arXiv subjects

Jianjun Cao

Publications and source records attributed to Jianjun Cao.

8 recordsLinked to original sources

arXiv2026

Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes

Deep neural networks exhibit periodic loss spikes during unregularized long-term training, a phenomenon known as the "Slingshot Mechanism." Existing work usually attributes this to intrinsic optimization dynamics, but its triggering mechanism remains unclear. This paper proves that this phenomenon is a result of floating-point arithmetic precision limits. As training enters a high-confidence stage, the difference between the correct-class logit and the other logits may exceed the absorption-error threshold. Then during backpropagation, the gradient of the correct class is rounded exactly to zero, while the gradients of the incorrect classes remain nonzero. This breaks the zero-sum constraint of gradients across classes and introduces a systematic drift in the parameter update of the classifier layer. We prove that this drift forms a positive feedback loop with the feature, causing the global classifier mean and the global feature mean to grow exponentially. We call this mechanism Numerical Feature Inflation (NFI). This mechanism explains the rapid norm growth before a Slingshot spike, the subsequent reappearance of gradients, and the resulting loss spike. We further show that NFI is not equivalent to an observed loss spike: in more practical tasks, partial absorption may not produce visible spikes, but it can still break the zero-sum constraint and drive rapid growth of parameter norms. Our results reinterpret Slingshot as a numerical dynamic of finite-precision training, and provide a testable explanation for abnormal parameter growth and logit divergence in late-stage training.

cs.LG↗

arXiv2025

Antarctic TianMu Staring Observation Project I: Overview and Implementation of the Prototype Telescope

Wide-field rapid sky surveys serve as critical observational methods for time-domain astronomical research. The Antarctic region, with several months of continuous dark nights annually, is an ideal site for time-domain astronomical observations. The Antarctic TianMu Staring Observation Project aims to deploy a fleet of small telescopes, adopting an array observation model to conduct time-domain optical observations in Antarctica, featuring wide-sky coverage, high-cadence sampling, long-period staring, and simultaneous multi-band measurements. Considering the severe challenges optical telescopes face in Antarctica, including extremely low temperatures, unattended operation, and limited power supply and network transmission, we have designed and developed the Antarctic TianMu prototype telescope based on drift-scan charge-coupled device technology. In October 2022, our prototype (with an aperture of 18 cm), named AT-Proto was transported to Zhongshan Station in Antarctica aboard China's 39th Antarctic Research Expedition. It has since operated stably and reliably in the frigid environment for over two years, demonstrating the significant advantages of this technology in polar astronomical observations. The experimental observation results of AT-Proto provide a solid foundation for the subsequent construction of a time-domain astronomy observation array in Antarctica.

astro-ph.IM↗

arXiv2021

Far-Field Super-Resolution Imaging By Nonlinear Excited Evanescent Waves

Abbe's resolution limit, one of the best-known physical limitations, poses a great challenge for any wave systems in imaging, wave transport, and dynamics. Originally formulated in linear optics, this Abbe's limit can be broken using nonlinear optical interactions. Here we extend the Abbe theory into a nonlinear regime and experimentally demonstrate a far-field, label-free, and scan-free super-resolution imaging technique based on nonlinear four-wave mixing to retrieve near-field scattered evanescent waves, achieving sub-wavelength resolution of $λ/15.6$. This method paves the way for application in biomedical imaging, semiconductor metrology, and photolithography.

physics.optics↗

arXiv2018

Multi-Modal Coreference Resolution with the Correlation between Space Structures

Multi-modal data is becoming more common in big data background. Finding the semantically similar objects from different modality is one of the heart problems of multi-modal learning. Most of the current methods try to learn the inter-modal correlation with extrinsic supervised information, while intrinsic structural information of each modality is neglected. The performance of these methods heavily depends on the richness of training samples. However, obtaining the multi-modal training samples is still a labor and cost intensive work. In this paper, we bring a extrinsic correlation between the space structures of each modalities in coreference resolution. With this correlation, a semi-supervised learning model for multi-modal coreference resolution is proposed. We firstly extract high-level features of images and text, then compute the distances of each object from some reference points to build the space structure of each modality. With a shared reference point set, the space structures of each modality are correlated. We employ the correlation to build a commonly shared space that the semantic distance between multi-modal objects can be computed directly. The experiments on two multi-modal datasets show that our model performs better than the existing methods with insufficient training data.

