Event
Ph.D. Dissertation Defense: Benjamin Klimko
Wednesday, August 7, 2024
10:00 a.m.
ERF 1207
Maria Hoo
301 405 3681
mch@umd.edu
ANNOUNCEMENT: Ph.D. Dissertation Defense
Name: Benjamin Klimko
Committee:
Professor Yanne Chembo (Chair)
Professor Thomas Murphy
Professor Thomas Antonsen
Professor Kevin Daniels
Professor Rajarshi Roy (Dean's representative)
Date/time: Wednesday, August 7, 2024 at 10:00 AM - 12:00 PM
Location: ERF 1207
Title: Development of Photonic Reservoir Computers for Radiofrequency Spectrum Awareness
Name: Benjamin Klimko
Committee:
Professor Yanne Chembo (Chair)
Professor Thomas Murphy
Professor Thomas Antonsen
Professor Kevin Daniels
Professor Rajarshi Roy (Dean's representative)
Date/time: Wednesday, August 7, 2024 at 10:00 AM - 12:00 PM
Abstract:
In this dissertation, we study the use of several optoelectronic oscillator architectures for
physical reservoir computing tasks. While optoelectronic oscillator-based reservoir computers have been reported in the literature for over a decade, all reported experimental results have been processed using wideband systems with baseband data sets. Our work focuses on two major innovations for physical reservoir computing: (i) narrowband reservoir computers allowing computing tasks to be performed natively on radiofrequency signals and (ii) illustrating that “simplified” optoelectronic oscillators, without external optical modulators, are capable of meeting or exceeding the results from more complex photonic reservoir computers.
By their nature, optoelectronic oscillators operate in the radio passband regime and reservoir computers have been shown to be capable on time-series tasks such as waveform prediction and classification data sets. We demonstrated that the optoelectronic oscillator-based reservoir computer can effectively process signals in the radio passband, which is a novel result that could provide an enabling technology for next-generation communication methods such as cognitive networks. The benefits of this innovation would scale with increasing frequency, such as potential use with millimeter-wave cellular networks.
In our second physical reservoir innovation, we have shown that external optical modulators, nearly ubiquitous devices in optoelectronic oscillators, may be excluded from the design of a physical reservoir computer without decreasing its accuracy. This is a major result as a reservoir without active optical components could be built on a single integrated circuit using modern semiconductor manufacturing processes. Such integration and miniaturization would be a large step towards photonic reservoir systems that could be more easily put into an operational environment. Up to this point, there has been minimal work on transitioning the optoelectronic oscillator from a benchtop, experimental system to one useful in the real world.
Lastly, we investigated the relationship between computational power of the reservoir computer and task error. This is a crucial finding since reservoir computing is often touted as an alternative computing paradigm that is less resource-intensive than other computing methods. By determining a threshold on computational needs for a photonic reservoir computer, we ensure that such systems are utilized efficiently and do not unnecessarily use resources.
physical reservoir computing tasks. While optoelectronic oscillator-based reservoir computers have been reported in the literature for over a decade, all reported experimental results have been processed using wideband systems with baseband data sets. Our work focuses on two major innovations for physical reservoir computing: (i) narrowband reservoir computers allowing computing tasks to be performed natively on radiofrequency signals and (ii) illustrating that “simplified” optoelectronic oscillators, without external optical modulators, are capable of meeting or exceeding the results from more complex photonic reservoir computers.
By their nature, optoelectronic oscillators operate in the radio passband regime and reservoir computers have been shown to be capable on time-series tasks such as waveform prediction and classification data sets. We demonstrated that the optoelectronic oscillator-based reservoir computer can effectively process signals in the radio passband, which is a novel result that could provide an enabling technology for next-generation communication methods such as cognitive networks. The benefits of this innovation would scale with increasing frequency, such as potential use with millimeter-wave cellular networks.
In our second physical reservoir innovation, we have shown that external optical modulators, nearly ubiquitous devices in optoelectronic oscillators, may be excluded from the design of a physical reservoir computer without decreasing its accuracy. This is a major result as a reservoir without active optical components could be built on a single integrated circuit using modern semiconductor manufacturing processes. Such integration and miniaturization would be a large step towards photonic reservoir systems that could be more easily put into an operational environment. Up to this point, there has been minimal work on transitioning the optoelectronic oscillator from a benchtop, experimental system to one useful in the real world.
Lastly, we investigated the relationship between computational power of the reservoir computer and task error. This is a crucial finding since reservoir computing is often touted as an alternative computing paradigm that is less resource-intensive than other computing methods. By determining a threshold on computational needs for a photonic reservoir computer, we ensure that such systems are utilized efficiently and do not unnecessarily use resources.
