Software Competence Center Hagenberg (Austria)

Recent findings suggest that consecutive layers of neural networks with the ReLU activation function fold the input space during the learning process. While many works hint at this phenomenon, an approach to quantify the folding was only recently proposed by means of a space folding measure based on the Hamming distance in the discrete activation space. We generalize the space folding measure to a wider class of activation functions through the introduction of equivalence classes of input data. We then analyze its mathematical and computational properties. Lastly, we link the folding to geometry of adversarial attacks. We underpin our claims with an experimental evaluation.

Waddington's landscape provides a conceptual model for developmental processes. It is the basis of various mathematical models describing cell maturation and development at cell and population levels. Yet, these mathematical models mostly disregard cell-to-cell communication, an essential process that modulates cellular decision-making and population dynamics. In this study, we provide a dynamical model for cell maturation and development which can be seen as an extension of Waddington's landscape. The coupled system of partial and ordinary differential equations describes cell density along the cell state together with ligand concentrations. Cell-state-dependent ligand production determines ligand availability, which controls population-level processes. We provide proof of the existence and uniqueness of solutions for our coupled differential equation system and demonstrate the model's validity by analyzing single-cell transcriptomics data. Our results show that cell-to-cell communication is essential for accurately depicting biological recovery processes, such as the regeneration of stem cells in the intestine's crypt and the response of immune cells upon LSP stimulation. Our findings underscore the importance of incorporating cell-to-cell communication into mathematical models of biological development. By doing so, we unlock the potential for deeper insights into complex processes such as tissue regeneration and immune responses, offering new avenues for understanding and predicting the dynamics of biological recovery and cell activation.

Abstract Hand gesture recognition (HGR) has gained significant attention in human-computer interaction, enabling touchless control in various domains, such as virtual reality, automotive systems, and healthcare. While deep learning approaches achieve high accuracy in gesture classification, their lack of interpretability hinders transparency and user trust in critical applications. To address this, we extend MIRA, an interpretable rule-based HGR system, with a novel gesture onset detection method that autonomously identifies the start of a gesture before classification. Our onset detection approach achieves 90.13% accuracy on average, demonstrating its robustness across users. By integrating signal processing techniques, MIRA enhances interpretability while maintaining real-time adaptability to dynamic environments. Additionally, we introduce a background class, enabling the system to differentiate between gesture and non-gesture frames and expand the dataset with new users and recordings to improve generalization. We further analyze how feature diversity affects performance, showing that low diversity can suppress personalization due to early misclassifications. Using a foundational and personalized rule framework, our approach correctly classifies up to 94.9% of gestures, reinforcing the impact of personalization in rule-based systems. These findings demonstrate that MIRA is a robust and interpretable alternative to deep learning models, ensuring transparent decision-making for real-world radar-based gesture recognition.

Quantum low-density parity-check codes are a promising candidate for fault-tolerant quantum computing with considerably reduced overhead compared to the surface code. However, the lack of a practical decoding algorithm remains a barrier to their implementation. In this work, we introduce localized statistics decoding, a reliability-guided inversion decoder that is highly parallelizable and applicable to arbitrary quantum low-density parity-check codes. Our approach employs a parallel matrix factorization strategy, which we call on-the-fly elimination, to identify, validate, and solve local decoding regions on the decoding graph. Through numerical simulations, we show that localized statistics decoding matches the performance of state-of-the-art decoders while reducing the runtime complexity for operation in the sub-threshold regime. Importantly, our decoder is more amenable to implementation on specialized hardware, positioning it as a promising candidate for decoding real-time syndromes from experiments.

Abstract Artificial Intelligence (AI) systems, which are currently reshaping the global employment landscape, require data to be annotated for training and evaluation – a task performed mainly by annotation workers located in the Global South. That AI-related companies are located in Western countries, but annotation work is performed in regions that are culturally and socioeconomically different implies low-quality, potentially culturally biased annotation under unacceptable working conditions. It would thus be desirable to move annotation work to Western countries. This study investigated whether annotation could be a promising occupation for individuals on the autism spectrum (AS), an underprivileged group in the labor market in Western countries. Our participatory research project included adults on the AS (experimental group, EG; N = 22) and a neurotypical comparison group (CG; N = 17). Job-related self-perceptions were collected from all participants, neurocognitive parameters (nonverbal IQ, reading comprehension, visual attention) were assessed for EG participants, and correlations with annotation outcomes were analyzed. Findings confirmed the hypothesized potential of autistic individuals to perform high quality data annotation. Good reading comprehension supported annotation outcomes, while higher IQ scores (> 110) were related to a tendency towards lower annotation quality, which might be explained by annotation tasks being insufficiently demanding, although neither correlation was significant in the limited study sample. Annotation performance might be supported by particular working conditions (i.e., working in small groups, minimal distraction, visual guidelines). Future studies should explore long-term trajectories (i.e., staying in work, income development, well-being) for annotation workers on the AS to confirm the reality of opportunities for this population.