Birefringence, the polarization-dependent splitting of light in anisotropic crystals, enables diverse optical phenomena and advanced functionalities such as optical communication, nonlinear optics, and quantum optics. However, conventional methods for controlling birefringence typically rely on engineering the optical crystal structure or applying external stimuli such as electric fields, mechanical stress, or thermal variations, which are often constrained by limited tunability, challenges in integration with compact photonic devices, or slow response time. Here, we introduce a new degree of freedom to manipulate the birefringence of light propagation in optical crystals through programming the spatiotemporal spectral phase of the incident light wave. We demonstrate that this approach achieves continuous tuning of birefringence across a spectrum more than 100 times broader than that achievable with conventional birefringence tuning, spanning from positive through zero to negative values, irrespective of the crystal's optical sign and without inherent physical limitations. This unique optical behavior provides a versatile platform for investigating the complex dynamics of wave flow in anisotropic media, while the broad tunability of this space-time birefringence will spur innovations in ultrafast optical manipulation, optical computation, and quantum information processing-applications that demand rapid and flexible device reconfiguration.

Tapered optical fibers (TFs), with diameters gradually reduced from hundreds of microns to the micron scale, offer key advantages over conventional flat optical fibers (FFs), including uniform illumination, efficient long-range signal collection, and minimal invasiveness for applications in high-sensitivity biosensing, optogenetics, and photodynamic therapy. However, high-fidelity, single-shot imaging through a single TF remains underexplored due to intermodal coupling from the tapering geometry, which distorts output speckle patterns and poses challenges for image reconstruction using existing deep learning methods. Here, we propose a physics-inspired TF-UNet architecture that augments skip connections with hierarchical grouped-MLP fusion to effectively capture non-local, cross-scale dependencies caused by intermodal coupling in TFs. We experimentally validate our method on both FFs and TFs, demonstrating that TF-UNet outperforms standard U-Net variants in structural and perceptual fidelity while maintaining competitive PSNR at quadratic complexity. Our study offers a promising approach for deep learning-based imaging through micron-sized, ultrafine optical fibers, enabling scanning-free single-shot reconstruction on a 512 & times; 512 reconstruction matrix, and further validating the framework on biologically meaningful neuronal and vascular datasets for physically interpretable characterization.

Abstract Intracortical microstimulation (ICMS) with ultraflexible neural electrodes enables low-threshold, chronically stable, and high-resolution modulation of neural circuits, providing a promising strategy for sensory restoration and closed-loop neuromodulation. However, the microscopic mechanisms delineating its safe and effective current range remain unclear. Here, we combine intravital two-photon (2P) imaging and electrophysiology in awake mice to examine the current-dependent neurovascular outcomes of charge-balanced stimulation via ultraflexible arrays. We observed gas bubbles formed along the electrode during ICMS, with bubble size increasing quadratically with current amplitude, consistent with a Faradaic bubble-growth model. Intravital 2P imaging reveals that at low-to-moderate currents (20–40 µA), vascular leakage is small, spatially confined, and largely reversible, whereas higher currents (≥60 µA) induce a sharp transition to extensive, field-dominated extravasation and secondary vessel disruption. This transition coincides with immediate, stimulus-locked motor responses and the onset of electrode degradation. Multiphysics simulations reproduce the observed nonlinear leakage–current relationship by incorporating gas bubble–induced electric field redistribution and voltage-dependent vessel wall permeability. The model indicates that gas bubbles act as local electric-field modulators, concentrating suprathreshold fields near the bubble boundary at lower currents while shielding more distant vessel segments; at higher currents, this confinement breaks down and the system enters a field-dominated damage regime. Collectively, these findings define a mechanistically informed safety window for ICMS with flexible neural interfaces and identify bubble-assisted vascular permeabilization as a key failure mode at high currents, crucial for the design of future bidirectional brain-computer interfaces and high-precision neuroprosthetic protocols.

