arXiv subjects

Shuaifeng Li

Publications and source records attributed to Shuaifeng Li.

16 recordsLinked to original sources

arXiv2026

In situ local learning of dynamic network materials

Functional materials realize functionalities through pre-designed structures, whereas physical neural networks are trained by prescribing input-output behavior. Unifying these views would allow materials to acquire functions directly from dynamical tasks. Here we introduce an in situ local learning framework for dynamic network materials. A forward drive and a time-reversed adjoint error drive are applied to the same mechanical network, enabling the material to compute gradients of a time-domain loss through its own dynamics. The resulting update rules are local, compatible with the principle of physical learning. Each bond or node requires only co-located forward and adjoint fields. The method trains spring constants, nodal masses, bond damping coefficients and nodal damping coefficients, and programs dynamical functions including broadband wave cloaking, evanescent-wave imaging, passive transient output enhancement and vowel classification. These examples involve distinct physical mechanisms, such as wave scattering, near-field information transfer, dissipative non-normal dynamics and machine learning, but are obtained by simply changing the loss function, rather than leveraging expert knowledge. Therefore, dynamic network materials can act both as programmable matter and as physical neural networks, providing a route to material systems that learn temporal responses in situ.

cond-mat.soft↗

arXiv2026

Reward-Guided Semantic Evolution for Test-time Adaptive Object Detection

Open-vocabulary object detection with vision-language models (VLMs) such as Grounding DINO suffers from performance degradation under test-time distribution shifts, primarily due to semantic misalignment between text embeddings and shifted visual embeddings of region proposals. While recent test-time adaptive object detection methods for VLM-based either rely on costly backpropagation or bypass semantic misalignment via external memory, none directly and efficiently align text and vision in a training-free manner. To address this, we propose Reward-Guided Semantic Evolution (RGSE), a training-free framework that directly refines the text embeddings at test time. Inspired by evolutionary search, RGSE treats text embedding adaptation as a semantic search process: it perturbs text embeddings as candidate variants, evaluates them via cosine similarity with current and historical high-confidence visual proposals as a reward signal, and fuses them into a refined embedding through reward-weighted averaging. Without any backpropagation, RGSE achieves state-of-the-art performance across multiple detection benchmarks while adding minimal computational overhead. Our code will be open source upon publication.

cs.CV↗

arXiv2025

Bayesian Test-time Adaptation for Object Recognition and Detection with Vision-language Models

Vision-language models (VLMs) such as CLIP and Grounding DINO have achieved remarkable success in object recognition and detection. However, their performance often degrades under real-world distribution shifts. Test-time adaptation (TTA) aims to mitigate this issue by adapting models during inference. Existing methods either rely on computationally expensive backpropagation, which hinders real-time deployment, or focus solely on likelihood adaptation, which overlooks the critical role of the prior. Our prior work, Bayesian Class Adaptation (BCA), addressed these shortcomings for object recognition by introducing a training-free framework that incorporates adaptive priors. Building upon this foundation, we now present Bayesian Class Adaptation plus (BCA+), a unified, training-free framework for TTA for both object recognition and detection. BCA+ introduces a dynamic cache that adaptively stores and updates class embeddings, spatial scales (for detection), and, crucially, adaptive class priors derived from historical predictions. We formulate adaptation as a Bayesian inference problem, where final predictions are generated by fusing the initial VLM output with a cache-based prediction. This cache-based prediction combines a dynamically updated likelihood (measuring feature and scale similarity) and a prior (reflecting the evolving class distribution). This dual-adaptation mechanism, coupled with uncertainty-guided fusion, enables BCA+ to correct both the model's semantic understanding and its contextual confidence. As a training-free method requiring no backpropagation, BCA+ is highly efficient. Extensive experiments demonstrate that BCA+ achieves state-of-the-art performance on both recognition and detection benchmarks.

cs.CV↗

arXiv2025

Topological edge states and amplitude-dependent delocalization in quasiperiodic elliptically geared lattices