cs.AI↗

arXiv2016

Optically Induced Transparency in a Micro-cavity

Electromagnetically induced transparency has the unique ability to optically control transparency windows with low light in atomic systems. However, its practical applications in quantum physics and information science are limited due to rigid experimental requirements. Here, we demonstrate a new mechanism of optically induced transparency in a micro-cavity by introducing a four-wave mixing gain to nonlinearly couple two separated resonances of the micro-cavity in an ambient environment. A signature Fano-like resonance was observed owing to the nonlinear interference of the two coupled resonances. Moreover, we show that the unidirectional gain of the four-wave mixing can lead to the remarkable effect of non-reciprocal transmission at the transparency windows. Optically induced transparency may offer a unique platform for a compact, integrated solution to all-optical and quantum information.

physics.optics↗

arXiv2015

Nonlinear Negative Refraction by Difference Frequency Generation

Negative refraction has attracted much interest for its promising capability in imaging applications. Such an effect can be implemented by negative index meta-materials, however, which are usually accompanied by high loss and demanding fabrication processes. Recently, alternative nonlinear approaches like phase conjugation and four wave mixing have shown advantages of low-loss and easy-to-implement, but associated problems like narrow accepting angles can still halt their practical applications. Here we demonstrate theoretically and experimentally a new scheme to realize negative refraction by nonlinear difference frequency generation with wide tunability, where a thin BBO slice serves as a negative refraction layer bending the input signal beam to the idler beam at a negative angle. Furthermore, we realize optical focusing effect using such nonlinear negative refraction, which may enable many potential applications in imaging science.

physics.optics↗

arXiv2014

Dielectric Optical-Controlled Magnifying Lens by Nonlinear Negative Refraction

A simple optical lens plays an important role for exploring the microscopic world in science and technology by refracting light with tailored spatially varying refractive index. Recent advancements in nanotechnology enable novel lenses, such as, superlens, hyperlens, Luneburg lens, with sub-wavelength resolution capabilities by specially designing materials' refractive indices with meta-materials and transformation optics. However, these artificially nano/micro engineered lenses usually suffer high losses from metals and are highly demanding in fabrication. Here we experimentally demonstrate for the first time a nonlinear dielectric magnifying lens using negative refraction by degenerate four-wave mixing in a plano-concave glass slide, obtaining magnified images. Moreover, we transform a nonlinear flat lens into a magnifying lens by introducing transformation optics into nonlinear regime, achieving an all-optical controllable lensing effect through nonlinear wave mixing, which may have many potential applications in microscopy and imaging science.

physics.optics↗

arXiv2014

Metal-Free Flat Lens Using Negative Refraction by Nonlinear Four-wave Mixing

A perfect lens with unlimited resolution has always posed a challenge to both theoretical and experimental physicists. Recent developments in optical meta-materials promise an attractive approach towards perfect lenses using negative refraction to overcome the diffraction limit, improving resolution. However, those artificially engineered meta-materials usually company by high losses from metals and are extremely difficult to fabricate. An alternative proposal using negative refraction by four-wave mixing has attracted much interests recently, though most of existing experiments still require metals and none of them has been implemented for an optical lens. Here we experimentally demonstrate a metal-free flat lens for the first time using negative refraction by degenerate four-wave mixing with a thin glass slide. We realize optical lensing effect utilizing a nonlinear refraction law, which may have potential applications in microscopy.

physics.optics↗