FPU2025 - Neurosymbolic Computation for an Embodied Approach to Abstraction and Analogy-Making

We set out to develop a novel neurosymbolic artificial intelligence (AI) system that combines the neural-network-based analysis of a diagram with an ‘embodied cognition’ layer implementing the conceptual blend of a diagram’s geometric entities (as identified by the neural network) with symbolic representations of image-schematic structure and dynamics. We expect that such a neurosymbolic system will be capable of showing some degree of embodied understanding of a diagram to support abstraction and analogy-making abilities currently missing in ANN-based architectures trained by way of deep learning. We will evaluate our proposed system on abstract diagrams, as those used in mathematics and computer science with respect to reasoning and query-answering tasks that require abstraction and analogy-making abilities.