We present a class of mechanical lattices based on elliptical gears with quasiperiodic modulation and geometric nonlinearity, capable of exhibiting topologically protected modes and amplitude-driven transitions. Starting from a one-dimensional chain of modulated elliptical gears, we demonstrate the emergence of localized edge states arising from quasiperiodic variation in the gears' moments of inertia, analogous to the topological edge modes of the Aubry-Andre-Harper model. Under increasing excitation amplitude, the system undergoes a nonlinear transition, where edge localization breaks down and energy delocalizes into the bulk. By coupling multiple such chains with varying modulation phase, we construct a two-dimensional lattice in which the phase acts as a synthetic dimension. This structure supports topological wave propagation along the synthetic dimension. Nonlinearity again induces a breakdown of topological states, leading to complex, amplitude-dependent wave propagation. We further propose a numerical continuation approach to analyzing the periodic orbits and their linear stability, effectively discovering the boundary of the basin of bounded motion and detecting the occurrence of delocalization under certain excitation amplitudes. Our results reveal that elliptical geared systems offer a passive, amplitude-dependent platform for exploring topological phenomena and synthetic dimensionality in mechanical metamaterials.

physics.app-ph↗

arXiv2025

Text-Driven Causal Representation Learning for Source-Free Domain Generalization

Deep learning often struggles when training and test data distributions differ. Traditional domain generalization (DG) tackles this by including data from multiple source domains, which is impractical due to expensive data collection and annotation. Recent vision-language models like CLIP enable source-free domain generalization (SFDG) by using text prompts to simulate visual representations, reducing data demands. However, existing SFDG methods struggle with domain-specific confounders, limiting their generalization capabilities. To address this issue, we propose TDCRL (\textbf{T}ext-\textbf{D}riven \textbf{C}ausal \textbf{R}epresentation \textbf{L}earning), the first method to integrate causal inference into the SFDG setting. TDCRL operates in two steps: first, it employs data augmentation to generate style word vectors, combining them with class information to generate text embeddings to simulate visual representations; second, it trains a causal intervention network with a confounder dictionary to extract domain-invariant features. Grounded in causal learning, our approach offers a clear and effective mechanism to achieve robust, domain-invariant features, ensuring robust generalization. Extensive experiments on PACS, VLCS, OfficeHome, and DomainNet show state-of-the-art performance, proving TDCRL effectiveness in SFDG.

cs.LG↗

arXiv2025

Bayesian Test-Time Adaptation for Vision-Language Models

Test-time adaptation with pre-trained vision-language models, such as CLIP, aims to adapt the model to new, potentially out-of-distribution test data. Existing methods calculate the similarity between visual embedding and learnable class embeddings, which are initialized by text embeddings, for zero-shot image classification. In this work, we first analyze this process based on Bayes theorem, and observe that the core factors influencing the final prediction are the likelihood and the prior. However, existing methods essentially focus on adapting class embeddings to adapt likelihood, but they often ignore the importance of prior. To address this gap, we propose a novel approach, \textbf{B}ayesian \textbf{C}lass \textbf{A}daptation (BCA), which in addition to continuously updating class embeddings to adapt likelihood, also uses the posterior of incoming samples to continuously update the prior for each class embedding. This dual updating mechanism allows the model to better adapt to distribution shifts and achieve higher prediction accuracy. Our method not only surpasses existing approaches in terms of performance metrics but also maintains superior inference rates and memory usage, making it highly efficient and practical for real-world applications.

cs.CV↗

arXiv2025

Topological mechanical neural networks as classifiers through in situ backpropagation learning

Recently, a new frontier in computing has emerged with physical neural networks(PNNs) harnessing intrinsic physical processes for learning. Here, we explore topological mechanical neural networks(TMNNs) inspired by the quantum spin Hall effect(QSHE) in topological metamaterials, for machine learning classification tasks. TMNNs utilize pseudospin states and the robustness of the QSHE, making them damage-tolerant for binary classification. We first demonstrate data clustering using untrained TMNNs. Then, for specific tasks, we derive an in situ backpropagation algorithm - a two-step, local-rule method that updates TMNNs using only local information, enabling in situ physical learning. TMNNs achieve high accuracy in classifications of Iris flowers, Penguins, and Seeds while maintaining robustness against bond pruning. Furthermore, we demonstrate parallel classification via frequency-division multiplexing, assigning different tasks to distinct frequencies for enhanced efficiency. Our work introduces in situ backpropagation for wave-based mechanical neural networks and positions TMNNs as promising neuromorphic computing hardware for classification tasks.

