Photonic Bayesian Neural Network Using Programmed Optical Noises
Photonic Bayesian Neural Network Using Programmed Optical Noises
复制标题
DOI：
10.1109/jstqe.2022.3217819
复制
发表时间：
2023-03
影响因子：
4.9
通讯作者：
Changming Wu;Xiaoxuan Yang;Yiran Chen;Mo Li
中科院分区：
文献类型：
--
作者：
Changming Wu;Xiaoxuan Yang;Yiran Chen;Mo Li
贝叶斯神经网络（BNN）结合了神经网络和统计建模的优势，因为它同时执行后验预测并量化预测的不确定性。集成光子学已经成为一种有前途的神经网络加速器硬件平台，能够实现节能，低延迟和并行计算。然而，迄今为止，光子神经网络大多是确定性的网络模型。在这里，我们扩展了光子神经网络的统计模型，并提出了一个光子贝叶斯神经网络（P-BNN）架构的基础上集成的光子平台和利用固有的光学噪声。贝叶斯神经元是通过控制信号放大自发辐射（信号ASE）拍频噪声的概率分布来实现的。我们显示了P-BNN的优势，在使用后验分布进行预测，通过模拟P-BNN进行手写数字分类任务。仿真结果表明，P-BNN不仅能够成功预测测试数据集上的期望图像，而且能够检测并剔除训练数据集外的非期望图像。P-BNN结构与片上光放大器兼容，并且可以使用当前和新兴的集成光子技术进行放大，因此对于实际的神经网络应用是有希望的。
The Bayesian neural network (BNN) combines the strengths of neural networks and statistical modeling in that it simultaneously performs posterior predictions and quantifies the uncertainty of the predictions. Integrated photonics has emerged as a promising hardware platform of neural network accelerators capable of energy-efficient, low latency, and parallel computing. However, photonic neural networks demonstrated to date are mostly deterministic network models. Here, we extend the photonic neural network to a statistical model and proposed a photonic Bayesian neural network (P-BNN) architecture based on the integrated photonic platform and harnessing the inherent optical noises. The Bayesian neuron is realized by controlling the probability distribution of the signal-amplified spontaneous emission (signal-ASE) beat noise. We show the P-BNN's advantages in making predictions using the posterior distribution by simulating a p-BNN to perform handwritten number classification tasks. The simulation results show that the proposed P-BNN not only makes successful predictions on the expected images from the test dataset but also detects and rejects the unexpected images outside the training datasets. The P-BNN architecture is compatible with on-chip optical amplifiers and can be scaled up using current and emerging integrated photonics technologies, thus is promising for practical neural network applications.