High-precision in vivo monitoring of ion fluxes is essential yet challenging for the study of plant electrophysiology, including growth regulation, signal transduction and stress responses. Existing methods for probing ion dynamics are limited by low sensitivity, high invasiveness that interferes physiological processes, and the inability to accurately resolve intracellular ion homeostasis with sufficient spatial and temporal resolution. Here, we introduce plant intracellular nanoelectrode (PINE) arrays manufactured on 1-μm-thick polymer substrates, which enable ultrasensitive and selective measurement of ionic current via scalable nanofabrication techniques. The fabricated PINE arrays possess dimensions smaller than those of typical plant cells and possess reduced mechanical stiffness, facilitating minimally invasive integration with living plant cells. This subcellular-scale plant-electronic interface allows for reliable, selective intracellular detection of K+ flux with a detection limit as low as ~10-8 M. We demonstrate continuous, stable monitoring in tomato stem cells over six weeks, accurately capturing dynamic potassium fluctuations throughout all key growth stages. The proposed approach supports long-term, real-time tracking of ion-specific intracellular dynamics without disrupting plant cellular structures or altering endogenous ion concentrations. By providing unprecedented access to intracellular ion homeostasis and signaling networks, PINE represents a powerful platform for advancing precision agriculture and enabling future digital plant engineering. ### Competing Interest Statement The authors have declared no competing interest. National Natural Science Foundation of China, 12388102 Zhangjiang Laboratory Youth Innovation Project, ZJYI2022A01, S202420005 CAS Pioneer Hundred Talents Program Shanghai Science and Technology Committee Program, 23560750200

Small and flexible penetrating neural electrodes have recently emerged as a promising technology for both fundamental brain science and advanced brain-computer interfaces. These neural electrodes are designed with exceptional flexibility and adaptability to seamlessly interact with delicate neural tissue, enabling sub-millisecond recording of individual neurons and stimulation of small neuronal clusters in the brain. In this review, we analyze the fundamental physical constraints imposed on electrical neural interfacing and highlight the significant progress achieved by those minimally invasive neural probes over the past decade. Furthermore, we discuss the research needs in manufacturing techniques, materials science, as well as surface functionalization processes for improving the long-term stability and reliability of neural electrodes. Finally, we summarize the emerging trends and outline the technological challenges faced in this rapidly evolving field.

Tapered optical fibers (TFs), with diameters gradually reduced from hundreds of microns to the micron scale, offer key advantages over conventional flat optical fibers (FFs), including uniform illumination, efficient long-range signal collection, and minimal invasiveness for applications in high-sensitivity biosensing, optogenetics, and photodynamic therapy. However, high-fidelity, single-shot imaging through a single TF remains underexplored due to intermodal coupling from the tapering geometry, which distorts output speckle patterns and poses challenges for image reconstruction using existing deep learning methods. Here, we propose a physics-inspired TF-UNet architecture that augments skip connections with hierarchical grouped-MLP fusion to effectively capture non-local, cross-scale dependencies caused by intermodal coupling in TFs. We experimentally validate our method on both FFs and TFs, demonstrating that TF-UNet outperforms standard U-Net variants in structural and perceptual fidelity while maintaining competitive PSNR at quadratic complexity. Our study offers a promising approach for deep learning-based imaging through micron-sized, ultrafine optical fibers, enabling scanning-free single-shot reconstruction on a 512x512 reconstruction matrix, and further validating the framework on biologically meaningful neuronal and vascular datasets for physically interpretable characterization.

