AI-generated image (Copilot)
In recent decades, we have been repeatedly surprised by the performance of deep learning systems in challenging image, text, and speech processing tasks. However, developing effective deep learning solutions is notoriously challenging, as such solutions require large amounts of data and computational processing (often beyond the limited budgets of typical users), are very sensitive to domain changes and spurious correlations in the data, and occasionally produce undesirable results that impair final performance and user confidence in the system. The good old artificial intelligence techniques, based on knowledge representation and symbol manipulation, are data-efficient, generalizable, and most importantly, produce verifiable behavior; however, they are not scalable, require expensive knowledge acquisition procedures, and have difficulty handling noise and uncertainty that are ubiquitous in cognitive tasks (question answering, object recognition, argumentation, etc.). Neurosymbolic approaches have recently resurged as a means to leverage the best of both approaches, providing systems that are both expressive and scalable, as well as interpretable, generalizable, data-efficient, and reliable.
A group of researchers at KEML is actively developing a computational framework that will enable the rapid development of neurosymbolic solutions specified using modern deep learning tools (e.g., PyTorch) and domain-specific specialized language for knowledge representation (e.g., Answer Set Programming). The framework is expected to facilitate the construction of systems that combine deep learning routines and specialized/common sense knowledge, integration of classification systems (e.g., different neural classifiers trained for different purposes), and various new forms of learning that combine data and knowledge, such as distant learning and learning with logical constraints. The group is also committed to creating case studies to showcase the system’s capabilities in challenging tasks such as probabilistic argumentation, automatic essay annotation, and question answering.
Learn a bit more about ….
- Silveira, I.C., Barbosa, B., Mauá, D. D.. A New Benchmark for Automatic Essay Scoring in Portuguese. Proceedings of the 16th International Conference on Computational Processing of Portuguese, 2024.
- Mauá, D. D., Cozman, F. G. Specifying credal sets with probabilistic answer set programming. Proceedings of the Thirteenth International Symposium on Imprecise Probability: Theories and Applications, 2023.
- Other publications here!