The rapid advancement of machine learning is transforming the sciences and technology. Taking advantage of optics and photonics, optical artificial intelligence (AI) can realize large-bandwidth and high-energy-efficiency computing. Nonetheless, both optical AI and AI for optics are implemented in silico on electronic computers and thus require strict modelling and large amounts of training data to extract the system information. To address the challenges associated with offline modelling, it would be ideal to implement the machine learning algorithms on site. However, achieving high accuracy and high efficiency in parallel onsite learning of large-scale systems remains challenging.
The research team led by Professor Lu Fang from the Department of Electronic Engineering and the team led by Academician Qionghai Dai from the Department of Automation at Tsinghua University have presented fully forward mode (FFM) learning for onsite machine learning in free-space and integrated optical systems. This method maps optical systems to parameterized onsite neural networks, and enables self-learning guided by targets of applications. By leveraging spatial symmetry and Lorentz reciprocity, the necessity of backward propagation in the gradient descent training is eliminated. Consequently, the optical parameters can be self-designed directly on the original physical system.
This research paper, titled "Fully forward mode training for optical neural networks," was published online in the journal Nature on the evening of August 7, Beijing time.
Nature reviewers noted that "the ideas presented in the paper are novel, and such an ONN training process has not been previously demonstrated. The proposed method is effective and relatively easy to implement. Accordingly, it could become a widely adopted tool for training ONNs and other optical computing systems."
The versatility of the FFM method are demonstrated in advancing distinct fields at both the free-space and integrated photonic scales to realize deep optical neural networks (ONNs), high-resolution scattering imaging, dynamic all-optical non-line-of-sight systems, and model-free exceptional point searching in non-Hermitian systems. In deep optical neural networks, FFM achieves accuracy that closely approaches ideal in silico results, improving the accuracy of representative intelligent tasks by 40%. In complex scene imaging, FFM aligns well with the resolution limits of the imaging configurations, delivering an impressive energy efficiency of 5.40×10^6 TOPS/W. Furthermore, in the field of topological photonics, FFM can automatically search for non-Hermitian exceptional points without relying on any prior models.
Fully forward mode onsite machine learning (Left panel); FFM applications in advancing distinct fields (Right panel)
Damien Querlioz, a research director at the Center for Nanosciences and Nanotechnologies of Université Paris-Saclay and CNRS, commented in Nature that "systems that emulate biological neural networks offer an efficient way of running AI algorithms, but they can’t be trained using the conventional approach. The symmetry of these ‘physical’ networks provides a neat solution."
In summary, FFM providing a new route to model-free high-performance self-design of optical systems and self-learning physics, and suggesting the possibilities of large-scale, high-efficiency physical AI in the post-Moore era.
Editor: Li Han
