Speaker: Daniele Bottazzi
(Heidelberg University)
10 Febbraio 2026 | 14:15
DEIB, Sala Seminari "A. Alario" (Ed. 21)
Per maggiori informazioni: Silvia Cascianelli | silvia.cascianelli@polimi.it
Sommario
Tuesday, February 10th, 2026 at 2:15 pm a new appointment of Data Science Seminars: Bioinformatics focus will take place in DEIB "Alessandra Alario" Seminar Room (Building 21) organized by the Data Science for Bioinformatics group.The seminar will be held by Daniele Bottazzi, master student in Computer Science at Saezlab of the Heidelberg University, on the following subject: ""Knowledge-Based Machine Learning via Semi-Amortized Neural Network and Differentiable Convex Optimization Layers".
Integrating domain knowledge into machine learning models remains challenging, particularly in biological applications where physical constraints, conservation laws, and specific mechanistic relationships are known but difficult to incorporate into powerful and expressive neural network architectures.
We propose a general purpose end-to-end differentiable framework that couples neural amortization with a structured convex optimization layer formulated as a single quadratic program (QP). The amortizer network generates condition-specific outputs from input features, which are then refined by the QP through a proximal-style optimization, that incorporate knowledge-derived constraints while minimizing deviation from the network's raw predictions. With mild strong convexity, the resulting solution map is unique and locally Lipschitz, ensuring stable gradients and exact feasibility.
We validate the proposed framework with (i) a proof-of-concept on a classical max-flow problem, illustrating how the convex layer enforces feasibility while the amortizer learns to produce accurate solutions, and (ii) a biological application: cell growth prediction from media composition via flux balance analysis (FBA) stoichiometric and reaction-bound constraints. Experiments show near-zero constraint violations, competitive predictive accuracy, and improved interpretability compared with purely data-driven baselines, while scaling efficiently with GPU-batched training.