The digital transformation of ambient environments relies on autonomous systems capable of interpreting complex human behavior. However, activity recognition from multi-modal, noisy sensor data faces significant hurdles: real-world domains are intrinsically hybrid, blending discrete elements, such as specific activity types, with continuous variables like spatial position. Over time, the interaction of these variables produces an exponentially growing set of possible states, often exceeding the memory constraints of standard hardware. Simultaneously, while human behavior follows rigid logical structures, purely data-driven models (e.g., neural networks) struggle to learn these rules without massive datasets — resources that are rarely available in most applications.

To mitigate these constraints, research is increasingly focusing on Neurosymbolic (NeSy) systems, which synthesize the learning power of neural networks with the formal logic of symbolic reasoning. By leveraging symbolic background knowledge, these frameworks can compensate for data deficits while enhancing interpretability and generalization. However, state-of-the-art implementations frequently encounter scalability bottlenecks in high-dimensional settings and rely on highly specialized inference algorithms tailored to specific mathematical representations and underlying logical formalisms. The ProbNeSy4Activities project seeks to overcome these hurdles by developing advanced probabilistic-neurosymbolic models designed for dynamic, hybrid, and structured probability spaces.

The project focuses on extending the CCBM framework, originally developed by the HAIML research group for efficient reasoning in complex and dynamic spaces. The objective is to enable representation-agnostic probabilistic inference and modeling in combination with the integration of neural networks for data-driven learning. The introduction of a representation-agnostic API allows for expressing probabilistic computations without making assumptions about the underlying distributional representation. This flexibility facilitates the combination of continuous and discrete variables within a single (hybrid) probabilistic model, as well as the integration of arbitrary distibutional representations such as neural networks and symbolic/procedural descriptions. Furthermore, the research seeks to facilitate the simultaneous learning of probabilistic and neural parameters by combining gradient-based and gradient-free optimization. Ultimately, this unified approach aims to achieve robust, data-efficient activity recognition in complex environments while remaining computationally tractable.