Embedding-space systems usually retrieve, rank, or recommend from a single point representation, even when the task would benefit from uncertainty or controlled diversity. We introduce DESS, a lightweight uncertainty layer that augments an existing embedding model with a predicted mean vector and an independent per-dimension spread vector. DESS is trained with deterministic element-wise losses, requires no sampling during training, and maps inference-time samples back to available outputs by nearest-neighbor selection, leaving the trained predictor decoupled from a fixed output dictionary. The central question is whether such a simple diagonal uncertainty layer is meaningful and safe in high-dimensional embedding spaces. We show that the learned spread captures real local geometry: on 20 Newsgroups sentence-transformer embeddings, it is strongly rank-correlated with empirical within-class variation (Spearman ρ = 0.94 ). However, it is not epistemic uncertainty: the spread norm performs at chance for held-out-class OOD detection (AUROC 0.49). We then show that inference-time sampling is governed by concentration of measure, with sample displacement concentrating near the spread-vector norm. This yields a training-set geometric rule for choosing the spread multiplier β , which predicts the safe sampling regime on AG News. On SQID/ESCI reranking, the strongest DESS variant preserves the frozen SBERT query embedding as the mean and learns only local spread; with β = 0.09 , sampled reranking matches the SBERT-text baseline without surpassing multimodal fusion or cross-encoder systems. DESS therefore provides a modular, geometry-aware uncertainty layer for embedding-space models, provided its spread is calibrated to local embedding geometry.
Over several decades, research on socially interactive agents has transformed human–machine interaction, evolving from early script-based systems to today’s generative, multimodal agents. Tracing this development through the ages, this article reflects on major conceptual and technical shifts that shaped the field, drawing on advances across multiple areas of AI, and discusses how progress in natural language processing and multimodal behavior analysis and generation has redefined AI as a social interface. Rather than presenting this trajectory as a narrative of continuous progress, we also examine persistent limitations, including weak grounding, difficult evaluation, opaque decision processes, safety risks in sensitive contexts, and the gap between convincing social simulation and genuine social understanding.
The Cognitive Robot Abstract Machine (CRAM) was developed to enable robots to perform everyday manipulation tasks in human environments by transforming underspecified task instructions into context-sensitive, executable actions. Rather than proposing a single planning algorithm, CRAM represents a cognitive architecture grounded in symbolic knowledge representation, generalized action plans, and introspective execution. This paper traces the evolution of the artificial intelligence ideas that shaped CRAM, from early work on frames, scripts, program interpretation and action-centered planning to contemporary developments in cognitive robotics. We show how these ideas are operationalized in the CRAM architecture and how successive extensions, including semantic digital twins, episodic memories, and dual-process reasoning, support robust execution, explanation and adaptation. In addition, we discuss the Virtual Research Building as an experimental substrate that enables systematic and reproducible evaluation of CRAM components. Finally, we outline future challenges for CRAM, highlighting the need to complement its top-down task-driven paradigm with more interaction-centered and enactive approaches for long-term deployment in human environments.
The research of the Image Processing group at the Mathematics/Computer Science Department of the University of Bremen was initially driven by requirements from the field of image retrieval, but it progressed very quickly to video and multimedia document analysis including texts, images, videos, and sound. By adopting symbolic image and video feature descriptions, it could also cover different challenging application areas such as Internet image search, search in TV archives and digital media art libraries, support for blind people, RoboCup vision technology, intercultural analysis of TV broadcasts, and next generation visual archives. In a definitely ambitious (and most successful) endeavor in 2005/2006, eleven highly engaged master students of a one-year student course project integrated the available combined image, video, and sound research results (and OpenCV!) with the goal to implement a movie trailer generator for different Hollywood movie genres. Their result was improved by 26 students in a second two-year student course project, especially in the trailer composition phase.
One aim of neurosymbolic AI is to enhance subsymbolic learning approaches with symbolic information, e.g., in the form of background knowledge, to increase result quality, interpretability, and trustworthiness. To model subsymbolic information on facts and their similarities, embeddings into some vector space have been proven useful. Symbolic information can then be modeled directly in the space by representing concepts as some geometric objects and logical operators as geometrical operations. These embeddings provide the first step towards filling the gap between qualitative, Tarskian style semantics, which is used for deductive reasoning over the facts, and quantitative structures, which are used for representing objects, relations, and concepts for learning purposes. To enable meaningful reasoning, embeddings of concepts are not allowed to be arbitrarily shaped. Particularly convex sets turned out to be useful due to their computational advantages and due to their foundation in cognition. However, in this context, the geometric modeling of logical operators turns out to be challenging, particularly the modeling of negation. Therefore, this paper discusses a general framework for embeddings able to model an expressive negation. It is general in the sense that it allows for relaxations of the convexity constraint and for different strengths of negation. Thus, the proposed framework can act as a basis for interpretable embedding approaches of adaptable strength and thus fills a gap in the current embedding landscape by allowing for embeddings of expressive logics.
