Recent advances in machine learning offer new opportunities to enhance searches for Beyond the Standard Model (BSM) physics by enabling more powerful, flexible, and computationally efficient data analysis strategies. In this seminar, I will present recent results from two projects focused on the application of neural networks to collider physics. In the first part, I will discuss the development of surrogate likelihood models that accurately approximate full experimental likelihoods, enabling a substantial acceleration of reinterpretation studies and an efficient exploration of high-dimensional BSM parameter spaces, with direct applications to phenomenological analyses. In the second part, I will introduce a novel anomaly detection framework for data-driven searches for new physics with minimal theoretical assumptions, based on unsupervised learning techniques to identify deviations from the Standard Model directly in data, thereby complementing traditional, signal-driven search strategies and improving sensitivity to unexpected signatures.