Symposium on the Physics of Information in Matter

Arvind Murugan

Learning without neurons

When we think of learning, we think of neural networks, i.e., networks of linear threshold devices. Even physical realizations of neural computation, such as molecular or electrical circuits, often mimic these network architectures at an element-by-element level. Here, we explore an alternative paradigm – inevitable collective physical processes that can learn in Hebbian-inspired ways without being designed to mimic a neural network element-by-element. Through theory and experiment, we show how nucleation of crystals and bifurcations in elastic materials can learn to recognize complex patterns in chemical or mechanical stimuli. Our work suggests that ubiquitous and inevitable physical phenomena, such as nucleation, might be a powerful underexplored route to neural computation without neurons.

Arvind Murugan obtained his BS in mathematics from Caltech, PhD in high energy physics from Princeton University. He joined the physics faculty at the University of Chicago after postdoctoral research at the Institute for Advanced Study (Princeton) and Harvard University. The major thrust of his research is how physical and biological systems can learn from their environmental history and show neural network-like behavior. His current work focuses on learning in molecular self-assembly and mechanical systems and on the biological side, the evolution of molecular error correction and evolution in changing environments. He is currently a Simons Investigator in the Mathematical Modeling of Living Systems.

Patty Stabile

InP Photonic Integrated Neural Networks

Artificial neural networks have been employed in a plethora of applications. However, the extraction of meaningful information requires enormous power and processing time when these models run on conventional electronics. Neuromorphic photonics is an emerging research field that develops an alternative approach to electronics with the attempt to set a milestone in increasing computing speed and energy efficiency using photons, instead of electrons. We map artificial neural models on chip to realize photonic integrated neural networks, using a combination of semiconductor optical amplifiers and array waveguide gratings technology. An all-optical monolithically integrated neuron and a multi-layer optical neural network are demonstrated. An in-depth scalability analysis and projection open to the possibility to develop all-optical neural networks on-chip. The best-in-class technologies are finally identified and a new concept for a 3D neuron and neural network is shown, together with predicted performance, opening to a promising and feasible technology for neuromorphic photonics.

Patty Stabile is Associate Professor at TU/e within the Electro-Optical Communication system group. She moved to TU/e in 2009 for working on large-scale photonic integrated circuits, for which she was awarded the “Early Career Women in Photonics Special Recognition” from the European Optical Society. She is co-applicant within several European Projects (e.g. Passion, Twilight, …) and PI within large national (e.g. Gravitation program) and international (e.g. KDT) programs. She is author and co-author of more than 150 papers and chair of the IEEE Photonics Benelux Chapter. She has been visiting scientist at University of Cambridge in UK and at MIT.

Corentin Coulais

Corentin Coulais

Reinventing the wheel

Controlling how waves propagate, attenuate and amplify in simple, scalable geometric structures is a daunting challenge for science and technology. In this talk, I will discuss how odd active matter—in which microscopic interactions are asymmetric and non-conservative—can be used to steer mechanical waves in unprecedented ways. Combining experiments active mechanical metamaterials with wave physics and continuum and topological mechanics, I will discuss the emergence of the non-Hermitian skin effect, of odd elasticity and of one-way solitons. I will further show how these odd waves can be used to induce locomotion and unusual responses to impacts and hence pave the way towards a novel generation of robots.

Corentin Coulais is Associate Professor at the University of Physics of the University of Amsterdam. Coulais’ Machine Materials group investigates designer soft materials, with a particular emphasis on how mechanical metamaterials can programmed to achieve advanced mechanical tasks. He has recently pioneered robotic materials, which combine the notions of emergence and symmetries inherent to condensed matter with the capabilities of robotics. This has led to early experimental observations of non-Hermitian wave phenomena such as unidirectional amplification (2019), non-Hermitian topology (2020) and odd elasticity (2021). Coulais has received the NWO VENI (2015), ERC Starting (2019), NWO VIDI (2022) and leads multiple collaborations with industry.

Hans Hilgenkamp

Routes for storing and processing information in matter, towards more energy-efficient information technologies

Information technologies have developed tremendously in the last decades, with down scaling of key hardware components like transistors, according to Moore’s law, as the leading paradigm. Speed and data-volumes were the main drivers. Meanwhile the focus of attention has shifted to the energy consumption of information technologies, which is presenting problems both from the viewpoints of sustainability and available infrastructure, as well as for desired functionality.

Various routes are being explored worldwide to process and store information in more energy-efficient ways, ideally circumventing the dissipative transfer of information as much as possible.

I will discuss several concepts that are currently being employed and will introduce research we do at the BRAINS Center in Twente on this topic, including dopant network processing units and Mott insulator-to-metal [2] based systems.

[1]: T. Chen, …, W.G. Van der Wiel, et al., Nature 577, 7790 (2020).
[2]: X. Gao, C.M.M. Rosário and H. Hilgenkamp, AIP Advances 12, 015218 (2022).

Hans Hilgenkamp completed his M.Sc. in Applied Physics (1991) and Ph.D (1995) at the University of Twente. He then worked at the IBM Zurich Research Lab and the University of Augsburg. In 2000 he returned to Twente, where he is a PI in the Nanoelectronic Materials cluster. In 2014 – 2018 he was dean of the Faculty of Science and Technology, after which he became co-director of the UT-BRAINS Center. He has co-founded the Global Young Academy, is member of the Board of Reviewing Editors of Science and of the Dutch Physics Council (DPC), vice-chair of the Netherlands’ Physical Society (NNV), and Fellow of the American Physical Society.