cond-mat.dis-nn↗

arXiv2024

On-demand realization of topological states using Miura-folded metamaterials

Recent advancements in topological metamaterials have unveiled fruitful physics and numerous applications. Whereas initial efforts focus on achieving topologically protected edge states through principles of structural symmetry, the burgeoning field now aspires to customize topological states, tailoring their emergence and frequency. Here, our study presents the realization of topological phase transitions utilizing compliant mechanisms on the facets of Miura-folded metamaterials. This approach induces two opposite topological phases, leading to topological states at the interface. Moreover, we exploit the unique folding behavior of Miura-folded metamaterials to tune the frequency of topological states and dynamically toggle their presence. Our research not only paves the way for inducing topological phase transitions in Miura-folded structures but also enables the on-demand control of topological states, with promising applications in wave manipulation and vibration isolation.

physics.app-ph↗

arXiv2024

Training all-mechanical neural networks for task learning through in situ backpropagation

Recent advances unveiled physical neural networks as promising machine learning platforms, offering faster and more energy-efficient information processing. Compared with extensively-studied optical neural networks, the development of mechanical neural networks (MNNs) remains nascent and faces significant challenges, including heavy computational demands and learning with approximate gradients. Here, we introduce the mechanical analogue of in situ backpropagation to enable highly efficient training of MNNs. We demonstrate that the exact gradient can be obtained locally in MNNs, enabling learning through their immediate vicinity. With the gradient information, we showcase the successful training of MNNs for behavior learning and machine learning tasks, achieving high accuracy in regression and classification. Furthermore, we present the retrainability of MNNs involving task-switching and damage, demonstrating the resilience. Our findings, which integrate the theory for training MNNs and experimental and numerical validations, pave the way for mechanical machine learning hardware and autonomous self-learning material systems.

cs.LG↗

arXiv2024

Topological pumping in origami metamaterials

In this study, we present a mechanism of topological pumping in origami metamaterials with spatial modulation by tuning the rotation angles. Through coupling spatially modulated origami chains along an additional synthetic dimension, the pumping of waves from one topological edge state to another is achieved, where the Landau-Zener transition is demonstrated by varying the number of coupled origami chains. Besides, the inherent nonlinearity of origami metamaterials enable the excitation-dependent Landau-Zener tunneling probability. Furthermore, with the increase of nonlinearity, the topological states tend to localize in several regions in a way reminiscent of discrete breathers. Our findings pave the way towards inter-band transitions and associated topological pumping features in origami metamaterials.

physics.app-ph↗

arXiv2023

Elastic chiral Landau level and snake states in origami metamaterials

In this study, we present a method for generating a synthetic gauge field in origami metamaterials with continuously varying geometrical parameters. By modulating the mass term in the Dirac equation linearly, we create a synthetic gauge field in the vertical direction, which allows for the quantization of Landau levels through the generated pseudomagnetic field. Furthermore, we demonstrate the existence and robustness of the chiral zeroth Landau level. The unique elastic snake state is realized using the coupling between the zeroth and the first Landau levels. Our results, supported by theory and simulations, establish a feasible framework for generating pseudomagnetic fields in origami metamaterials with potential applications in waveguides and cloaking.

cond-mat.soft↗

arXiv2023

Geometry-informed dynamic mode decomposition in origami dynamics

Origami structures often serve as the building block of mechanical systems due to their rich static and dynamic behaviors. Experimental observation and theoretical modeling of origami dynamics have been reported extensively, whereas the data-driven modeling of origami dynamics is still challenging due to the intrinsic nonlinearity of the system. In this study, we show how the dynamic mode decomposition (DMD) method can be enhanced by integrating geometry information of the origami structure to model origami dynamics in an efficient and accurate manner. In particular, an improved version of DMD with control, that we term geometry-informed dynamic mode decomposition~(giDMD), is developed and evaluated on the origami chain and dual Kresling origami structure to reveal the efficacy and interpretability. We show that giDMD can accurately predict the dynamics of an origami chain across frequencies, where the topological boundary state can be identified by the characteristics of giDMD. Moreover, the periodic intrawell motion can be accurately predicted in the dual origami structure. The type of dynamics in the dual origami structure can also be identified. The model learned by the giDMD also reveals the influential geometrical parameters in the origami dynamics, indicating the interpretability of this method. The accurate prediction of chaotic dynamics remains a challenge for the method. Nevertheless, we expect that the proposed giDMD approach will be helpful towards the prediction and identification of dynamics in complex origami structures, while paving the way to the application to a wider variety of lightweight and deployable structures.

