Chongfan Technology
All Categories >

08

2024

-

08

Tsinghua University's latest Nature!

Author:

Optical computing promises to improve the speed and energy efficiency of machine learning applications. However, current methods for effectively training these models are limited by computer simulation on a digital computer.

On August 7, 2024, Fang Lu and Dai Qionghai of Tsinghua University jointly communicated inNature Published online entitled“Fully forward mode training for optical neural networks”The research paper,The studyA method called full forward mode (FFM) learning was developed, which implements a computationally intensive training process on a physical system. As a result, most machine learning operations can be performed efficiently in parallel in the field, alleviating the limitations of numerical modeling.

In free-space and integrated photonics, we experimentally demonstrate state-of-the-art performance of optical systems for a given network size. FFM learning shows that training the deepest optical neural network with millions of parameters can achieve an accuracy comparable to an ideal model. It supports all-optical focusing through a scattering medium with a diffraction-limited resolution; it can also perform parallel imaging of objects hidden out of sight at a frame rate of more than kilohertz, and can perform light intensity down to sub-photon per pixel at room temperature (5.40 × 1018-Energy efficiency per watt per second) of all-optical processing. It is further demonstrated that FFM learning can automatically search for non-Hermitian outliers without an analytic model.FFM learning not only facilitates orders of magnitude faster learning processes, but can also advance applied and theoretical fields such as deep neural networks, hyper-sensitive knowledge, and topological photonics.

The rapid development of machine learning is changing science and technology. Optical artificial intelligence (AI) takes advantage of optics and photonics to enable large bandwidth and energy efficient computing. At the same time, the application of artificial intelligence in the field of optics can improve the design and control of optical systems. Nevertheless, both optical artificial intelligence and optical artificial intelligence are implemented on electronic computers, so rigorous modeling and large amounts of training data are required to extract system information. Optical artificial intelligence methods are mainly concerned with the optical simulation of electronic artificial neural networks (ANNs) and the execution of photonic structure design on electronic computers. Therefore, correcting experimental systematic errors has always been a challenge, requiring a lot of effort to characterize the space and time of light propagation.

In artificial intelligence for optical design, the system must also be analytically or implicitly modeled. As system complexity increases, both analysis and numerical modeling become time-consuming, and the computational resources required increase exponentially.More importantly, due to the imperfection of the system and the complexity of light wave propagation, the perfect and accurate modeling of the general optical system is impossible, and there is always a mismatch between the offline model and the actual system.

To address the challenges associated with offline modeling, it would be desirable to implement machine learning algorithms on the site. However, it is still a challenge to achieve high accuracy and high efficiency in parallel field learning of large-scale systems. Perturbation-based optimization methods have limited efficiency. More effective gradient-based methods are usually limited to small-scale and simple problems, because the requirements for backward-propagating optical fields require strict system calibration and alignment. A hybrid approach combining in-situ forward propagation with off-line backpropagation using an electronic computer is also explored to implement the gradient descent algorithm. However, in general optical systems, the off-line modeling that is still required in the backpropagation process may not always be achievable.
Here, the authors propose full forward mode (FFM) learning for in-situ machine learning in free-space and integrated optical systems. The method maps the optical system to a parameterized field neural network and can be self-learning according to the goals of the application. Using spatial symmetry and Lorentz reciprocity, the necessity of backpropagation in gradient descent training is eliminated. Therefore, the optical parameters can be designed directly on the original physical system. The method calculates the gradient based on the measured light field data and error propagation, and then uses the gradient descent algorithm to effectively update the parameters. This study demonstrates the versatility of the FFM learning method to advance different fields on the free-space and integrated photon scales to achieve deep optical neural networks (ONNs), high-resolution scattering imaging, dynamic all-optical non-line-of-sight systems and non-Hermitian systems. Model-free anomaly search.

Chongqing Chongfan Technology Co., Ltd. Optical platform film thickness meter OCT optical coherence tomography LabVIEW control

Previous

Hepatology, Army Military Medical University Chai Jin and other teams found that SLC35C1 can reduce cholestatic liver injury and inflammation.

Nat Immunol | Fudan University Zhou Zhaacai/Jiao Shi shows that p97-Npl4 complex plays an important role in controlling the balance of Treg-TH17 cells in tumors

Next

Previous

Hepatology, Army Military Medical University Chai Jin and other teams found that SLC35C1 can reduce cholestatic liver injury and inflammation.

Next

Nat Immunol | Fudan University Zhou Zhaacai/Jiao Shi shows that p97-Npl4 complex plays an important role in controlling the balance of Treg-TH17 cells in tumors

LATEST NEWS