Objective The nervous system serves as the primary communication system in animals. Neurons, the fundamental structural and functional units of this system, communicate through a combination of electrical and chemical signals. Deciphering and comprehending diverse neural activities and circuit functions are of paramount importance in the realms of fundamental brain science, the diagnosis and treatment of neurological disorders, and brain-computer interface applications. Integrating optogenetics and electrophysiology into an optoelectric neural interface offers a synergistic approach to studying complex brain circuits and unraveling their intricate dynamics, enabling researchers to observe and modulate neuronal activity with precision. This capability opens new avenues for investigating fundamental questions about how different brain regions communicate and contribute to behavior. By combining optogenetics and electrophysiology to create advanced optoelectric neural interfaces, researchers can gain unprecedented insights into brain functions. Methods An optogenetic stimulation system was integrated with an electrophysiological recording system to measure photoelectric artifacts. A 473 nm fiber-coupled laser served as the light source for optogenetic stimulation, with a pulse generator employed to control the laser pulses. Electrophysiological signals were recorded using an Intan 1024-channel electrophysiological recording system. The optoelectrodes were fabricated using tapered optical fibers and ultra-flexible electrodes, which were then implanted into either the mouse brain or an agarose gel phantom to capture photoelectric artifacts in vivo or in vitro. The ultra-flexible neural electrode was fabricated using planar semiconductor technology, incorporating a polyimide insulation layer and a gold wire layer, as described in our previous publications. In addition, the electrode surface was modified with PEDOT & ratio;PSS to enhance electrophysiological recording performance. The tapered optical fiber, supplied by Optogenix, featured a numerical aperture (NA) of 0.39, core size of 200 mu m, and an active length of 2.5 mm. The optoelectrode probe was assembled by temporarily bonding the optical fiber to the ultra-flexible electrode using polyethylene glycol (PEG, m(PEG)=400000 u). The optical performance of the fabricated optoelectrode was characterized through theoretical calculations using LightSpread software, as well as experimental verification in vitro and in vivo. In the in vitro measurements, powdered milk, agarose, and sodium fluorescein were used to simulate tissue scattering and assess the light-field distribution of both tapered and flat port fibers in a scattering medium. For the in vivo demonstrations, optoelectrodes were implanted into the mouse CA1 brain region to perform concurrent optogenetic stimulation and electrical recording. Electrophysiological signals were filtered and analyzed using MATLAB software. The peak value of the photoelectric artifact was defined as the maximum absolute voltage observed during the optical pulse. The power spectral density (PSD) of the local field potentials during optical stimulation was obtained using a short-time Fourier transform, and the Mountainsort4 algorithm was employed for peak potential cluster analysis to isolate the waveform and timestamp of the action potentials. Results and Discussions We designed and fabricated a novel optoelectrode that combines a tapered fiber with an ultra-flexible electrode (Fig. 1). The tapered fibers exhibit an extended illumination range and a more uniform intensity distribution compared to flat- port fibers in a scattering medium (Fig. 2). In vitro experiments reveal variations in photoelectric artifacts across different channels, optical powers, and media types, with the peak values of photoelectric artifacts increasing alongside higher concentrations of milk powder and laser power. Power spectral density analysis indicates that photoelectric artifacts predominantly occur within the frequency range below 10 Hz (Fig. 3). During in vivo experiments, we analyzed the impact of light stimulation on the frequency bands of local field potentials (LFP), ranging from 0 to 300 Hz, and action potentials (AP), ranging from 300 to 7500 Hz. Our findings indicate that photoelectric artifacts primarily affect the LFP signals. Additionally, we longitudinally assessed the impedance evolution of the optoelectrode post-implantation and observed a gradual increase in average impedance during the first week, followed by stabilization over the subsequent three weeks. The peak value of photoelectric artifacts initially increases during the first two weeks, followed by a gradual decline over the next two weeks. Power spectral density analysis reveals that light stimulation predominantly influenced electrophysiological signals below 10 Hz, consistent with the in vitro testing results (Fig. 4). Finally, we validated the capabilities of optogenetic stimulation and synchronous electrophysiological recordings using the optoelectrode. Conclusions In this study, we present the design and fabrication of a novel optoelectrode that combines a tapered fiber with an ultra-flexible neural electrode. A comparative analysis of the optical power density and optical field distributions in a scattering medium was conducted between the tapered flat-port fibers, revealing the superior optogenetic stimulation performance of the tapered fiber. Moreover, we investigated the impact of photoelectric artifacts from the optoelectrode on electrophysiological recordings. The in vitro test results reveal variations in photoelectric artifacts across different electrode channels, environmental conditions, and laser powers. In vivo experiments demonstrate that optical stimulation primarily influences the LFP band, whereas the electrochemical impedance of the optoelectrode gradually increases and eventually stabilizes over time. The peak value of photoelectric artifacts varies depending on the duration of implantation. Photoelectric artifacts primarily induce interference within the frequency range below 10 Hz. To mitigate these artifacts, future studies could explore the utilization of coating materials such as PBK, PGO, and Pt-Black/PEDOT- GO. In addition, incorporating principal component analysis or machine learning techniques during data processing, employing a longer-wavelength excitation light source, or adjusting the electrode distance from the light source are all avenues worth investigating.

