RESEARCH SPOTLIGHT
Hi, I am
Barishan Das
HOST INSTITUTE
FEMTO-ST, CNRS
PROJECT TITLE
Optimizer Free Training of Physical Neural Networks
KEY WORDS
machine learning, reinforcement learning, neural networks, meta learning
RESEARCH AIM
My research aims to develop gradient-free learned optimization methods for neural networks, so that machine learning models can be trained in situations where gradient information is unavailable or difficult to obtain.
RESEARCH OVERVIEW
My research explores new ways of training artificial intelligence systems when traditional learning methods are not available. Most AI models learn using a technique called backpropagation, which requires detailed mathematical information about the system. However, this is often difficult in emerging technologies such as photonic and neuromorphic systems. I am developing a learning framework where a network learns to optimize itself. By studying this approach on benchmark tasks, I aim to create training methods that can eventually be applied directly to physical computing systems where conventional optimization techniques are impractical.
RESEARCH CHALLENGE
My research aims to solve the problem of how to train intelligent systems when the information normally required by standard AI training methods is unavailable. Conventional techniques such as backpropagation rely on detailed mathematical knowledge of the system being optimized, which is often difficult or impossible to obtain in emerging computing technologies such as photonic and neuromorphic hardware. The challenge is to develop learning methods such that a physical device can learn to optimize itself instead of relying on a training algorithm.
RESEARCH INNOVATION
The innovative aspect of my research is that it treats the optimization process itself as a learning problem. Even though there has been previous work in this niche, most of them rely on gradient information, which is not possible to obtain when we deal with physical systems. Instead of relying on a predefined algorithm such as gradient descent, I train an AI agent to discover how to improve another neural network through experience. This approach makes it suitable for physical computing platforms where conventional training methods are difficult to apply. The project aims to develop optimizers that can adapt to different tasks and eventually learn directly from measurements obtained from real hardware systems.
RESEARCH IMPACT
The computational cost of modern AI is concentrating capability in the hands of a few organisations able to afford large-scale digital infrastructure. Physical computing substrates offer a route to far cheaper and more efficient inference, but only if they can be trained. Making these systems trainable is a step towards AI hardware that is more widely accessible and less dependent on scarce, energy-intensive resources.
RESEARCH SUMMARY
My research focuses on developing advanced optimization and learning strategies for systems where conventional training methods are difficult to apply. The project addresses situations in which detailed internal information about a system may be unavailable, unreliable, or expensive to access, making standard gradient-based approaches less suitable. I investigate reinforcement-learning-based methods that can improve system performance through feedback from measured outputs and objective evaluations, rather than relying entirely on explicit analytical gradients. A longer-term motivation is to support learning and optimization in emerging computational platforms, including hardware-based and physics-inspired systems, where standard digital training pipelines may not directly transfer. The project aims to contribute to more flexible, adaptive, and robust optimization methods for complex real-world systems.
RESEARCH VISUAL
RESEARCH JOURNEY
RESEARCH SECONDMENT
The first secondment will be hosted by CSIC, Spain, for three months, from September to November 2026. The main objective is to prepare a dynamical oscillator scheme for the existing meta-learner implementation. The secondment is expected to provide exposure to nonlinear dynamical systems and an understanding of how they operate.
The second secondment will be hosted by HPE, Belgium, for three months, from September to November 2027. The main objective is to gain hands-on industry experience and practical exposure to an industrial environment. The specific activities and expected outcomes are yet to be determined.
LEARNING GOALS
What I am most looking forward to is the gap between simulation and physical reality. My framework was developed and validated in simulation, where a system is noiseless, stationary and perfectly observable. Real photonic systems are none of these. I want to understand, from people who work with them daily, how such systems actually behave: how they drift, how noisy a measurement really is, and which of my assumptions will not survive contact with hardware.
RESEARCH BENEFITS
I believe every group has a specific strength, and by interacting and working with them I can get hands on experience on their strengths which might go beyond the scope of my lab. Also, I believe it promotes international collaboration.
CAREER DEVELOPMENT
Working productively inside a second research group, in a different country and scientific culture, is a skill in itself and one that is difficult to acquire without doing it. The secondment also broadens my network beyond my host institution and gives me a perspective on the field that is not solely my supervisor’s: both of which matter for whatever comes after the doctorate, in academia or in industry.
MORE THAN PHD
Why did you decide to join POSTDIGITAL+ ?
My research question sits precisely between two communities: machine learning, which has developed sophisticated training methods, and photonics, which is building hardware those methods cannot train. Working on that question in isolation is difficult, because it requires fluency in both. POSTDIGITAL+ was attractive because it is explicitly built across that boundary — the network brings together groups working on substrates, architectures, training and benchmarking, with secondments and industrial contact built in rather than added on. It offered the context the problem actually needs.
What do you think is the biggest advantage of being part of the International Doctoral Network?
Interacting with international experts from the fields in an intimate environment, sharing ideas and knowledge with fellow doctoral candidates in the network
What skill have you developed the most so far?
Scientific communication: Over the first year I have had to present the same work as a technical deliverable, a conference poster, an abstract and a general-audience article, each for a different reader. Learning to change the level without distorting the science — and to be clear about what the results do and do not show — has probably improved more than any technical skill.
Critical evaluation of my own work: The hardest and most useful thing I have learned is to look for where my method breaks before someone else does: identifying the scaling limitation in my current formulation, recognising which claims my data genuinely supports, and distinguishing a result from a preliminary observation.
What are your career ambitions after completing your doctorate?
Research-oriented: I would like to continue working at the interface of machine learning and physical computing hardware, most likely through a postdoctoral position. The field is at an early stage where the fundamental questions — how learning should work in physical systems at all — are still open, and I would like to keep working on them.
Industry-oriented: I am drawn to the point where this research becomes real hardware. Photonic computing is moving from laboratory demonstration into deployment, and I would like to work in an industrial R&D setting on making such systems trainable and practical.
Open: I intend to keep working on physical and neuromorphic computing, though whether that is in an academic group or in industrial research is something I expect the next two years to clarify. What matters more to me is staying close to problems where the algorithms and the hardware have to be designed together.
What inspired you to pursue this research Area?
During my Bachelor and Master studies, I majored in Physics and minored in Data Science and Engineering, and that is when I discovered my interest in the area that lies in the intersection of physics and artificial intelligence
What surprised you most since starting your PhD?
How much of research is communication rather than computation. I expected to spend my time writing code; a great deal of it has gone into writing, figures and explaining the work to different audiences, and that effort has repeatedly clarified my own thinking.
The distance between “it works” and “it is a result.” Getting the method to train something was the easy part. Establishing that the effect is real, robust to the choices I happened to make took far longer and taught me more