{"id":"arxiv/2607.05457","url":"https://arxiv.org/abs/2607.05457","tags":["cs.LG","cs.AI","math.DS"],"score":0,"title":"Empirical Minimal-Realisation Compression of Deep Neural Networks via Controllability-Observability Tests","domain":"cs.LG","source":"arxiv","authors":["Anis Hamadouche","Amir Hussain"],"reviews":0,"subject":"cs.LG","summary":"Deep neural networks often contain substantial hidden-state redundancy, but most compression methods operate directly on weights, neurons, or quantised representations without explicitly characterising the dynamical role of internal states. This paper proposes a controllability-observability framework for empirical state-order reduction of deep neural networks. By viewing a trained network as a depth-indexed nonlinear dynamical system, we construct data-driven reachability, observability, and balanced Gramians from hidden-state snapshots and output Jacobians. The resulting A/B/C tests…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-07-05","reactions":{}}