We present a novel modification for an ultra-flexible neural probe capable of simultaneous electrophysiological recording and dopamine detection within cerebral tissue. Following electrode modification, the electrochemical stability of the electrode coating was enhanced, and its specific surface area was increased, thereby significantly improving its sensitivity to dopamine. Building upon this foundation, the ultra-flexible design minimizes damage to brain tissue, enabling stable simultaneous detection of electrophysiological and electrochemical signals for over six weeks. The high-density electrode array design facilitates concurrent monitoring of activities across multiple brain regions. This novel approach provides new methodologies for neuroscience research and the treatment of brain disorders.

Microinfarcts, the "invisible lesions", are prevalent in aged and injured brains and associated with cognitive impairments, yet their neurophysiological impact remains largely unknown. Using a multimodal chronic neural platform that combines functional microvasculature imaging with spatially resolved neural recording, the neurovascular effect of a single microinfarct is investigated. Unlike larger strokes, microinfarcts induced only temporary suppression of neural activity with minimal cell death, with recovery paralleling vasculature remodeling at the infarct core. Neural activity is more severely suppressed at the shallower cortical layer despite milder vascular damage compared to deeper layers, and the excitability of fast-spiking interneurons attenuation is accompanied by heightened bursting of regular spiking neurons. Spike phase locking at the low-gamma band is disrupted, indicating a lasting impairment of long-range assembly communication. These results highlight the subtle yet significant neurovascular disruptions of a single microinfarct.

Ultraflexible neural electrodes have shown superior stability compared with rigid electrodes in long-term in vivo recordings, owing to their low mechanical mismatch with brain tissue. It is desirable to detect neurotransmitters as well as electrophysiological signals for months in brain science. This work proposes a stable electronic interface that can simultaneously detect neural electrical activity and dopamine concentration deep in the brain. This ultraflexible electrode is modified by a nanocomposite of reduced graphene oxide (rGO) and poly(3,4-ethylenedioxythiophene):poly(sodium 4-styrenesulfonate) (rGO/PEDOT:PSS), enhancing the electrical stability of the coating and increasing its specific surface area, thereby improving the sensitivity to dopamine response with 15 pA/mu M. This electrode can detect dopamine fluctuations and can conduct long-term, stable recordings of local field potentials (LFPs), spiking activities, and amplitudes with high spatial and temporal resolution across multiple regions, especially in deep brain areas. The electrodes were implanted into the brains of rodent models to monitor the changes in neural and electrochemical signals across different brain regions during the administration of nomifensine. Ten minutes after drug injection, enhanced neuronal firing activity and increased LFP power were detected in the motor cortex and deeper cortical layers, accompanied by a gradual rise in dopamine levels with 192 +/- 29 nM. The in vivo recording consistently demonstrates chronic high-quality neural signal monitoring with electrochemical signal stability for up to 6 weeks. These findings highlight the high quality and stability of our electrophysiological/electrochemical codetection neural electrodes, underscoring their tremendous potential for applications in neuroscience research and brain-machine interfaces.

Lead-free V2O5-TeO2 glass systems are promising new sealant materials for electronics vacuum packaging, with low-melting temperature and unique light absorption. The glass forming regularity, structure and corresponding properties of the V2O5-TeO2-RO (R=Ca, Sr, Ba) glass system are systematically studied and analyzed in this paper. The results show that there is a wide glass forming region for the V2O5-TeO2-RO glass system, and with the increase of TeO2/V2O5 molar ratio, the thermal stability of V2O5-TeO2-SrO glass with the RO content of 20 mol% increases and the coefficient of thermal expansion increase from 11.2 to 15.86 x 10-6/degrees C due to transformation process of glass network structure from the [VO4] to transition state [TeO3 + 1] and [TeO3] group. The higher initial crystallization temperature of 40 mol%V2O5-40 mol%TeO2-20 mol%SrO glass is conducive to the flow of the glass slurry during the sealing process, obtaining a sealing device with good airtightness at extremely low temperatures of 350 degrees C.

Abstract The presence of grand minima, characterized by significantly reduced solar and stellar activity, brings a challenge to the understanding of solar and stellar dynamo. The Maunder Minimum (1645–1715 AD) is a representative grand solar minimum. The cyclic variation of solar activity, especially the cycle length during this period, is critical to understand the solar dynamo but remains unknown. By analyzing the variations in solar activity‐related equatorial auroras recorded in Korean historical books in the vicinity of a low‐intensity paleo‐West Pacific geomagnetic anomaly, we find clear evidence of an 8‐year solar cycle rather than the normal 11‐year cycle during the Maunder Minimum. This result provides a key constraint on solar dynamo models and the generation mechanism of grand solar minima.

