Integrated electro-optic attention nonlinearities for transformers

Transformers have emerged as the dominant neural-network architecture, achieving state-of-the-art performance in language processing and computer vision. At the core of these models lies the attention mechanism, which requires a nonlinear, non-negative mapping using the Softmax function. However, although Softmax operations account for less than 1% of the total operation count, they can disproportionately bottleneck overall inference latency. Here, we use thin-film lithium niobate (TFLN) Mach-Zehnder modulators (MZMs) as analog nonlinear computational elements to drastically reduce the latency of nonlinear computations. We implement electro-optic alternatives to digital Softmax and Sigmoid, and evaluate their performance in Vision Transformers and Large Language Models. Our system maintains highly competitive accuracy, even under aggressive 4-bit input-output quantization of the analog units. We further characterize system noise at encoding speeds up to 10 GBaud and assess model robustness under various noise conditions. Our findings suggest that TFLN modulators can serve as nonlinear function units within hybrid co-packaged hardware, enabling high-speed and energy-efficient nonlinear computation.

Nonlinear Optical Data Transformer for Machine Learning

Modern machine learning models use an ever-increasing number of parameters to train (175 billion parameters for GPT-3) with large datasets to obtain better performance. Bigger is better has been the norm. Optical computing has been reawakened as a potential solution to large-scale computing through optical accelerators that carry out linear operations while reducing electrical power. However, to achieve efficient computing with light, creating and controlling nonlinearity optically rather than electronically remains a challenge. This study explores a reservoir computing (RC) approach whereby a 14 mm long few-mode waveguide in LiNbO3 on insulator is used as a complex nonlinear optical processor. A dataset is encoded digitally on the spectrum of a femtosecond pulse which is then launched in the waveguide. The output spectrum depends nonlinearly on the input. We experimentally show that a simple digital linear classifier with 784 parameters using the output spectrum from the waveguide as input increased the classification accuracy of several databases compared to non-transformed data, approximately 10$\%$. In comparison, a deep digital neural network (NN) with 40000 parameters was necessary to achieve the same accuracy. Reducing the number of parameters by a factor of $\sim$50 illustrates that a compact optical RC approach can perform on par with a deep digital NN.