorders, such as schizophrenia, very difficult. The standard approach splits the frequency of EEG oscillations focusing on five classical bands: delta, theta, alpha, beta, and gamma. In this paper, we investigate three main questions: (i) Are there certain frequencies that will allow for a reliable diagnosis of schizophrenia, and (ii) if they are, is there a connection to what is known about their neural basis? (iii) How long segments of EEG recordings are sufficient for reliable classification? We filter EEG signals, obtaining a set of 64 very narrow 1Hz frequency bands. A genetic algorithm, with a simple k-NN classifier, finds an optimal combination of these bands. The method is sufficiently simple to be used in clinical settings. A public schizophrenia EEG data set, containing 60 seconds of EEG recordings in the resting state, with only 16 electrodes, for 45 adolescent patients and 39 healthy controls, was used for testing. Signal duration of about 30 seconds enables to reach an accuracy of over 96% in 5-fold cross-validation tests. We compare our results to much more sophisticated state-of-the- art methods and discuss insights gained by our analysis into the brain basis of schizophrenia.

Human Activity Recognition (HAR) plays an important role in the automation of various tasks related to activity tracking in such areas as healthcare and eldercare (telerehabilitation, telemonitoring), security, ergonomics, entertainment (fitness, sports promotion, human–computer interaction, video games), and intelligent environments. This paper tackles the problem of real-time recognition and repetition counting of 12 types of exercises performed during athletic workouts. Our approach is based on the deep neural network model fed by the signal from a 9-axis motion sensor (IMU) placed on the chest. The model can be run on mobile platforms (iOS, Android). We discuss design requirements for the system and their impact on data collection protocols. We present architecture based on an encoder pretrained with contrastive learning. Compared to end-to-end training, the presented approach significantly improves the developed model’s quality in terms of accuracy (F1 score, MAPE) and robustness (false-positive rate) during background activity. We make the AIDLAB-HAR dataset publicly available to encourage further research.

Sciendo provides publishing services and solutions to academic and professional organizations and individual authors. We publish journals, books, conference proceedings and a variety of other publications.

Electroencephalography (EEG) allows recording of cortical activity at high temporal resolution. Creating features useful for the analysis of the EEG recording can be challenging. Here we introduce a new method of pre-processing the time-series for the analysis of the resting state and binary task classification using recurrence quantification analysis (RQA) and compare it with the existing state-of-the-art approach based on signal embedding. To reveal patterns that unfold brain dynamics, we present a new pipeline that does not rely on selection of embedding parameters for RQA. Instead of using EEG time-series signals directly, Short-term Fourier transform (STFT) is used to generate new time-series, based on the power spectra from sliding, overlapping windows. Recurrence plots are created in a standard way from embedded EEG signals, and the STFT vectors. The efficiency of RQA features extracted from such plots is compared in classification of EEG segments that correspond to open and closed eye conditions. In contrast to the common approaches to such analysis, no filtering into separate frequency bands was needed. Differences between the two representations of EEG signals are illustrated using histograms of RQA features and UMAP plots. Classification results at the 95.9% level were obtained using selected features for less than 10 electrodes.