In this paper, the modification and strengthening dual functionality of rare-earth oxide is authenticated by the effect La2O3/BaO on composition-structure-property correlations of new type La2O3-BaO-SiO2 (LBS) glass-ceramics sealing material with high softening point, high coefficient of thermal expansion and ex-cellent high temperature resistivity. As the network modifier, the La2O3 can create more non-bridging oxygen than the BaO, resulting in the increase of Q0 and Q1 units and the decrease of Q2 and Q3 units in the La2O3-BaO-SiO2 glass according to Raman and 29Si Solid-State Nuclear Magnetic Resonance. The strengthening functionality of La2O3 is declared by Static 139La WURST-QCPMG NMR spectra that La3+ ions form La-O bonds with O2-ions in the glass structure and La3+ spectra widens slightly with the increasing of La2O3. Correspondingly, physical properties including Tg, Tc, rho, and MV of La2O3-BaO-SiO2 glass gradually increase with increasment of effective cation field strength. The main crystal phase of LBS glass-ceramic transforms from Ba2Si3O8 to Ba5Si8O21 and furthur to Ba3Si5O13 phase with high coefficient of thermal expansion and there is not lanthanum-contained phase precipitation. The coefficients of thermal expansion (CTE) of LBS glass-ceramic increases from 12.48 x 10-6/degrees C to 13.18 x 10-6/degrees C (30-1000 degrees C) and the softening temperature are between 1231.1 degrees C and 1305.1 degrees C. The direct-current (DC) resistivity of the LBS glass -ceramics is higher 106 omega center dot cm at a high temperature of 700 degrees C, which has good high-temperature electrical insulation. All in all, the high CTE, high Ts and high electrical insulation "Tri-high" properties of rare-earth oxide La2O3-BaO-SiO2 glass-ceramic will be an excellent candidate for solid oxide fuel cell, solid oxide electrolysis cell, oxygen sensors, and so on.(c) 2023 Elsevier B.V. All rights reserved.

