Photonic integrated circuits or optical chips potentially offer numerous advantages over their electronic counterparts, such as reduced power consumption and lower latency in computations. This is why many researchers believe they could be extremely effective in machine learning tasks and artificial intelligence (AI) development. Intel also sees significant potential for the application of silicon photonics in this area. Its research team has detailed new methods that could bring optical neural networks a step closer to reality.
In a recent , focused on machine learning, it explains how research in optical neural networks began. The scientific work of David A. B. Miller and Michael Reck demonstrated that a type of photonic circuit known as a Mach-Zehnder interferometer (MZI) can be configured to perform 2 × 2 matrix multiplications; furthermore, if the MZI is placed in a triangular grid for multiplying larger matrices, it can create a circuit that implements the matrix-vector multiplication algorithm—a fundamental computation used in machine learning.
A new study by Intel focused on understanding what happens when various defects occur in optical chips during production (as computational photonics is inherently analog), causing differences in computational accuracy between different chips of the same type. Although similar studies have been conducted in the past, they primarily focused on post-manufacturing optimization to eliminate potential inaccuracies. However, this approach has poor scalability as networks increase in size, leading to a rise in the computational power required for tuning optical networks. Instead of post-manufacturing optimization, Intel explored the possibility of one-time training of chips before manufacturing using a noise-resistant architecture. A benchmark optical neural network was trained once, after which the learning parameters were distributed across several fabricated instances of the network with variations in their components.
The Intel team examined two architectures for building AI systems based on MZI: GridNet and FFTNet. GridNet predictably arranges MZIs in a grid, while FFTNet places them in a 'butterfly' configuration. After training both models on a benchmark deep learning task of handwritten digit recognition (MNIST), researchers found that GridNet achieved higher accuracy than FFTNet (98% vs. 95%), but the FFTNet architecture proved to be 'significantly more robust.' In fact, the performance of GridNet dropped below 50% with the addition of artificial noise (imitating potential defects in the production of optical chips), while the performance of FFTNet remained almost constant.
Scientists assert that their research lays the groundwork for AI training methods that will eliminate the need for fine-tuning optical chips after production, saving valuable time and resources.
Just like any manufacturing process, certain defects arise that indicate minor differences between chips, which can affect computational accuracy," says Intel AI Senior Director Casimir Wierzynski. "If optical neural cores are to become a viable part of the AI hardware ecosystem, they will need to transition to larger chips and industrial-grade manufacturing technologies. Our research suggests that selecting the right architecture in advance can significantly increase the likelihood that the resulting chips will achieve the desired performance, even in the presence of manufacturing variations.
At the same time, while Intel is primarily conducting research, PhD candidate Yichen Shen from the Massachusetts Institute of Technology founded the Boston-based startup Lightelligence, which has raised $10.7 million in venture funding and a prototype optical chip for machine learning that is 100 times faster than current electronic chips and significantly reduces power consumption, further showcasing the potential of photonic technologies.
Source: 3dnews.ru