Neurons along the visual pathway–from the retina to the primary visual cortex (V1) and higher visual areas–process information related to objects, such as their shape, texture, color, and direction of movement.The representation of spatial and temporal information in these areas is divergent and functional.Understanding how different features of input stimuli are represented throughout the visual system is important for comprehending how the brain computes spatial and temporal stimulus information. Although there is a substantial body of literature identifying how simple cell selectivity emerges with the help of synaptic plasticity, most of these papers focus on the role of excitatory synapses in selectivity emergence. These studies often implement only lateral inhibitory synapses between simple cells, which reduces the role of inhibition to decorrelating neuronal activity, leaving the role of plasticity in inhibitory synapses unclear.To address this knowledge gap, a spiking neural network (SNN) resembling the early visual pathway has been developed.We evaluated the emerging feature representation by measuring the diversity of the receptive fields and by testing the SNN on different object recognition datasets. Additionally, the roles of tuned lateral inhibition and synaptic delays in stimulus movement direction selectivity were investigated.
This dissertation abstract is a summary of the published doctoral thesis "Fast Neuro-Symbolic Approaches for Class Expression Learning". The thesis focuses on the development of efficient methods for class expression learning on knowledge graphs. The abstract presents four key contributions that achieve.
Artificial Intelligence has firmly established itself in the energy sector. It sees a multitude of applications, from seemingly simple and well-defined tasks such as forecasting and optimization to safety-critical applications that aim to increase an energy system's resilience. It is an active field of research and technical innovation, with scientists being quick to adopt the most up-to-date technologies, such as Large Language Models, while more established approaches, such as traditional agent systems, see continuous improvements towards practical application. However, the energy sector presents its own challenges. Not all aspects of artificial intelligence can be easily transferred from other domains to the energy sector. As critical infrastructure and a lifeline for modern society, the energy sector must guarantee the highest levels of reliability. Here, we present a survey on the current state of the art in artificial intelligence in the energy sector, highlighting current developments and major research and technology gaps.
Heating systems account for a significant share of residential energy consumption, and rising energy prices call for intelligent, cost-aware control strategies. Traditional methods, such as rule-based or model predictive control (MPC), often require detailed system modeling or lack adaptability to dynamic price signals. This work explores the use of deep reinforcement learning (DRL) to control heat pumps in a way that balances occupant comfort with energy-cost minimization. We evaluate deep Q-network (DQN) and proximal policy optimization (PPO) methods across discrete and continuous action spaces. The agents are trained in simulation using real weather and electricity price data, with a model representing the thermal dynamics of the building. Short-term electricity price forecasts are included to enable anticipatory heating strategies. Reward functions combine price penalties with piecewise-linear or quadratic comfort penalties. Among the DRL variants, a DQN agent with discrete actions and a piecewise-linear comfort reward achieves the best overall trade-off between comfort and cost. MPC still performs best in absolute cost terms because it uses an exact model, while the DQN policy approaches MPC performance and retains the model-free, adaptive advantages of RL. The findings highlight the potential of DRL for adaptive and price-aware heating control without the need for detailed physical modeling.
Ontology-driven conceptual models are generally considered a distinct class of conceptual models, as they draw on the semantics of foundational ontologies to provide a strong basis for modeling elements. These models are designed as tools for knowledge sharing and are therefore expected to be widely reused within their respective domains. However, studies indicate that understanding existing artefacts often poses challenges for users. To address this issue, we adopted a pragmatic approach to explanation, in which a model is clarified through a dedicated model view. We further demonstrated how such views, including model abstractions, can be systematically constructed by relying on the semantics provided by the underlying foundational ontology. User experiments confirmed that this approach can enhance the comprehension of ontology-driven conceptual models. The PhD dissertation with the corresponding title was defended at the Free University of Bozen–Bolzano [8]. Here, we summarize the key findings of this research and refer to the relevant literature for a more detailed discussion.
Accurate assessment of long-term wind power potential requires high-resolution wind speed data, yet most climate projections operate at spatial resolutions too coarse for reliable local estimates. Deep learning methods developed for image super-resolution offer a promising solution by learning statistical mappings from coarse to fine-resolution wind fields. In this paper, we systematically compare four state-of-the-art deep learning approaches namely purely convolutional, attention-based, spatio-temporal, and diffusion-based across two wind speed downscaling tasks of increasing complexity, evaluating their performance not only on standard pixel-based metrics but also on physically informed and distributional metrics directly relevant to wind power estimation. We show that pixel-based metrics alone are insufficient for model selection in energy applications: models that appear comparable on reconstruction accuracy can differ by up to 15 percentage points in recovered cumulative wind power. Diffusion-based downscaling consistently outperforms competing approaches on physical and distributional metrics and most accurately recovers long-term wind power in both experimental settings, with the advantage becoming more pronounced on the more challenging cross-dataset task.
