Photonic integrated circuits, or optical chips, have the potential to offer many advantages over their electronic counterparts, such as lower power consumption and reduced latency in computing. That is why many researchers believe that they can be extremely effective in the tasks of machine learning and the creation of artificial intelligence (AI). Intel also sees great prospects for the use of silicon photonics in this area. A group of researchers in described in detail new methods that can bring optical neural networks one step closer to reality.
In recent , dedicated to machine learning, tells how research in the field of optical neural networks began. The scientific work of David AB Miller and Michael Reck has demonstrated that a type of photonic circuit known as a Mach-Zehnder interferometer (MZI) can be configured to perform 2 Γ 2 matrix multiplication, whereby, if placed MZI in a triangular grid for multiplying large matrices, you can get a circuit that implements the matrix-vector multiplication algorithm - the main calculation used in machine learning.
Intel's new research focused on what happens when various defects that optical chips are exposed to during manufacturing (because computational photonics is analog in nature) cause differences in computational accuracy between different chips of the same type. Although similar studies have already been carried out, in the past they have focused more on post-manufacturing optimization to eliminate possible inaccuracies. But this approach has poor scalability as networks become larger and larger, resulting in more processing power required to set up optical networks. Instead of post-manufacturing optimization, Intel looked at the possibility of one-time training of chips before fabrication through the use of a noise-tolerant architecture. The reference optical neural network was trained once, after which the training parameters were distributed over several fabricated network instances with differences in their components.
The Intel team considered two architectures for building artificial intelligence systems based on MZI: GridNet and FFTNet. GridNet places MZIs in a predictable grid, while FFTNet places them in a butterfly pattern. After training both in handwritten digit recognition deep learning reference task (MNIST) simulations, the researchers found that GridNet achieved higher accuracy than FFTNet (98% versus 95%), but the FFTNet architecture was "significantly more robust." In fact, the performance of GridNet fell below 50% with the addition of artificial noise (interference that mimics possible defects in the production of optical chips), while for FFTNet it remained almost constant.
The scientists say their research lays the groundwork for AI training methods that can eliminate the need to fine-tune optical chips after they are manufactured, saving valuable time and resources.
βAs with any manufacturing process, certain defects will occur, which means that there will be small differences between the chips, and these will affect the accuracy of the calculations,β writes Casimir Wierzynski, senior director of the Intel AI Product Group. βIf optical neural essences become a viable part of the AI ββhardware ecosystem, they will need to move to larger chips and industrial manufacturing technologies. Our research shows that choosing the right architecture upfront can significantly increase the likelihood that the resulting chips will achieve the desired performance, even in the presence of manufacturing variations.β
While Intel is primarily doing research, MIT Ph.D. Yichen Shen founded Boston-based startup Lightelligence, which has raised $10,7 million in venture capital funding and a prototype optical chip for machine learning, which is 100 times faster than modern electronic chips, and also reduces power consumption by an order of magnitude, which once again clearly demonstrates the promise of photonic technologies.
Source: 3dnews.ru