Photonic integrated circuits or optical chips potentially have many advantages over electronic counterparts, such as reducing power consumption and reducing computational delay. That is why many researchers believe that they can be extremely effective in problems of machine learning and the creation of Artificial intelligence (AI). Intel also sees great promise for applying silicon photonics in this direction. A group of its researchers in a scientific article described in detail the new methods that can bring optical neural networks one step closer to reality.
A recent Intel machine learning blog post tells how research began in the field of optical neural networks. The scientific work of David Miller (David AB Miller) and Michael Reck (Michael Reck) demonstrated that the type of photon chain, known as the Mach-Zehnder interferometer (MZI), can be configured to perform multiplication of 2 × 2 matrices; MZI in a triangular grid for multiplying large matrices, you can get a scheme that implements the matrix-vector multiplication algorithm – the basic calculation used in machine learning.
Intel’s new study focused on studying what happens when various defects appear that optical chips are exposed to during production (since computational photonics is analog in nature), causing differences in the accuracy of calculations between different chips of the same type. Although such studies have already been carried out, in the past they have focused more on post-production optimization to eliminate possible inaccuracies. But this approach has poor scalability as the networks become larger in size, which leads to an increase in the computing power necessary to tune the optical networks. Instead of optimizing after fabrication, Intel considered the possibility of one-time learning of chips before fabrication through the use of an interference-resistant architecture.
The Intel team reviewed two architectures for building artificial intelligence systems based on MZI: GridNet and FFTNet. GridNet predictably places MZI in the grid, and FFTNet places them in the form of butterflies. After learning both in the modeling on the reference problem of deep learning to recognize handwritten numbers (MNIST), the researchers found that GridNet achieved higher accuracy than FFTNet (98% versus 95%), but the FFTNet architecture turned out to be “much more reliable.” In fact, the GridNet performance dropped below 50% with the addition of artificial noise (noise simulating possible defects in the production of optical chips), while for FFTNet it remained almost constant.
Scientists claim that their research lays the foundation for artificial intelligence teaching methods that will help get rid of the need to fine-tune optical chips after their production, saving valuable time and resources.
“ As in any production process, certain defects arise which mean that there will be small differences between the chips and they will affect the accuracy of the calculations ,” writes the senior director of the Intel AI product group, Casimir Wierzynski. “ If optical neural essences become a viable part of the hardware ecosystem of artificial intelligence, they will need to switch to larger chips and industrial production technologies. Our research suggests that choosing the right architecture in advance can significantly increase the likelihood that the resulting chips will achieve the desired performance even with production variations . ”
At the same time, while Intel is mainly conducting research, Ph.D. in Physics and Mathematics from Massachusetts Institute of Technology Ishen Shen (Yichen Shen) founded a startup Lightelligence based in Boston, which attracted venture capital funding of $ 10.7 million and recently demonstrated a prototype of an 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 prospects of photon technologies.