The inception of our collaboration with BrainCreators traces back to 2019 when Vitor Horta and Maarten Stol met at a conference. Vitor was just starting the second year of his PhD research then, but the research was so timely it did not take long to publish and to be approached for a research chat after his conference presentation. This meeting ignited a shared interest in extracting knowledge graphs from neural networks to enhance explainability. Since then, Vitor has earned his PhD, and our long-term research partnership has flourished, resulting in several joint papers and groundbreaking research on taxonomy extraction from convolutional neural networks (CNNs). This work is recognized as a promising avenue in the NeuroSymbolic community, so much that following its publication I have been asked to join the NeuroSymbolic AI Journal as part of the Editorial Board.
Over two years of collaborative research, Maarten secured access to BrainCreators’ data on road cracks, leading to an innovative in-kind project about to start. We will be focusing on extracting semantic knowledge and using deductive rules to characterise the spatial relationships between cracks. These efforts have been instrumental in advancing our understanding and application of Neuro-Symbolic integration via knowledge graphs in practical scenarios.
Our monthly research discussions with Maarten have been exceptionally forward-looking. They have enabled us to identify a novel theoretical perspective on data-driven AI, particularly deep learning, and to set up a comprehensive research agenda. This agenda aims to match formal logics with neuro symbolic requirements, driven by specific use cases. Following recent acceptance of this position paper, Maarten has been invited to give a talk for a workshop at VU Amsterdam for the Marie Skłodowska-Curie Actions doctoral network called “Learning with Multiple Representations” (https://lemur-dn.github.io/). His talk scheduled for July 10th 2024 is titled “IID Relaxation by Logical Expressivity: A Research Agenda for Fitting Logics to Neurosymbolic Requirements”.
Looking ahead, Maarten and I plan to leverage the theoretical insights and practical outcomes of our collaboration to address challenges in BrainCreators’ flagship product for harbour surveillance, Seagull. Our goal is to demonstrate the robustness and effectiveness of knowledge-enhanced AI approaches within neural symbolic integration. By doing so, we aim to make a compelling case to the company leaders that these approaches can offer more contextual robustness, computational efficiency, and reduced opacity compared to traditional computer vision methods.
Our research has also benefited greatly from the contributions of several talented interns. Two of our former interns, Robin Sobczyk and Alexandre Monnier, have played key roles in our taxonomy extraction and adaptation of the graph extraction methods from images to language processing. Both Robin and Alexandre are now pursuing their PhDs, continuing their academic journeys. Additionally, Vitor, who started with us as a PhD student, has now advanced his career to become a shareholder and Data Science Manager at a company in Brazil, and we are currently exploring new opportunities for collaboration as targeted projects within INSIGHT.
The team: Alessandra Mileo, Maarten Stol, Vitor Horta, Robin Sobczyk, Alexandre Monnier and Meem Arafat Manab
