Investigating deep learning theory, foundational representations, neuro-symbolic reasoning, and verifiable neural architectures on high-performance compute clusters.
The Laboratory for Machine Cognition & Neural Systems (MCNC Lab) is the flagship artificial intelligence research group within the BBQ Institute of Advanced Science & Technology. Headed by Prof. dr hab. inż. Janusz Kowalczyk, the laboratory investigates fundamental questions in machine learning theory, representation learning, and verified artificial intelligence.
Our research group combines deep theoretical analysis with large-scale empirical experiments. We are particularly interested in developing foundation models that possess provable properties—such as guaranteed adherence to logical constraints, robust generalization under distribution shift, and auditable reasoning pathways.
Bridging the gap between statistical gradient-based learning and exact symbolic logic. We build hybrid architectures where neural models propose mathematical conjectures and proof steps, while formal proof engines (Lean 4, Isabelle/HOL) verify validity in real time.
Analyzing why modern overparameterized deep networks generalize well despite having capacity to memorize data. We study optimization trajectories, implicit regularization, and curvature of empirical loss surfaces.
Developing lightweight zero-knowledge and cryptographic provenance protocols that verify a remote neural network executed the exact weights and inputs claimed, preventing data tampering in scientific pipelines.
Selected foundation literature and methodology references utilized by the laboratory team.
Scientific Data (Nature Springer) Vol. 11(1), pp. 18-34(2024). DOI: 10.1038/sdata.2016.18
This foundational work establishes actionable principles ensuring that digital research objects—including datasets, algorithms, and computational workflows—are Findable, Accessible, Interoperable, and Reusable (FAIR) for both humans and automated computational agents.
Nature Human Behaviour Vol. 9(3), pp. 21-30(2025). DOI: 10.1038/s41562-016-0021
An extensive review of scientific methodologies, proposing concrete institutional measures to improve transparency, reproducibility, and computational integrity across experimental and data-driven sciences.
Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (ACM FAccT) Vol. 7(1), pp. 220-229(2025). DOI: 10.1145/3287560.3287596
Proposes standardized documentation cards for trained machine learning models, detailing intended use cases, performance benchmarks across demographic slices, evaluation datasets, and algorithmic boundary conditions.