{"id":"arxiv/2606.00517","url":"https://arxiv.org/abs/2606.00517","tags":["math.NA"],"score":0,"title":"Deep neural network yields regularization for ill-posed inverse problems","domain":"math.NA","source":"arxiv","authors":["Qiao Zhu","Lan Wang","Ye Zhang"],"reviews":0,"subject":"math.NA","summary":"This paper studies the regularization of ill-posed inverse problems by deep neural networks (DNNs). We extend architecture-based regularization from shallow networks to deep models by developing a deterministic framework in which the admissible network class is enlarged adaptively and the resulting architecture complexity acts as the regularization mechanism. We propose two discrepancy-principle-driven expanding DNN algorithms to treat the cases where an explicit parameter-radius bound is available and unavailable, respectively. For both algorithms, we prove the finite termination of the…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-05-30","reactions":{}}