Smart Grids (SG) enhance efficiency and centralised control by enabling networked device communication, but these capabilities expose them to cyberattacks. Machine Learning (ML) and Deep Learning (DL) based Intrusion Detection Systems (IDS) have been employed to detect these threats. Yet, their adoption introduces new adversarial risks: specifically, attacks designed to fool IDS into misclassifying malicious activity as benign. In this study, we propose ADVIS-G, a novel, adversarially defended IDS framework for smart grids utilising deep learning. Our approach begins by training a high-accuracy (macro F1 96+ ∼ 0.5 (at ϵ =0.1), adversarial training improves robustness to 0.7. Our proposed autoencoder-based blocking further increases the F1-score to ∼ 0.92 with RDU-Net, and ∼ 0.9 with U-Net. But the U-Net performed comparatively better under heavier attacks and normal images. Moreover, combining adversarial training with autoencoder defence achieves the highest resilience under stronger attacks. Additionally, MAE thresholding on reconstructions enables adversarial detection with an Area Under Curve (AUC) of 0.914 using RDU-Net and of 0.865 using U-Net. These results suggest that ADVIS-G significantly enhances IDS robustness against adversarial attacks, offering a promising direction for future smart grid security research.
Artificial Intelligence has firmly established itself in the energy sector. It sees a multitude of applications, from seemingly simple and well-defined tasks such as forecasting and optimization to safety-critical applications that aim to increase an energy system’s resilience. It is an active field of research and technical innovation, with scientists being quick to adopt the most up-to-date technologies, such as Large Language Models, while more established approaches, such as traditional agent systems, see continuous improvements towards practical application. However, the energy sector presents its own challenges. Not all aspects of artificial intelligence can be easily transferred from other domains to the energy sector. As critical infrastructure and a lifeline for modern society, the energy sector must guarantee the highest levels of reliability. Here, we present a survey on the current state of the art in artificial intelligence in the energy sector, highlighting current developments and major research and technology gaps.
Machine learning models are increasingly deployed in critical energy infrastructure, where domain experts require transparent explanations for decision-making. SHAP (SHapley Additive exPlanations) has become a popular method for energy systems applications. However, energy data exhibit inherent correlations due to physical constraints, operational relationships, and market dynamics, posing challenges for interpreting SHAP-based explanations. This work investigates how feature correlations influence SHAP-based explanations using controlled synthetic experiments and real-world power grid data. Our analysis shows that only correlation-aware methods can attribute importance to economically linked features, such as solar generation in predicting fossil fuels, which may reflect genuine systemic interdependencies that are valuable for prediction and scientific understanding. Our findings highlight the tradeoff between true to the model explanations that reflect model behavior and true to the data approaches that consider real-world dependencies. In complex energy systems with circular dependencies, temporal dynamics, and hidden constraints, explanation validity cannot be universally defined. We emphasize the need for practitioners’ awareness of the trade-offs between model analysis, scientific discovery, and operational understanding.
Semantic maps allow a robot to reason about its surroundings to fulfill tasks such as navigating known environments, finding specific objects, and exploring unmapped areas. Traditional mapping approaches provide accurate geometric representations but are often constrained by pre-designed symbolic vocabularies. The reliance on fixed object classes makes it impractical to handle out-of-distribution knowledge not defined at design time. Recent advances in Vision-Language Foundation Models, such as CLIP, enable open-set mapping, where objects are encoded as high-dimensional embeddings rather than fixed labels. In LIEREx, we integrate these VLFMs with established 3D Semantic Scene Graphs to enable target-directed exploration by an autonomous agent in partially unknown environments.
While deep learning has significantly advanced robotic object recognition, purely data-driven approaches often lack semantic consistency and fail to leverage valuable, pre-existing knowledge about the environment. This report presents the ExPrIS project, which addresses this challenge by investigating how knowledge-level expectations can serve as priors to improve object interpretation from sensor data. Our approach is based on the incremental construction of a 3D Semantic Scene Graph (3DSSG). We integrate expectations from two sources: contextual priors from past observations and semantic knowledge from external graphs like ConceptNet. These are embedded into a heterogeneous Graph Neural Network (GNN) to create an expectation-biased inference process. This method moves beyond static, frame-by-frame analysis to enhance the robustness and consistency of scene understanding over time. The report details this architecture, its evaluation, and outlines its planned integration on a mobile robotic platform.
The development of tools based on Artificial Intelligence (AI) provides significant potential to enhance industrial production, but many small and medium-sized enterprises (SMEs) in more rural areas, such as the study area Upper Franconia in Germany, encounter difficulties in implementing them due to limited resources and expertise. The SME AI Experience Center at the Cleantech Innovation Park offers practical support for AI integration through interactive workshops and a production plant simulating manufacturing processes. It showcases possibilities for the digital transformation by using different types of sensors and the potential of applying machine learning for tasks such as intelligent supply chains, predictive maintenance and quality control. The particular focus is on AI for sustainability. Furthermore, SME managers and workers will be aware that most AI technologies do not result in fully autonomous systems that replace the human workforce but that AI tools often rely on human-AI collaboration.
