Computing enabled by electronics has been improved so extensively as to make machine learning algorithms such as deep neural network so powerful than ever before. Data communications enabled by optics have been one of the cornerstones of the modern society built upon Internet. The rapidly increasing demands for communication bandwidth due to numerous emerging applications such as AI-based cloud computing have significantly increased the bisection bandwidth of intra-datacenter networks. This trend will necessitate the optics-electronics co-packaging on board, which can only be realized by the substantial development of integrated photonics such as silicon photonics. This opportunity, making optics and electronics so close to each other, will in turn offer a chance to reconsider building EO-hybrid computational and intelligent systems. Although optical computing and optical neural networks have been proposed since a while ago, recent demonstrations of deep neural networks implemented on silicon photonics reactivate the study of exploiting photonics for machine learning. The current program-based deep neural networks are powerful, but consume huge computational resources. Suppose that just optical propagation in compact photonic chips could realize similar functions, it would greatly decrease the power and latency. Even though the scalability and the integration of nonlinear activation functions on photonic chips are still challenging, for some applications such as classifier and digit recognition, photonics systems could be viable and beneficial from such aspects.
This talk will introduce our efforts to develop the topology concepts, algorithms, and applications in order to implement silicon photonics for calculation and machine learning applications. First, a high-bit reconfigurable DAC based on generic photonic circuits will be presented. Second, without using any nonlinear activation functions or building deep neural networks, we demonstrate a photonic classifier based on only linear optical components. At last, we will show several ways of implementing silicon photonic circuit to recognize digits.
Digital-to-analog converters (DAC) are indispensable functional units in optical signal transmission and processing. The photonic DAC that converts electrical digital signals to an optical analog one will offer advantages in lowering system complexity, power, and cost. Especially with the required bandwidth increasing, it could mitigate the problems faced by its electrical counterparts in dealing with higher sampling rate. Achieving such a photonic DAC in silicon photonics is promising due to the integration capability of both electronics and photonics and large scale DAC-based photonic circuits can be further realized for on-chip optical signal processing. In this work, we demonstrate 2-bit D/A conversion for the simple proof of concept utilizing only one single silicon Mach-Zehnder modulator (MZM), which is much simpler than previously reported segmented MZM and microring resonator based DACs. One-single MZM capable of 2-bit DAC merits future higher bit resolution design and meanwhile guarantees wide spectral bandwidth. One arm of MZM is used for the MSB bit input, while the other for the LSB, both of them being accomplished by only one phase shifter. For each bit input, we utilize amplitude modulation, instead of phase modulation, by applying the carrier injection induced absorption in the phase shifters. For principle, by setting different bias points for two phase shifters, we can produce the condition at which the amplitude weighting ratio of LSB to MSB is 1/2 in order to obtain the linear amplitude DAC output. In other words, the output optical field has the analog linear amplitude levels (0,1,2,3) which corresponds to the power levels of (0,1,4,9) at the full extinction condition. For fabrication, this device was fabricated on a 220-nm SOI wafer with a 3-m buried-oxide layer at the AIST SCR 300-mm CMOS foundry. The 430-nm-wide fully etched channel waveguide was used for the components except for the pn phase shifter which adopted the shallow-etched rib waveguide structure with a slab thickness of about 110 nm and a width of about 600 nm. The doping density in the weak p/n regions was about 1.61018 cm-3. This MZM was arm-balanced with 2-mm-long phase shifters, adopting GSGSG configuration. Two 50- terminators were also integrated on-chip at the ends of two signal electrodes. For measurement, a two-channel pulse pattern generator produced bit sequences at various frequencies for both MSB and LSB which was applied to the signal electrodes through bias-tees and high-speed probes. The 1.55-m cw light at TE polarization was coupled into the chip via a tapered fiber and the optical output passed an EDFA and a bandpass filter and then was sent to a high-speed oscilloscope for examining DAC analog output. Using this device, we successfully achieved correct D/A conversions with the sampling rates up to 3 GS/s with <1 V peak-to-peak voltages. Note that this speed can be further enhanced to <10 GS/s by constructing the pn phase shifter into a MZM structure or replacing it with a SiGe electro-absorption modulator. In summary, this work verified the feasibility to realize high-sampling-rate 2-bit D/A conversion utilizing a single silicon MZM modulator.