math.DS↗

arXiv2022

Characterization of elastic topological states using dynamic mode decomposition

Elastic topological states have been receiving increased intention in numerous scientific and engineering fields due to their defect-immune nature, resulting in applications of vibration control and information processing. Here, we present the data-driven discovery of elastic topological states using dynamic mode decomposition (DMD). The DMD spectrum and DMD modes are retrieved from the propagation of the relevant states along the topological boundary, where their nature is learned by DMD. Applications such as classification and prediction can be achieved by the underlying characteristics from DMD. We demonstrate the classification between topological and traditional metamaterials using DMD modes. Moreover, the model enabled by the DMD modes realizes the prediction of topological state propagation along the given interface. Our approach to characterizing topological states using DMD can pave the way towards data-driven discovery of topological phenomena in material physics and more broadly lattice systems.

physics.app-ph↗

arXiv2020

Topological transition in spiral elastic valley metamaterials

Elastic valley metamaterials offer an excellent platform to manipulate elastic waves and have potential applications in energy harvesting and elastography. Here we introduce a series of strategies to realize topological transition in spiral elastic valley metamaterials by parameter modulations. We show the evolution of Berry curvature and valley Chern number as a function of inherent parameters of spiral, which further results in a general scheme to achieve topological valley edge states. Our strategy leverages multiple degrees of freedom in spiral elastic valley metamaterials to provide enhanced opportunities for desired topological states.

physics.app-ph↗

arXiv2020

Digital image processing to detect subtle motion in stony coral

Coral reef ecosystems support significant biological activities and harbor huge diversity, but they are facing a severe crisis driven by anthropogenic activities and climate change. An important behavioral trait of the coral holobiont is coral motion, which may play an essential role in feeding, competition, reproduction, and thus survival and fitness. Therefore, characterizing coral behavior through motion analysis will aid our understanding of basic biological and physical coral functions. However, tissue motion in the stony scleractinian corals that contribute most to coral reef construction are subtle and may be imperceptible to both the human eye and commonly used imaging techniques. Here we propose and apply a systematic approach to quantify and visualize subtle coral motion across a series of light and dark cycles in the scleractinian coral Montipora capricornis. We use digital image correlation and optical flow techniques to quantify and characterize minute coral motions under different light conditions. In addition, as a visualization tool, motion magnification algorithm magnifies coral motions in different frequencies, which explicitly displays the distinctive dynamic modes of coral movement. We quantified and compared the displacement, strain, optical flow, and mode shape of coral motion under different light conditions. Our approach provides an unprecedented insight into micro-scale coral movement and behavior through macro-scale digital imaging, thus offering a useful empirical toolset for the coral research community.

q-bio.QM↗

arXiv2019

Valley Anisotropy in Elastic Metamaterials

Valley, as a new degree of freedom, raises the valleytronics in fundamental and applied science. The elastic analogs of valley states have been proposed by mimicking the symmetrical structure of either two-dimensional materials or photonic valley crystals. However, the asymmetrical valley construction remains unfulfilled. Here, we present the valley anisotropy by introducing asymmetrical design into elastic metamaterials. The elastic valley metamaterials are composed of bio-inspired hard spirals and soft materials. The anisotropic topological nature of valley is verified by asymmetrical distribution of the Berry curvature. We show the high tunability of the Berry curvature both in magnitude and sign enabled by our anisotropic valley metamaterials. Finally, we demonstrate the creation of valley topological insulators and show topologically protected propagation of transverse elastic waves relying on operating frequency. The proposed topological properties of elastic valley metamaterials pave the way to better understanding the valley topology and to creating a new type of topological insulators enabled by an additional valley degree of freedom.

physics.app-ph↗