2025-10-02 –, Coffee room
Language: English
With the ever increasing computational cost of machine learning applications in mind, the need to explore alternative learning paradigms becomes apparent. One such paradigm would be to exploit the massive parallelism of physical systems. We introduce a metamaterial based on coupled nonlinear resonators that learns in-situ, and show how we are able to study it seamlessly across various levels of abstraction, building on Julia and the extensive ecosystem for scientific computing.
There exist many physical systems that show learning behaviour in response to external stimuli spanning large ranges of complexity, from slime moulds to the human brain. These systems are generally able to learn complex tasks at a fraction of the energy required by modern machine learning algorithms, simply by following a trajectory imposed by physical laws. As such, investigating the constraints and possibilities that arise from these systems existing in a physical world, compared to their in-silico counterparts which are (only) limited by the properties of the hardware they are implemented on, may open an avenue to new and highly energy-efficient computing and learning paradigms.
We introduce a system of two-dimensional, coupled nonlinear resonators, exploiting naturally occurring degeneracies to store short- and long-term memory, the two main ingredients required for learning, in its physical dynamics. By carefully choosing the couplings between the resonators, we are able to create a feedback loop between short- and long-term memory, which causes the system to learn.
In this poster, we show how we use Julia to bring together many different packages from different corners of the ecosystem to study and optimise the learning dynamics of this metamaterial across many levels of abstraction, from stochastic differential equations over dynamical systems theory to Markov chains. We argue that this ability to easily look at the problem from many perspectives, combined with Julia's general speed is highly beneficial to computationally study systems that might not be optimised to run efficiently on classical computers, but in the real, physical world.
I was trained as a theoretical physicist and complex systems researcher, with a heavy focus on interdisciplinary research. After working mostly on the complexity of human mobility, I now study ways in which we can use physics to build learning metamaterials.