In this work, we realize the weighted addition of neuron on a monolithically integrated 8x8 cross-connect chip, based on semiconductor optical amplifier (SOA) and arrayed waveguide grating (AWG) technology, with four10Gb/s on-off key WDM input channels. We analyze the weight tuning range of each channel. The results show that a weight reading error below 0.08 can be achieved with a maximum weight tuning range of 20dB. It opens the way towards a high fidelity cross-connectivity of up to 8 neurons for input layer to next hidden-layer with a data precision of at least 3bits for photonic integrated deep neural networks.
|Publication status||Published - 2018|
|Event||23rd Annual Symposium of the IEEE Photonics Society Benelux Chapter - Paleis der Academiën, Brussels, Belgium|
Duration: 15 Nov 2018 → 16 Nov 2018
|Conference||23rd Annual Symposium of the IEEE Photonics Society Benelux Chapter|
|Period||15/11/18 → 16/11/18|
Shi, B., Calabretta, N., & Stabile, R. (2018). Continuous weight tuning for WDM-based neuron addition in an SOA-based InP cross-connect. Paper presented at 23rd Annual Symposium of the IEEE Photonics Society Benelux Chapter, Brussels, Belgium.