Skip navigation
Please use this identifier to cite or link to this item: http://arks.princeton.edu/ark:/88435/dsp01w3763965v
Title: Dynamics and Operations of Photonic Neurons
Authors: Nahmias, Mitchell Aaron
Advisors: Prucnal, Paul R
Contributors: Electrical Engineering Department
Keywords: Artificial Intelligence
Lasers
Neural Networks
Neuromorphic Photonics
Photonics
Unconventional Computing
Subjects: Optics
Computational physics
Issue Date: 2019
Publisher: Princeton, NJ : Princeton University
Abstract: Photonic communication channels—which code information on lightwaves, rather than electrons—compose both the vast networks that underlie the internet and the fiber links that connect servers at datacenters. Electronics, in contrast, has dominated information processing: digital logic gates, especially those based on complementary metal oxide semiconductor (CMOS) technology, has driven the computing landscape for nearly sixty years with a Moore’s law progression towards faster, higher through- put processors. Electronic computing, however, is running up against fundamental physical limitations that are increasingly harder to circumvent. The first major limitation is data movement. More than sixty percent of both the energy and area costs in electronic hardware result from interconnects: the energy lost from capacitively charging and discharging the wires that move information from one point to another. This cost is exacerbated in highly parallel processing models such as field programmable gate arrays (FPGAs), for example, which sometimes have greater than ninety percent of the space or energy costs in the interconnects alone. The second is in data processing. High performance computing (HPC), which involves large-scale simulations of physical phenomena (i.e., weather), solutions to complex systems problems (i.e., social networks), and artificial intelligence (i.e., deep learning), is rapidly becoming a cornerstone in data industries and many fields of sci- ence. It is primarily bottlenecked by linear operations such as matrix multiplications and fourier transforms. The demand for deep learning training, for example, appears to be doubling every 3.5 months, far outpacing Moore’s law typical progression of transistor density doubling performance every 18 months [13]. This enormous gap between supply and demand presents a significant opportunity for unconventional approaches. Decades ago, photonic computing was unable to match the performance scaling of digital electronics, but today, the landscape has changed tremendously. Moore’s law scaling is slowing down at a time when computing demand is expanding more than ever. Scaling technologies now exist for photonics—for example, it is now possible to integrate high efficiency photonic components directly into modern microelectronic chips with only several modifications to the fabrication processes [143]. The energy efficiency (per bit) of optical links are beginning to match or exceed those of electronic chip-to-chip interconnects (<1 pJ/bit). Further developments in both photonic scala- bility and miniaturization is expected to lead to better performance as the technology matures. Photonics has the potential to directly address many well-known bottlenecks in electronic computing. For example, an optical communication link only requires charging or discharging optoelectronic transducers, and optical multiplexing allows for enormous on-chip bandwidth capabilities between interacting processors with an energy cost that is nearly constant with respect to the length of the data link. A second useful property is the ability of optical systems to perform linear operations efficiently: a matrix multiplication with N channels will in total perform N2 matrix operations, but the energy required to do this in a passive system scales only with the number of channels N.1 Motivated by these many advantages, our group has spearheaded a field now known as neuromorphic photonics, in which neural network models are directly in- stantiated with photonic components. This thesis focuses on the neurons themselves, designed to combine the best properties of electronic and photonic signals while re- maining compatible with photonic integrated circuit (PIC) platforms to assure future scalability and compatibility. We discuss several neuromorphic photonic units: the first model is a fully functional laser neuron in a photonic integrated circuit platform: it involves an integrated laser driven by a photodetector through a short, recieverless electronic link, exhibiting biologically-relevant spiking behavior at a sub-nanosecond timing resolution. We also discuss new models based on modulator-class systems, together with the use of novel materials (graphene) or nonlinear effects (the quantum confined stark effect). We end with a detailed comparison of neural network-inspired photonic integrated circuits with current systems in digital and analog electronics, showing significant advantages in the photonic domain for matrix-like operations in artificial intelligence.
URI: http://arks.princeton.edu/ark:/88435/dsp01w3763965v
Alternate format: The Mudd Manuscript Library retains one bound copy of each dissertation. Search for these copies in the library's main catalog: catalog.princeton.edu
Type of Material: Academic dissertations (Ph.D.)
Language: en
Appears in Collections:Electrical Engineering

Files in This Item:
File Description SizeFormat 
Nahmias_princeton_0181D_13193.pdf68.44 MBAdobe PDFView/Download


Items in Dataspace are protected by copyright, with all rights reserved, unless otherwise indicated.