Could you please introduce yourself and your area of research?
Francesco Da Ros: I am an associate professor at the Technical University of Denmark, focusing on the synergies between machine learning and photonics.
I both work to solve photonic problems with machine learning – modelling, inverse design, and in-situ and in-silico optimization/training of photonic systems – and to investigate photonic computing hardware for supporting machine learning applications – photonic reservoir computing, photonic matrix-vector multipliers etc.
Tell me a bit about photonics – how would you describe the field to an outsider?
Francesco Da Ros: Photonics is a field in which physics and engineering come together to understand the natural phenomena related to light and its interaction with matter and/or other particles.
More specifically, a major subfield of photonics deals with using light to transmit and process information, e.g. through optical communication systems and networks which supply the backbone of our whole communication infrastructure. Such infrastructure relies on the usage of low-loss optical fibres to transmit the information, as well as on photonic integrated circuits (PICs) to provide transmitters and receivers by converting the information from the electronic domain onto the optical carriers.
More recently, optical communication is showing advantages over e.g. electronics in shorter and shorter distances, down to the communication within the same integrated circuit. This increased coverage of optical interconnectivity is especially driven by computing applications, e.g. machine learning.
Why might photonics be a good choice for machine learning?
Francesco Da Ros: Photonics has demonstrated its utmost superiority when it comes to transmitting information over long distances and at high data rates. Currently, its dominance is progressing even at shorter distances, down to interconnections within the same integrated circuit. Machine learning suffers from the data transfer bottleneck caused by the separation of memory and computing units within a conventional von Neumann architecture, i.e. the architecture used in the vast majority of digital computing hardware.
As ML algorithms require a higher level of data transfer between memory and processing units than traditional computing algorithms, this bottleneck causes slower computations and higher energy consumption. Photonic computing may provide substantial advantages by considering some of the key features it already effectively provides for optical communication: low-loss propagation, broad bandwidth, and multiplexing (frequency, time, mode, etc.).
In the short term, optical interconnections, i.e. linear pluggable optics, co-packaged optics, etc. are actively investigated to provide substantially higher connectivity within the computing chips by e.g. decreasing latency and increasing throughput.
Mid-term, specific operations may be more efficient to be implemented in photonics, e.g. matrix-vector multiplications (MVMs) for in-memory computing can be implemented at lower energy consumption in PICs, thanks to (1) the ability to fan-in (summing) and fan-out (splitting) effectively in the optical domain, (2) the low-loss waveguide crossing allowing for compact designs, and (3) recent progress on programmable PICs.
Finally, longer-term, additional features of photonics, e.g. its slow and fast nonlinear dynamics allowing for tuning the memory of the computing system, may provide the tools to implement both linear and nonlinear accelerators increasing the role of photonics within ML hardware.
Are there certain tasks (or applications) where photonics might perform better than other technologies?
Francesco Da Ros: Aside from providing interconnectivity and specific acceleration (e.g. MVMs) mentioned above, using photonic for computing/processing data would allow for seamless integration with applications in which data is already in the photonic domain, e.g. optical communications, fibre-based or free-space optics (e.g. Lidar) sensing.
Such a choice could potentially allow for decreasing the number of domain conversions, e.g. through analog-to-digital/digital-to-analog, as well as electro-optic and optoelectronic conversions, which currently limit the transmission rates and introduce substantial energy consumption.
What are the primary difficulties in researching and implementing photonic chips?
Francesco Da Ros: Research on PICs is mainly limited but [there is an issue with the] relatively lower technological maturity of PICs compared to electronic integrated circuits (EICs). Furthermore, and related to that, fabrication runs are costly in terms of actual financial resources and time for a tape-out from submitting a design to receiving the PIC. Multi-project wafer runs and the increasing number of smaller foundries (chip manufacturing facilities) offering fabrication services over the last few years improved this challenge, but difficulties remain.
What are the major challenges of implementing ML on photonic hardware?
Francesco Da Ros: Switching from digital hardware with its well-defined programming rules and inherently higher tolerance to noise and distortion, to analog hardware brings several challenges:
- How to accurately train the photonic hardware,
- what is the best material platform for such hardware
- how to enhance the control of noise.
Training photonic computing circuits is a very active research field with several interesting methods reported, however, none can yet rival the tools at the disposal of ML on digital hardware. This is a very exciting topic that requires a strong collaboration between experts in photonics, machine learning, and possibly even neuroscientists if inspiration from learning in biological systems is considered.
The are many material platforms being investigated that bring different advantages in terms of integration with electronics, robustness, high-energy efficiency, etc. but there is yet no clear winner and hybrid material platforms will be likely needed.
Noise can be a challenge in analog systems but can also be turned into an advantage for probabilistic ML systems that rely heavily on stochastic behaviour (randomness), such as generative/diffusion models.
How would photonic chips interact with existing hardware and software?
Francesco Da Ros: For them to be effective, photonic chips need to be fully integrated with existing chips and the software they are running. As such, a substantial research effort is now being dedicated to both photonic-electronic co-integration (even by major industrial players like TSMC), and software-hardware co-design. Photonics are not expected to replace existing integrated circuits but rather to support them in the operations that are more effectively performed in the photonic domain, e.g. data transfer, or even in-memory computing.
Envisioning the near future, how do you think the field will evolve?
Francesco Da Ros: The field of photonic computing is already evolving towards closer interaction with researchers working on the theory of ML and electronic hardware. Such an interaction will necessarily increase over time, especially once the current challenges of defining a common language across disjoint disciplines are fully addressed.
Practically, the deployment of optical interconnects for intra-chip communications are already being actively discussed. Linear and nonlinear accelerators are also active research directions, where likely the nonlinear hardware will be a hybrid combination of analog photonic and electronics linked with digital electronics.
What support do you think the field needs now to achieve its goals?
Francesco Da Ros: Access to foundries (specialised manufacturing facilities) and effective collaboration with computer scientists and electronic hardware experts are perhaps the two major current challenges.
Along the first direction, the European Chips Act provides good support but there is still a general shortage of researchers/engineers skilled in PIC designs, and fabrication runs can still require up to a year from design to delivery of packaged PIC.
Along the second direction, dedicated funding support could be highly beneficial. EU Programs like MSCA doctoral networks reward interdisciplinary but are limited in research resources (e.g. to cover for technology development or intensive fabrication at foundries). EU programs like RIAs can be more focused on technological innovation but are still limited by the number of fabrication runs that can be realistically included within a project timeframe.