Neuromorphic computation with Photonic integrated circuits

Neural networks in the recent applications have enabled transmissive metasurface antennas, terahertz image super-resolution, resolution-enhanced plenoptic all-in-focus imaging, high-bandwidth SDM-WDM visible light communication and high-speed computer generated holography.

The Von Neumann has long been a reliable albeit inefficient computation architecture. As its strengths, it allows the design to have remarkable flexibility and area efficiency, can handle random complex deployments without the need to upscale the processor in terms of transistor count in proportion to the size and complexity of the problem.

Von Neumann architectures proved to be a good technology to piggyback on until the exponential drop in transistor sizes as predicted by Moore’s Law. Now, the advantages of the architecture which relied on implementation of reasonably complex computation with a minimum number of on-chip transistors is diminishing. The technology nodes available in the current industry allows the integration of almost infinite transistors in a chip, thus negating the biggest advantages of the Von Neumann structures and exposing its weaknesses when it comes to designing computational challenges that can only be approached by high-density architectures.

The recent AI revolution has accelerated the development of alternatives to the Von Neumann approaches. Some applications such as AIs based on convolutional neural networks are extremely computationally intensive for Von Neumann based architectures. The industry is already approaching the stagnation by hybridizing Von Neumann with non-Von Neumann architectures.

Neuromorphic computing is one of the leading alternatives to Von Neumann architectures. The spurt in miniaturization of transistors and advances in bulk CMOS fabrication technologies has reinvigorated the neuromorphic movement that was shelved in the late 1970s despite its tremendous capabilities. Neuromorphic implementations are not only of great interest in the electronic device domain but also in the field of photonics which has seen tremendous growth in recent years due to its ultra-high bandwidth and thoughput capabilities. At the same time, photonic interconnects are less prone to thermal effects and have lower losses compared to their electronic counterparts.

The implementation of a neural network in photonics requires the elements of a generalized neural network to be substituted by photonic devices that perform the same mathematical operations. A general neural network can be segmented into an input layer, a set of hidden layers and an output layer. Each of these layers are constituted by several artificial neurons that serve the purpose of an elementary computational unit in the network. The computational units are meant to perform the weighting operation on the inputs and subsequently the summation of the weighted inputs. After that, a bias value is added to the weighted sum and ultimately, the weighted sum is passed through a transfer function that compresses the fed term to a nominal range and outputs a value that usually ranges from zero to one.

In an optical integrated circuit, the inputs are usually in a multiplexed state. The input signals can be time division multiplexed (TDM) or wavelength multiplexed (WDM). Wavelength multiplexing is a more popular choice since it offers more advantages when compactness and speed are of essence. In WDM, the multiplexing is carried out with optical combiners that put them onto the same waveguide and are demultiplexed with optical filters. For the weighting of the inputs, one can utilize the many degrees of freedom with electromagnetic waves, such as its amplitude, phase, frequency, polarization, etc. In a WDM based model, the weighting can be carried out with tunable ring resonators that can filter and determine the amplitude of the input signal at its output.

Once the weighted input is ready, the resultant signal is passed through a device with a non-linear transfer function to obtain the activation signal. Micro-ring resonators, Mach-Zehnder interferometers, CROW filters, etc. can be used for implementing this in the optical domain. Another alternative is to transfer the signal to the electronic domain to utilize the non-linear transfer function of numerous electronic devices such as operational amplifiers, transimpedance amplifiers, inverters and various filters.

Numerous mathematical operations can be carried out with a photonic integrated circuit (PIC). A neuromorphic PIC would generally require addition and product of different parameters such as the inputs and weights. To achieve that, the concepts of superposition of light signals and intensity summation need to be used extensively. Mixing of signals can be carried out in non-linear devices to obtain higher order harmonics of the said signals. With these basic operations in place, the challenge of implementing more complex operations lie on the layout and design of the higher level device. Despite that, the footprint can still be kept reasonably small, such that a single chip remains within the dimensions of a square centimeter.

In the coming decade, photonic devices are therefore bound to see peaking interest and adoption in numerous industries including quantum computing, traditional computing, communications, defence and the likes of such industries. The bottleneck however remains in the development and adoption of newer production/fabrication technologies that allows an expansion of the horizons of the devices and applications of PICs.