Parameter Space Reconstruction

Parameter space reconstruction aims to recover the internal parameters of a system, such as a neural network or a physical model, based on its observable input-output behavior. Current research focuses on developing efficient algorithms, often employing derivative-free optimization or novel query generation strategies, to overcome the challenges posed by high-dimensionality and complex non-linear relationships within the parameter space. These techniques are applied to diverse models, including convolutional neural networks (like ResNet50), neural ordinary differential equations, and even compact models for semiconductor devices. Successful reconstruction has significant implications for improving model interpretability, enhancing security against model extraction attacks, and accelerating the design and optimization of complex systems.