Marc Serra-Garcia

Nanomechanics as an emergent platform for information technologies

When thinking about information technologies, the first things to come to mind are electronics, fiber optics or magnetic disks; rarely one things of mechanical systems. Mechanical systems, however, have very interesting properties such as near-zero energy losses and strong geometric nonlinearities, that make them appealing for low-power computing applications. In this talk, I will discuss two research directions connecting information and mechanics. First, I will talk about our work in implementing zero power speech recognition, a technology aimed at building battery-less smart devices. Second, I will talk about our research in the fundamental energy limits of computing, where we use nanomechanical systems to discover the absolute minimum amount of energy necessary to perform a computation.

Marc Serra-Garcia is tenure-track group leader of the Hypersmart Matter group at AMOLF, working at the intersection between mechanics, information techologies and thermodynamics. His group combines theory and experiment to design novel information-processing devices, while at the same time asking fundamental questions such as ‘what is the minimum amount of energy necesary to perform a computation?’ and ‘How can intelligent behavior emerge from a large number of simple elements?‘. Before AMOLF, he was a PhD student and postdoc at ETH Zurich, and a master student at Caltech. His research has been recognised by an ERC starting grant, and resulted in a spin-off company at ETH.

Elisabetta Chicca

Spike-based local synaptic plasticity models and BEOL compatible memristive devices

Synaptic plasticity is considered to be the basis of learning and memory in the brain. It goes from low level task-specific learning to high level cognition. Understanding the computational foundations of synaptic plasticity is therefore a growing research that inspires progress in the design of autonomous adaptive systems. In that perspective, a large number of brain-inspired learning rules have been modeled and implemented. Locality, a fundamental computational principle of biological synaptic plasticity, is a key requirement for physical implementation of learning rules. In this talk we provide an overview of models and circuits for spike-based local synaptic plasticity. This overview provides the background for presenting our recent work aimed at the implementation of learning systems based on CMOS and BEOL compatible memristive devices.

Elisabetta Chicca is a full professor at the Zernike Institute for Advanced Materials (ZIAM) and the Groningen Cognitive Systems and Material Center (CogniGron), University of Groningen, The Netherlands. She is Chair of Bio-Inspired Circuit and Systems (BICS) since 2020. Her current interests are in the development of CMOS models of cortical circuits for brain-inspired computation, learning in spiking CMOS neural networks and memristive systems, bio-inspired sensing (vision, olfaction, audition, touch) and motor control. She combines these research approaches with the aim of understanding neural computation by constructing behaving agents which can robustly operate in real-world environments.

Eliška Greplová

Multiverse of engineered quantum topology

The field of condensed matter physics is currently being transformed by a series of exciting theoretical discoveries of intriguing properties of quantum materials and remarkable experimental progress that allows us to test the novel theories contemporaneously. In this talk, I will illustrate this condensed matter renaissance with two examples that also connect condensed matter to the two emerging fields of artificial intelligence and quantum computing. First, I will discuss a top-down example that considers solving existing models with the aid of artificial intelligence tools. In a second, bottom-up approach, I will discuss on-chip engineering of topological features using scalable quantum computing building blocs.

Eliška Greplová is an assistant professor at Kavli Institute of Nanoscience at TU Delft in the Netherlands, where she leads “Quantum Matter and AI” lab. Eliska is also a visiting researcher at Microsoft Research Amsterdam. Eliska completed her PhD at Aarhus University, Denmark, and postdoctoral fellowship at ETH Zurich, Switzerland.

Greg Stephens

From partial observations to long timescales through maximally predictive states

In many physical settings, isolating slower dynamics from fast fluctuations has proven remarkably powerful. But how do we proceed in complex systems when the underlying equations are unknown or uninformative? Here, we integrate information theory, dynamical systems and statistical physics to extract understanding directly from partial, dynamical, measurements. We introduce maximally predictive states, a symbolic, delay-embedded encoding constructed to maximize short-time predictive information. Transitions between these states yield a simple approximation of the transfer operator, which we use to reveal timescale separation and long-lived collective modes. Applicable to both deterministic and stochastic systems, we illustrate our approach through the the Lorenz system and the thermal dynamics of a particle in a double-well potential. Applied to the behavior of the nematode worm C. elegans, we bridge sub-second posture fluctuations and long range effective diffusion in foraging behavior, recovering the “ballistic-to-diffusive” transition in the worm’s centroid trajectories. Finally, we use our operator perspective to reveal long-lived “run-and-pirouette” behaviors, and to predict additional subtle subdivisions of the worm’s “run'” dynamics.

Greg Stephens has a background in theoretical physics and is an associate professor in the department of physics and astronomy at the VU Amsterdam and an adjunct professor at OIST Graduate University in Japan. Work in his group is helping to pioneer a new field – the physics of behavior: from individual organisms to entire societies. While the science of the living world is mostly focused on the microscopic, such as the expression of genes or the pattern of electrical activity in our brains, all of these processes serve the greater evolutionary goals of the organism: to find food, avoid predators and reproduce. This is the behavioral scale, and despite it’s importance, our quantitative understanding of behavior is much less advanced. But how do we quantify the emergent dynamics of entire organisms? What principles characterize living movement? The Stephens group addresses these questions with a modern biophysics approach combining ideas from statistical physics, information theory and dynamical systems with model systems ranging from the nematode C. elegans to zebrafish and honeybee collectives.