• Bachelor in Technical Physics at Technische Universitaet Ilmenau (Germany) and Bachelor Thesis at Chemical Physics department at Lund University (Sweden).
◦ Bachelor Thesis title: “Defocused Imaging of Perovskite Quantum Dots”
◦ finished 2017
• Two years (2017-2019) working student as Software Developer at Hella Aglaia Mobile Vision GmbH
◦ the company develops intelligent driving safety systems towards autonomous driving
◦ working in the Vision Solutions department
• Master in Physics at Technische Universitaet Berlin with thesis at the Institute of theoretical physics about “Deep Time-Delay Reservoir Computing” in 2020
◦ simulations of unidirectionally coupled opto-electronic reservoirs
◦ investigation of hyperparameter settings and their relation to the conditional Lyapunov exponents of single layers and the overall influence on memory capacity of the deep system
Based on increasing data sets and faster processors, artificial intelligence finds its way into industry and science. With this, the demand for computational resources increases massively. Nevertheless, hardware based on the von-Neumann architecture exhibits several shortcomings while used for AI, such as high energy consumption and long run times. Therefore, unconventional computing methods based on photonic hardware gain more and more interest. One promising concept called reservoir computing utilizes the natural information processing of dynamical systems and enables machine learning with photonic hardware. In this talk, we investigate the information encoding into a dynamical system based on optoelectronic hardware. An information-carrying input signal drives a delayed dynamical system and triggers its transient information processing. We analyze the abilities of the reservoir to recall the past and predict the future of a chaotic time series. In a second step, we introduce a concept where several optoelectronic delay systems are coupled in a row to utilize their interplay and overcome trade-offs known in reservoir computing. We present a configuration based on the modular design principle where only the internal order in the deep reservoir is varied, providing an effective way to utilize an ensemble of layers for several different tasks without even changing their hyperparameters.
Meeting ID: 830 4334 7066
Detalles de contacto:
Miguel C. Soriano Contact form