Neuro-Symbolic Reasoning

Neuro-Symbolic Reasoning is field at the intersection of statistical machine learning techniques and computational logic. Our work focuses on the integration of the learning capabilities of neural networks with the structured, interpretable reasoning of symbolic systems. By combining these paradigms, we aim to create intelligent systems that are not only capable of processing vast amounts of unstructured data but also of reasoning and explaining decisions. In particular, our research targets hybrid forms of reasoning, explainable AI systems, and visual question answering.

Overview

Hybrid Reasoning

Real-life data is generally unstructured and messy, thus neural networks and similar statistical systems are generally more succesful at handling this data than symbolic approaches. However, unlike symbolic systems, they struggle with logical reasoning and abstraction. Research into hybrid forms of reasoning, which combine the strengths of both statistical and symbolic AI is thus paramount.

Explainable AI Systems

The decisions made by black-box machine learning algorithms are often not comprehesible for humans. Explainable AI systems seek to mitigate this by offering justifications, accountability, and human-interpretable reasoning. Neuro-symbolic reasoning can be seen as a way to achieve explainable AI systems, since the symbolic component is better to interpret and formal explanation strategies can be studied more easily.

Visual Question Answering

Visual Question Answering (VQA) combines computer vision, natural language processing and reasoning, where systems need to answer questions based on visual input like images or videos. It requires understanding both the content of the visual data, the context of the question and logical reasoning capabilities to generate accurate and coherent responses. Hence, VQA is a prime use-case for neuro-symbolic reasoning and as such, is an ongoing topic of research in our group.