Periodic Reporting for period 1 - ARIADNE (Artificial intelligence enabled by automatic dynamic exploration of integrated photonic spiking neural networks)
Berichtszeitraum: 2023-04-01 bis 2025-03-31
Machine learning (ML) and artificial intelligence (AI) have shown incredible usefulness across science and technology, transforming markets and society. However, their increasing adoption and development come with significant energy costs, which undermines their sustainability. What's more, their current applications barely scratch the surface of their full potential. One way to boost AI's energy efficiency, speed, and overall sustainability and applicability is to host ML models on dedicated, non-digital hardware, like hardware neural networks with physical artificial neurons and connections, instead of simulating them on conventional digital computers.
For example, processing optical data (like the internet data carried by light in fiber optics) with a photonic neural network keeps the processing in the optical domain. This avoids the energy consumption, latency, and information loss that happen when data is converted and digitized. Relevant examples of optical data include telecom signals through optical fibers and signals from numerous optical sensing applications, such as for biomedical or environmental monitoring. However, while photonic circuits are generally great for efficient linear operations and data transfer, they face challenges in performing the nonlinear operations and long-term memory tasks crucial for many neural network and ML implementations.
ARIADNE has developed ML applications using a new type of on-chip photonic neural network. This network leverages highly complex optical dynamics within a system of microresonators. This system allows us to process optical data for ML, including both nonlinear operations and long-term memory, in a fast and energy-efficient way and it finds application in enhancing a wide variety of optics-based technologies.
For example, processing optical data (like the internet data carried by light in fiber optics) with a photonic neural network keeps the processing in the optical domain. This avoids the energy consumption, latency, and information loss that happen when data is converted and digitized. Relevant examples of optical data include telecom signals through optical fibers and signals from numerous optical sensing applications, such as for biomedical or environmental monitoring. However, while photonic circuits are generally great for efficient linear operations and data transfer, they face challenges in performing the nonlinear operations and long-term memory tasks crucial for many neural network and ML implementations.
ARIADNE has developed ML applications using a new type of on-chip photonic neural network. This network leverages highly complex optical dynamics within a system of microresonators. This system allows us to process optical data for ML, including both nonlinear operations and long-term memory, in a fast and energy-efficient way and it finds application in enhancing a wide variety of optics-based technologies.
ARIADNE pursued two primary goals through its research and development activity. First, it focused on creating effective and computationally efficient machine learning (ML) algorithms and strategies designed to leverage the unique capabilities of the considered photonic neural network. It conducted proof-of-concept experiments to assess the full ML system's computational performance, testing it on widely used ML benchmarks like image classification and time series classification and prediction. The system showed outstanding performance compared to other state-of-the-art photonics-based ML systems. Crucially, it achieved this while reducing the number of required trainable parameters and digital operations compared to similarly performing conventional artificial neural networks.
The second main objective involved experimentally demonstrating practical applications that rely on optical signal processing. To achieve this, a significant hurdle in photonic computing was overcome: establishing an energy-efficient, long-term (much longer than microseconds), resettable optical memory effect. This was vital for processing time-dependent optical signals from sensing applications, where sensed quantities often change over milliseconds to seconds. The successful demonstration of ML-exploitable long-term memory within their photonic neural network opens the door for a vast array of new applications in optical smart sensing, which we plan to explore in future research. Additionally, experiments (which are ongoing after ARIADNE's formal end) applied the ML system to data from a fiber sensor capable of detecting mechanical perturbations along its entire length, yielding promising results.
The second main objective involved experimentally demonstrating practical applications that rely on optical signal processing. To achieve this, a significant hurdle in photonic computing was overcome: establishing an energy-efficient, long-term (much longer than microseconds), resettable optical memory effect. This was vital for processing time-dependent optical signals from sensing applications, where sensed quantities often change over milliseconds to seconds. The successful demonstration of ML-exploitable long-term memory within their photonic neural network opens the door for a vast array of new applications in optical smart sensing, which we plan to explore in future research. Additionally, experiments (which are ongoing after ARIADNE's formal end) applied the ML system to data from a fiber sensor capable of detecting mechanical perturbations along its entire length, yielding promising results.
