Extreme mass ratio inspirals (EMRI) are one of the most exciting gravitational-wave sources the Laser Interferometer Space Antenna (LISA) will detect. They are made of compact objects evolving in the highly curved spacetime around Massive Black Holes (MBH). As such, they are exquisite probes of the strong regime of gravity. EMRIs emit gravitational-wave signals that typically feature 10⁵ orbit cycles and can last one year in the observation window. They have relatively low signal-to-noise ratios, so they must be observed long enough. Their signal has a highly complex structure as it contains the evolution of the harmonics of three fundamental frequencies: the orbital frequency, the perihelion precession frequency and the orbital plane precession frequency. Currently, state-of-the-art approaches for EMRI detection and parameter estimation involve semi-coherent searches, which combine analyses of the relatively short data segments. However, the problem of their detection in noisy data is challenging and has only been partially addressed, mainly because of computational complexity. New techniques can be applied to drive this effort forward. One of the keys to efficient detection is to represent the signal in the most compressed way. Such representations can use time-frequency transformations like wavelets, machine learning-driven algorithms like Normalising Flows, or a combination of both. This project will investigate ways to accumulate information from EMRI long-lasting signals. First, we will study different methods for data reduction. We will explore different types of Autoencoders, starting from linear principal component decomposition and simple fully connected networks, gradually ramping up to more complicated methods. The goal is to find the most compact data representations that preserve scientific information.