RESEARCH SPOTLIGHT
Hi, I am
Karanpreet Singh
HOST INSTITUTE
Université libre de Bruxelles
PROJECT TITLE
Information Processing in Physical Systems
KEY WORDS
Unconventional Computing ; Physical Learning ; Neuromorphic Photonics ; Chemical Learning ; Bio-Inspired Computing
RESEARCH AIM
My research aims to teach physical systems to learn in their own physics so that computation stops being the exclusive privilege of the digital chip, and we can find out how far the ability to learn actually travels across substrates.
RESEARCH OVERVIEW
Your phone learns by running software on a chip. But learning does not actually need a chip - every living cell makes decisions using nothing but chemical reactions, no brain involved. I take that idea literally: I build networks of reacting molecules and train them, the way you would train a neural network, until the chemistry itself can tell things apart, even when the differences are subtle. No code runs inside. The mixture is the computer. And I am ready to ask the same of light instead of molecules, computing in whatever medium will carry the idea.
RESEARCH CHALLENGE
Some approaches to teaching physical systems to learn treat them like a ball rolling to the bottom of a bowl. But many systems operate with a continuous flow of energy or matter. Molecules keep entering, interacting, and leaving - even when their overall concentrations become steady. In these conditions, the usual training signal can point the wrong direction, causing training to reach incorrect solutions. My challenge is getting a reliable learning signal out of such messy systems.
RESEARCH INNOVATION
The main innovation in this research lies in treating the physical system as the computer itself, rather than as a stand-in for one, and then asking the question most people step around: is the ability to learn a property of the algorithm, or can almost any substrate be taught given the right conditions? So far the answer seems to lie in both. I pursue that question across both chemistry and light, rather than betting everything on a single medium.
RESEARCH IMPACT
The big impact is showing that computing can happen where it is extremely hard for a chip to reach: inside a droplet, a material or a living cell. The computation lives in the body's own chemistry - imagine steering a cell by adjusting concentrations rather than by dosing it with a drug. That is distant, but the right direction to face. On the optical side the payoff is already concrete: light carries enormous bandwidth, performs the heavy arithmetic of a neural network almost for free as it propagates, runs many wavelengths down one channel at once, and does it on a fraction of the energy a digital chip burns.
RESEARCH SUMMARY
Physical learning trains a system's own dynamics to perform a task, using local contrastive rules in place of external backpropagation. Equilibrium Propagation is one of the foremost tools, but it assumes an energy function and a symmetric response — conditions that break the moment a network is driven out of equilibrium, where the response turns asymmetric and the naive gradient becomes biased. I have applied algorithms designed for such conditions to chemical reaction networks, and I work on developing further methods for training. Trained on their reaction rates alone, these networks reach around 98% on benchmark classification, implement every two-input logic gate including the non-separable ones, and learn nonlinear decision boundaries just through a change of readout. The same framework can potentially carry to photonic hardware - nonlinear, driven, non-conservative - which the optical half of my project builds on.
RESEARCH VISUAL
RESEARCH JOURNEY
RESEARCH SECONDMENT
1. Thales Research & Technology
Country: France
Duration: 2 months
Main objectives: Work on nonlinear photonics — the physics of how light behaves in nonlinear optical systems, and how that behaviour can be turned into computation.
Skills I expect to gain: hands-on nonlinear photonics, a closer view of industrial photonics R&D, and a first test of whether the learning framework maps onto a real optical platform.
---------------------------
2. VLC Photonics
Country: Spain
Duration: 2 months
Main objectives: Learn how a computing idea becomes an actual photonic chip — the design, layout, and testing of photonic integrated circuits at industrial scale.
Skills I expect to gain: photonic integrated circuit design and characterization, and the engineering path from a concept to fabricated optical hardware.
LEARNING GOALS
I have designed and fabricated photonic chips before, so the secondments are less about first contact with hardware and more about aiming that experience at a new question. At Thales I want to get properly into nonlinear photonics - the regime where light stops being a passive carrier and starts doing something computationally interesting. At VLC I want the industrial picture of how a photonic design becomes a manufactured, tested chip at volume, which is a different discipline from building one by hand in a university cleanroom.
RESEARCH BENEFITS
My project runs on two substrates, chemistry and light, and my master's research was on the optical side. The secondments let me bring that hardware instinct to the learning framework I am building now. Thales grounds it in the physics of real nonlinear optics; VLC grounds it in how these devices are actually manufactured and tested. Together they keep the theory tethered to what silicon photonics can, and cannot, be made to do.
CAREER DEVELOPMENT
Whichever direction I take - an industry lab, academia, or eventually a venture of my own - the useful thing is fluency on both sides of the line between an idea and working hardware. I already carry the design and fabrication background; the secondments add the industrial scale and the foundry relationships that turn a promising device into a real product. Being able to stand comfortably in that gap, between the theory and the chip, is exactly the position I want to occupy.
MORE THAN PHD
Why did you decide to join POSTDIGITAL+ ?
Honestly, light is what drew me. My earlier research was on photonic hardware, and POSTDIGITAL+ was built around exactly the question I wanted to keep chasing: whether computation can be carried by light rather than by electronics. What I did not expect was that the project would also hand me chemistry. That turned out to be the good fortune of the whole thing - the same question about learning, asked of a completely different kind of matter.
What do you think is the biggest advantage of being part of the International Doctoral Network?
Collaboration is the obvious answer, so let me name the less obvious part: it is not that you collaborate, it is who you collaborate with. I am the only one here currently working on chemistry, surrounded by people building things out of light and hardware. That is a very useful arrangement - I get to try my ideas on people who come at the same problem from a different substrate, which is a standing test of whether those ideas hold up outside my own corner. Close enough, too, that this happens over coffee at a meetup rather than over email.
What skill have you developed the most so far?
Translation - and not only the French-into-English kind that Brussels keeps insisting on. I came in from the experimental, fabrication side - cleanrooms and chips - and landed in a theory group, so the skill I have built most is moving between those two languages: taking a clean piece of theory and asking what it would actually do in a physical system. The research fundamentals I brought with me; the theory-to-hardware bilingualism is the newer muscle.
What are your career ambitions after completing your doctorate?
Open, and deliberately so. I arrived aiming at an industrial research lab; working in a theory group has made academia genuinely tempting too, so I am keeping both doors open. And somewhere down the line I fully intend to build something of my own: a company, when the right idea and the right moment collide. Probably not the day after I defend, but it is on the list.
What inspired you to pursue this research Area?
A habit that started early: taking an unreasonable question seriously and following how far the physics goes. My master's was about making sound compute along the surface of a chip, and the recurring lesson was to make the substrate do the work. Chemical networks were the natural next unreasonable question; if sound can carry information, why not a soup of molecules? The substrate keeps changing; the question I am chasing does not.
What surprised you most since starting your PhD?
How readily physical learning transfers between substrates. That chemistry can compute is not new - people have been showing that for decades. What surprised me is that the machinery of learning carries over almost intact: the same training ideas built for other systems, pointed at a network of molecules, and the thing learns from examples by adjusting nothing but its reaction rates. I expected the translation to be far more painful than it was. There is still something faintly absurd about watching a set of rate constants turning into a decisive computation.