Implementation of energy-efficient convolutional neural networks based on kernel-pruned silicon photonics.

Area of Science:

• Photonics
• Artificial Intelligence
• Computer Engineering

Background:

• Silicon-based optical neural networks promise high-performance computing on integrated photonic circuits.
• Scalability of on-chip optical depth networks is limited by energy and space constraints.

Purpose of the Study:

• To present a silicon-based photonic convolutional neural network (PCNN) integrated with kernel pruning.
• To address the scalability limitations of current optical neural networks.

Main Methods:

• Developed a PCNN with a tunable micro-ring resonator weight bank as the optical convolutional computing core.
• Investigated the impact of weight mapping accuracy on PCNN performance through numerical simulations.
• Experimentally demonstrated PCNN accuracy on the MNIST dataset with significant kernel pruning.

Main Results:

• PCNN performance degrades significantly with weight mapping accuracy below 4.3 bits.
• Experimental results show minimal accuracy loss on MNIST even after pruning 93.75% of convolutional kernels.
• Kernel pruning saves approximately 202.3 mW per removed kernel, with energy savings linearly proportional to the number of pruned kernels.

Conclusions:

• Kernel pruning is a viable strategy for enhancing energy efficiency in silicon-based PCNNs.
• The proposed methodology is scalable, offering a path towards faster and more energy-efficient large-scale optical convolutional neural networks on photonic integrated circuits.