ARIADNE produced the following advancements with respect the state of the art:
1. Photonic neuromorphic processing with volatile energy-efficient memory of extended duration. Direct experimental demonstrations showed a ML-exploitable memory duration of several tens microseconds, which is approximately a two order of magnitude improvement with respect to the state of the art [A Lugnan et al., Advanced Optical Materials 13 (11), 2403133]. However, this memory effect, which is based on information being stored into the self-sustained dynamics of our photonic neural network, is expected to have an even longer duration, as it lasts in principle forever under ideal conditions. This represents a breakthrough for ML-based optical computing, since it allows processing of relatively slow signals from optical sensors, which otherwise would require the employment of digital electronic memory, with a related increase in energy cost, complexity and latency.
2. Record high throughput generation of optical nonlinear representations exploitable for ML, by exploiting the space, time and wavelength dimensions. Our photonic neural network can produce hundreds of different transformations of input images in parallel, with a speed of around 10 ns per input pixel, consuming only a few milliwatts of power. These output representations were employed by our lightweight ML model achieving high accuracies in handwritten digits classification (97.5% for the MNIST dataset) and fashion images classification (90.6% for the Fashion-MNIST dataset). These accuracy levels are obtained with a significant increase in computational efficiency (in terms of digital operations and trainable parameters) with respect to conventional artificial neural networks hosted in digital electronics.
3. First application of a photonic neural network with memory to the data from an optical fiber sensor, where the sensed information can be stored in the network dynamical memory and retrieved by a simple machine learning system. This experiment, which is ongoing, demonstrates that the dynamical memory of our photonic network is sensitive and stable enough for practical applications in fiber sensing.
These advancements have shown that ARIADNE's photonic neural network's memory, computational power, stability and sensitivity are suitable for applications in real neuromorphic computing technologies. Further research is needed to fully demonstrate applications in fiber sensing and smart neuronal interfaces (where a photonic neural network communicates with light sensitive biological neurons). The Nanoscience Laboratory at University of Trento and its current collaborations have the necessary expertise to significantly progress along these research and development paths.
1. Photonic neuromorphic processing with volatile energy-efficient memory of extended duration. Direct experimental demonstrations showed a ML-exploitable memory duration of several tens microseconds, which is approximately a two order of magnitude improvement with respect to the state of the art [A Lugnan et al., Advanced Optical Materials 13 (11), 2403133]. However, this memory effect, which is based on information being stored into the self-sustained dynamics of our photonic neural network, is expected to have an even longer duration, as it lasts in principle forever under ideal conditions. This represents a breakthrough for ML-based optical computing, since it allows processing of relatively slow signals from optical sensors, which otherwise would require the employment of digital electronic memory, with a related increase in energy cost, complexity and latency.
2. Record high throughput generation of optical nonlinear representations exploitable for ML, by exploiting the space, time and wavelength dimensions. Our photonic neural network can produce hundreds of different transformations of input images in parallel, with a speed of around 10 ns per input pixel, consuming only a few milliwatts of power. These output representations were employed by our lightweight ML model achieving high accuracies in handwritten digits classification (97.5% for the MNIST dataset) and fashion images classification (90.6% for the Fashion-MNIST dataset). These accuracy levels are obtained with a significant increase in computational efficiency (in terms of digital operations and trainable parameters) with respect to conventional artificial neural networks hosted in digital electronics.
3. First application of a photonic neural network with memory to the data from an optical fiber sensor, where the sensed information can be stored in the network dynamical memory and retrieved by a simple machine learning system. This experiment, which is ongoing, demonstrates that the dynamical memory of our photonic network is sensitive and stable enough for practical applications in fiber sensing.
These advancements have shown that ARIADNE's photonic neural network's memory, computational power, stability and sensitivity are suitable for applications in real neuromorphic computing technologies. Further research is needed to fully demonstrate applications in fiber sensing and smart neuronal interfaces (where a photonic neural network communicates with light sensitive biological neurons). The Nanoscience Laboratory at University of Trento and its current collaborations have the necessary expertise to significantly progress along these research and development paths.