Periodic Reporting for period 1 - INSPIRE (Three dimensional INtegrated PhotonIcS to RevolutionizE deep Learning)
Reporting period: 2022-12-01 to 2025-05-31
Neural networks are artificial intelligence (AI) computing systems inspired by the intricate connectivity and processing power of the human brain. They already surpass human capabilities in certain tasks and are widely used in applications such as pattern recognition, classification, and optimisation. With the potential to reshape society, neural networks have become central to major technological strategies, as seen in Google’s "AI-first" approach, and are recognised by governments and policymakers worldwide as a key future technology.
However, despite their promise, existing neural network architectures suffer from fundamental inefficiencies, particularly when scaled to large and complex systems. Current hardware operates far below theoretical limits, posing a significant bottleneck to future AI advancements. Overcoming this challenge requires a paradigm shift in how neural networks are designed and implemented.
The EU-funded INSPIRE project will introduce a groundbreaking approach by harnessing advanced photonic integration. By leveraging three-dimensional photonic waveguides, **INSPIRE** will develop a biologically inspired, fully parallel, and highly scalable architecture. This pioneering technology will unlock unprecedented computational efficiency, paving the way for next-generation AI systems with orders of magnitude greater performance.
However, despite their promise, existing neural network architectures suffer from fundamental inefficiencies, particularly when scaled to large and complex systems. Current hardware operates far below theoretical limits, posing a significant bottleneck to future AI advancements. Overcoming this challenge requires a paradigm shift in how neural networks are designed and implemented.
The EU-funded INSPIRE project will introduce a groundbreaking approach by harnessing advanced photonic integration. By leveraging three-dimensional photonic waveguides, **INSPIRE** will develop a biologically inspired, fully parallel, and highly scalable architecture. This pioneering technology will unlock unprecedented computational efficiency, paving the way for next-generation AI systems with orders of magnitude greater performance.
the ERC INSPIRE (Three-dimensional INtegrated PhotonIcS to RevolutionizE Deep Learning) Consolidator Grant project, has achieved significant milestones in integrating three-dimensional photonic structures with deep learning applications. Results of the project are posted on the project website https://projects.femto-st.fr/INSPIRE/en(opens in new window).
During the first reporting period the ERC INSPIRE Consolidator Grant project has delivered transformative research and technological advancements, pioneering three-dimensional photonic integration, has substantially outperformed traditional electronic counterparts in various metrics. We have enabled real-time processing capabilities for complex machine learning with 10^10 inferences per second, and made substantial breakthroughs in training unconventional neuromorphic computing substrates.
During the first reporting period the ERC INSPIRE Consolidator Grant project has delivered transformative research and technological advancements, pioneering three-dimensional photonic integration, has substantially outperformed traditional electronic counterparts in various metrics. We have enabled real-time processing capabilities for complex machine learning with 10^10 inferences per second, and made substantial breakthroughs in training unconventional neuromorphic computing substrates.
The breakthroughs we have made in model free optimization do exceed initial expectation, in particular the early stage where they demonstrate their effectiveness. Furthermore, we have carried out substantial memory-load analysis comparing the current state-off-the-art to training unconventional neural networks following the digital twin approach . What this shows is that (i) model free optimization has the edge in terms of computational overhead while being on par in terms of accuracy, and (ii) most importantly that the training effort using the digital twin approach scales O(N4), while model free scales with O(N2) with N as the number of neurons.
Using the large-scale multimode vertical-cavity surface-emitting lasers (LA-VCSEL) with chaotic cavity geometry we have been able to push the number of ‘neurons’ implemented in a single LA-VCSEL to ~10.000! What this means in numbers is truly astonishing and revolutionary: we implement 10.000 neurons (demonstrated experimentally) in a single CMOS compatible device that consumes around 100 mW (demonstrated experimentally), that provides recurrent connections fully in parallel as well as neurons that have a transient time of around 0.5 ns (demonstrated experimentally), creating a large neural network providing 1010 results per second.
Using the large-scale multimode vertical-cavity surface-emitting lasers (LA-VCSEL) with chaotic cavity geometry we have been able to push the number of ‘neurons’ implemented in a single LA-VCSEL to ~10.000! What this means in numbers is truly astonishing and revolutionary: we implement 10.000 neurons (demonstrated experimentally) in a single CMOS compatible device that consumes around 100 mW (demonstrated experimentally), that provides recurrent connections fully in parallel as well as neurons that have a transient time of around 0.5 ns (demonstrated experimentally), creating a large neural network providing 1010 results per second.