This presentation described the construction of neural networks using light; it was an introduction to three papers that have been submitted on optical neural networks since 2018. In recent years, many attempts have been made to reproduce neural networks with light to increase speed and reduce power. In constructing a network, optimization is essential to generate appropriate outputs from inputs. Matrix computation in optical neural networks is based on the principle of Mach-Zehnder interferometry, and is represented by the intensity of light intensity due to phase shift. Optimization requires measurement of the gradient. While previous studies have used computer models for optimization, a new approach, called the associated transformation, is used to obtain the gradient by measuring the light intensity of the device. This allows the gradient to be acquired in the optical domain, shortening the procedure used in conventional methods.
Article: https://www.osapublishing.org/optica/abstract.cfm?uri=optica-5-7-864
: https://www.osapublishing.org/oe/abstract.cfm?uri=oe-28-8-12138
: https://advances.sciencemag.org/content/5/12/eaay6946
CLEO 2020 Takumasa Kodama
Research
CLEO 2020 Participation Report
11-15 May 2020, Online
Takumasa Kodama, 1st year master's student
About CLEO2020
CLEO (The Optical Society OSA) is the largest international photonics conference sponsored by the American Physical Society (APS) and the IEEE Photonics Society. The conference is usually held in San Jose, but this year, due to the spread of the new coronavirus (COVID-19), all presentations were made online. Although the situation was unprecedented, presenters from a total of 75 countries gathered, and more than 2000 presentations and 248 posters were made. As the presenter also participated in the conference, I would like to introduce some of the presentations I attended, including my impressions through my own presentation.
2. Presentation by the presenter
Title: High-resolution Spectrometer with Random Photonic Crystals
Presenter: Takumasa Kodama
Affiliation: Keio University
Presentation No.: FM2R.4 (Mon, May 11th)
We presented a microscopic spectrometer using a width-variant line-defect photonic crystal waveguide and a deep learning algorithm. The merging of photonics and deep learning was also seen at the conference, and we could reconfirm that deep learning is a recent trend. The presentation was given in a live format, and the chairperson read the questions received in the Q&A section during the Q&A session. I was anxious about listening to English, but I was able to overcome my anxiety because I was able to read the questions myself. Many of the questions were related to concepts such as the difference between deep learning and optimization algorithms used for data analysis, and why such algorithms can improve resolution. When presenting, the presenter shares his presentation materials on the screen, but if he does it incorrectly, the presenter's tools can be seen by the audience. To avoid this, start the slide show after sharing the PowerPoint presentation.
3. presentations attended
Title: Lasing up to T = 339K in Subwavelength Nanowire-Induced Photonic Crystal Nanocavities
Presenter: Sylvain Sergent
Affiliation: NTT Nanophotnics Center
Presentation No.: SM1J.6 (Mon, May 11th)
Low-threshold lasing using silicon-nitride photonic crystals and group III-V nanowires The photonics group at NTT has been actively working on photonic crystals and plasmonics in recent years. In this presentation, nanowires are placed in a photonic crystal waveguide, which serves as a gain medium, and excitation up to T = 360 K is confirmed. Many combinations of nanowire length, radius, and lattice constant were tested, and their Q-values and threshold values were measured, showing their relationship. The photonic crystals are used to directly insert light into the nanowires, making them compatible with integrated circuits, room temperature, and low power input (3.5 MW/cm2). The silicon substrate is very small in size, so there is little loss of light. The silicon substrate is very small in size, and can be operated at room temperature, resulting in very low power consumption. The presenters also mentioned the advantage of using various nanowires to control the laser wavelength and flexibility in design, although ZnO was used in this presentation.
Paper: https://pubs.acs.org/doi/pdf/10.1021/acsphotonics.0c00166
Title: Photonic neural networks: training, nonlinearity, and recurrent systems
Presenter: Shanhui Fan
Affiliation: Stanford University
Presentation No.: JF2A.3 (Fri, May 15th)
English 
日本語 
简体中文 
Deutsch 
Français 
한국어 
Español 
Italiano