Current breakthroughs in artificial intelligence stand on decades of progress in understanding artificial neural networks as trainable, nonlinear learning systems. Years of insights into complex dynamical systems in physics now offer a new starting point for intrinsically neuromorphic processors. Our manuscript from 2023 in Advanced Science explores the fundamental principles of neural-like computing using nonlinear wave dynamics in optical single-mode fibers and deepens understanding through experimental demonstration based on optical waves. Specifically, we demonstrate how broadband frequency generation through optical pulse decay processes in a single fiber enables the emulation of classification and prediction functions of various optical networks. The underlying training principle is based on the Extreme-Learning Machine framework, which allows for highly efficient training (linear regression) of a generalizable map of the high-dimensional feature projection of the fiber's output spectra and the data labels of a classification or regression task. Our findings reveal that nonlinear frequency mixing can be used for feature extraction to a point where a single fiber could replace numerous classification models. Our latest findings with the system even suggest that nonlinear inference capacity in optical systems can surpass deep neural networks with five hidden layers on nonlinear classification benchmarks — opening a new frontier in computing efficiency.

Interfacing evolutionary algorithms with adaptive photonic waveguide systems paves the way towards novel smart applications in optics. Yet, both algorithms and blueprints for all-optical integration are just starting to emerge into the field of optical sciences. In our recent work from 2021, we have presented a functional toolbox for autonomous shaping of telecom-relevant 10-100ps pulses. Through the unique use of a particle swarm optimization algorithm, the system could reutilize an alienated, computer-programmable on-chip interferometer for temporally coherent synthesis of various envelope shapes at system output. An all-optical sampling technique delivered the feedback to the algorithm, which smartly optimized the multi-path interferometer towards the best match between an optical output to a given target waveform. The entire patented scheme is potentially chip-integrable and might serve as a great template for future applications in all-optical switching and modulation control.

Liquids as optical media are unique in their extraordinarily strong, non-local nonlinearities. How such nonlinearities affect the common dynamics in nonlinear systems in not entirely known. In a key contribution from 2017, we reported on the first experimental indications and numerical verification of a modified fission dynamics of self-maintaining optical pulses (i.e., solitons) occurring in liquid-core fibers. This uncommon behaviour, which occurs as comet-like feature in the spectrogram of the supercontinuum output (see figure), is caused by the inimitable long-lasting molecular (Raman-like) nonlinearities of liquids. Through this work, I extended the early theory about non-instantaneous solitons by Conti et al., with more insights soon to come.