Significance The nervous system is composed of different types of neurons connected in a network. Neurons communicate with each other through electrochemical signals, and this dynamic interaction of neurons is the internal driving force of human perception, cognition, and behavior. Deciphering and understanding the meaning of various complex neural activities is of great significance in the frontiers of neurological disease diagnosis and treatment, neurological rehabilitation, fundamental brain science, and several brain -computer interface applications. To achieve this, it is critical to develop advanced neural interfaces capable of interacting with the dynamic neural activities and nervous system. Fundamental research in this field has rapidly increased over the past few decades as a result of advancements in neuroscience and neurotechnology. This research includes the development of innovative neural recording and modulation tools that have provided researchers with an early glimpse into previously unanswerable questions, such as determining how the mind works, or which have been recognized and funded by a host of initiatives. In recent years, the United States, European countries, Japan, and China have launched their own brain initiatives to support this emerging field. In the future, in order to completely understand the complex structural and functional nervous system, more powerful tools must be developed to record, transmit, and modulate signals using multiple approaches. These tools must have the ability to manipulate neuron cell types specifically while minimizing side effects such as "the observer effect." Neurological disorders affect more than one billion people worldwide, accounting for 7% of the total global disease burden, and this number is expected to increase substantially as human life expectancy increases and with increased population aging. This is largely due to neuropsychiatric diseases (including Alzheimer s disease, Parkinson s disease, epilepsy, and so on) and cerebrovascular diseases, which impose a heavy burden on society and individuals, while also promoting advances and developments in neuroengineering, biomedical science, and technology. Currently, the treatment of these diseases relies primarily on drug therapy or implantable electrical stimulation devices, such as injecting current into the target tissue through metal electrodes to activate or suppress the action potential of neurons, as well as to achieve therapeutic purposes, including using cochlear implants, deep brain stimulators, spinal cord electrical stimulation, and visual prostheses to reduce symptoms or restore nerve functions. Progress With the enrichment of multiple neural modalities, neural technologies and tools have been increasingly augmented, and have been widely used to collect neuron activities in vivo from individual neurons to neuron populations in different brain areas, with a variety of signal recording and modulation manners (Fig. 1). In addition, advances in genetic engineering, especially optogenetics, allow us to continuously control specific types of neurons with high accuracy and fidelity over a short period. The rapid development of genetically encoded neural probes provides new avenues for real-time and high-speed neural recording. Various optical, electrical, and chemical tools have been developed to record and modulate neural activities. Currently, the monolithic integration of multiple functional features has become a pressing demand and challenge in neural engineering, while flexible neural implants are expected to establish seamless integration with the soft biological tissue and achieve a high -bandwidth close-loop interaction with the nervous system. It will provide a powerful tool for identifying complex neural circuits, as well as diagnosing and treating neurological diseases. In order to more accurately understand the brain neural network and its working mechanism, and to treat neurological diseases with high selectivity, it is necessary to simultaneously monitor neural activities with high spatial and temporal resolution. The combination of electrophysiological and optical methods (for example, two -photon imaging and electrophysiological recording) can maximize the synergistic effect of the two methods, making up for the shortcomings of each method. Implantable multimodal neural interfaces integrate these approaches by maximizing their benefits and efficacy, providing neuroscientists with new access to the brain and revolutionizing applications such as the treatment and rehabilitation of neurological diseases. In order to achieve this, we need to understand the various neurotechnologies individually, how they function, as well as how they work together. Implantable neural interfaces have already been successfully employed in the long-term stable interrogation of large-scale neural activities and clinically restoring sensorimotor function in disabilities. However, they are limited in the long run by poor biocompatibility, mechanical mismatches between the device and neural tissue, and the risk of chronic inflammatory reactions after implantation. In addition, traditional neural probes are limited by spatial and temporal resolution and scalabilities, and still face challenges in the study of large-scale neural networks in situ (Fig. 1). To this end, long-term stability is achieved by matching the mechanical properties of the implanted device with those of the internal biological tissue. High spatial and temporal resolution and even specificity can be obtained by reducing the feature size of implants and mimicking the structural morphology of neurons. Multimodal neural interfaces are currently emerging, in particular, a variety of clinical multimodal implantable devices have been developed to treat neurodegenerative diseases such as Parkinson s disease, epilepsy, and depression. However, traditional device designs, such as electrophysiological readout, fluorescence cell imaging, and the structural dynamics of the brain, may conflict with each other during different data acquisition processes. Severe electrophysiological signal contamination caused by photoelectric (magnetic) artifacts can also occur. Over the past decade, there have been efforts to address these challenges, and many excellent results have emerged regarding the latest advances in neural technologies and applications. Conclusions and Prospects With the development of neural probe structures and materials, as well as the innovation of synthetic technologies for nanoparticles, dye molecules, and genetically encoded proteins, it is expected that neural technology will continue to be developed toward the limits of the lifetime, localization and specificity of neural recording and stimulation, and will eventually blur the boundary between living biological tissue and physical equipment and tools. These cutting-edge neural technologies, which combine advanced optical and nanoelectronic technologies, optogenetics, genetically encoded indicators, and acoustic and magnetic tools, provide us with unprecedented opportunities for new multimodal neural information interactions, with which powerful paradigms for multimodal inquiry of brain activity will be foreseen, and will even fundamentally alter how brain activity maps to the physical world.

The Scientific Experimental system in Near SpacE(SENSE)consists of different types of instruments that will be installed on a balloon-based platform to characterize near-space environmental parameters.As one of the main scientific payloads,the middle and near ultraviolet spectrograph(MN-UVS)will provide full spectra coverage from middle ultraviolet(MUV,200-300 nm)to near ultraviolet(NUV,300-400 nm)with a spectral resolution of 2 nm.Its primary mission is to acquire data regarding the UV radiation background of the upper atmosphere.The MN-UVS is made up of six primary components:a fore-optical module,an imaging grating module,a UV intensified focal plane module,a titanium alloy frame,a spectrometer control module,and a data processing module.This paper presents in detail the engineering design of each functional unit of the MN-UVS,as well as the instrument's radiometric calibration,wavelength calibration,impact test,and low-pressure discharge test.Furthermore,we are able to report ground test and flight test results of high quality,showing that the MN-UVS has a promising future in upcoming near-space applications.

