Learning Changes the World and Then What Was Learned has Changed: Training and Deploying Neural Networks in the Control Loop (Conference Presentation)

This deposit contains the official presentation slides delivered at the 23rd International Conference on Informatics in Control, Automation and Robotics (ICINCO 2026) in Angers, France, accompanying the paper: "Learning Changes the World and Then What Was Learned has Changed: Training and Deploying Neural Networks in the Control Loop" by Vittorio Lippi (Calejo Hybrid Intelligence AB).Published in: Proceedings of the 23rd ICINCO - Volume 1, ISBN: 978-989-758-870-9, ISSN: 2184-2809, pages 15–26.Preprint: arXiv:2610.22084 [stat.ML]Code & Data Archive: DOI: 10.5281/zenodo.22774300 Summary & Contributions When neural networks (NNs) are trained offline on static datasets and subsequently deployed within physical feedback control loops, their control actions fundamentally alter the physical state distribution upon which they were parameterized. This presentation formalizes this fundamental failure mode and provides actionable control-theoretic solutions: The Three-Part Failure Taxonomy: Disentangles sequential degradation into (1) the initial Closed-Loop Identification Error ($E_{ID}$), caused by baseline controller correlation with process noise during data collection; (2) the deployment-induced Action Error ($E_{Act}$), where replacing the baseline controller warps the operating manifold; and (3) the continuous Tracking Error ($E_{Tr}$). Discrete-Time Sensitivity Formalism (Theorem 1): Formalizes how offline-learned parameters locally guarantee trajectory divergence upon deployment, providing an exact analytical bound on state sensitivity accumulation ($S_k$) dictated by plant and controller Lipschitz constants. Physical Time-Scale Separation (Theorem 2): Identifies the exact physical bottleneck governing safe online adaptation (\(\lambda_{\max}(\Gamma) < 2 / L_\theta\)). Proves that parameter adaptation rates must be strictly bounded below the plant's mechanical relaxation time to avoid high-frequency chatter and closed-loop instability. Ablation Benchmarks on Nonlinear Duffing Oscillator: Demonstrates that static deployment and structural closed-loop identification frameworks (Dual IOP / SLP) fail under dynamic distribution shift, whereas bounded streaming adaptation (SGD with gradient clipping) successfully stabilizes tracking. Robotic Case Studies: Synthesizes real-world empirical observations from bipedal posture control (Robot Lucy) and gait phase classification on lower-limb exoskeletons (Exosmooth). Universal AI Feedback Implications: Connects the control-theoretic Action Error to broader socio-technical phenomena, including Generative Epistemic Injustice and autoregressive model collapse. Contents of this Deposit Lippi_ICINCO_2026_Presentation_Slides.pdf: Standard 16:9 widescreen presentation slides without speaker notes, formatted for clean viewing, reading, and citation. CITATION.bib: BibTeX citation entries for the presentation and published conference paper.