This paper demonstrates that deep neural networks, despite being trained on diverse tasks and initializations, converge to remarkably similar low-dimensional parametric subspaces. This finding offers significant implications for model reusability, multi-task learning, and reducing the computational and environmental costs of large-scale neural models.

This paper introduces an online methodology for real-time traffic density estimation using Physics-Informed Neural Networks (PINNs) with probe vehicle measurements. The proposed framework demonstrates robustness to model errors and noisy data, outperforming classical open-loop observers in scenarios with model mismatch and successfully reproducing traffic characteristics in high-fidelity simulations.