Although various methods have been proposed for pedestrian attribute recognition, most studies follow the same feature learning mechanism, \ie, learning a shared pedestrian image feature to classify multiple attributes. However, this mechanism leads to low-confidence predictions and non-robustness of the model in the inference stage. In this paper, we investigate why this is the case. We mathematically discover that the central cause is that the optimal shared feature cannot maintain high similarities with multiple classifiers simultaneously in the context of minimizing classification loss. In addition, this feature learning mechanism ignores the spatial and semantic distinctions between different attributes. To address these limitations, we propose a novel disentangled attribute feature learning (DAFL) framework to learn a disentangled feature for each attribute, which exploits the semantic and spatial characteristics of attributes. The framework mainly consists of learnable semantic queries, a cascaded semantic-spatial cross-attention (SSCA) module, and a group attention merging (GAM) module. Specifically, based on learnable semantic queries, the cascaded SSCA module iteratively enhances the spatial localization of attribute-related regions and aggregates region features into multiple disentangled attribute features, used for classification and updating learnable semantic queries. The GAM module splits attributes into groups based on spatial distribution and utilizes reliable group attention to supervise query attention maps. Experiments on PETA, RAPv1, PA100k, and RAPv2 show that the proposed method performs favorably against state-of-the-art methods.

Although various methods have been proposed for multi-label classification, most approaches still follow the feature learning mechanism of the single-label (multi-class) classification, namely, learning a shared image feature to classify multiple labels. However, we find this One-shared-Feature-for-Multiple-Labels (OFML) mechanism is not conducive to learning discriminative label features and makes the model non-robustness. For the first time, we mathematically prove that the inferiority of the OFML mechanism is that the optimal learned image feature cannot maintain high similarities with multiple classifiers simultaneously in the context of minimizing cross-entropy loss. To address the limitations of the OFML mechanism, we introduce the One-specific-Feature-for-One-Label (OFOL) mechanism and propose a novel disentangled label feature learning (DLFL) framework to learn a disentangled representation for each label. The specificity of the framework lies in a feature disentangle module, which contains learnable semantic queries and a Semantic Spatial Cross-Attention (SSCA) module. Specifically, learnable semantic queries maintain semantic consistency between different images of the same label. The SSCA module localizes the label-related spatial regions and aggregates located region features into the corresponding label feature to achieve feature disentanglement. We achieve state-of-the-art performance on eight datasets of three tasks, \ie, multi-label classification, pedestrian attribute recognition, and continual multi-label learning.

Although various methods have been proposed for pedestrian attribute recognition, most studies follow the same feature learning mechanism, \ie, learning a shared pedestrian image feature to classify multiple attributes. However, this mechanism leads to low-confidence predictions and non-robustness of the model in the inference stage. In this paper, we investigate why this is the case. We mathematically discover that the central cause is that the optimal shared feature cannot maintain high similarities with multiple classifiers simultaneously in the context of minimizing classification loss. In addition, this feature learning mechanism ignores the spatial and semantic distinctions between different attributes. To address these limitations, we propose a novel disentangled attribute feature learning (DAFL) framework to learn a disentangled feature for each attribute, which exploits the semantic and spatial characteristics of attributes. The framework mainly consists of learnable semantic queries, a cascaded semantic-spatial cross-attention (SSCA) module, and a group attention merging (GAM) module. Specifically, based on learnable semantic queries, the cascaded SSCA module iteratively enhances the spatial localization of attribute-related regions and aggregates region features into multiple disentangled attribute features, used for classification and updating learnable semantic queries. The GAM module splits attributes into groups based on spatial distribution and utilizes reliable group attention to supervise query attention maps. Experiments on PETA, RAPv1, PA100k, and RAPv2 show that the proposed method performs favorably against state-of-the-art methods.

Video instance segmentation (VIS) aims at segmenting and tracking objects in videos. Prior methods typically generate frame-level or clip-level object instances first and then associate them by either additional tracking heads or complex instance matching algorithms. This explicit instance association approach increases system complexity and fails to fully exploit temporal cues in videos. In this paper, we design a simple, fast and yet effective query-based framework for online VIS. Relying on an instance query and proposal propagation mechanism with several specially developed components, this framework can perform accurate instance association implicitly. Specifically, we generate frame-level object instances based on a set of instance query-proposal pairs propagated from previous frames. This instance query-proposal pair is learned to bind with one specific object across frames through conscientiously developed strategies. When using such a pair to predict an object instance on the current frame, not only the generated instance is automatically associated with its precursors on previous frames, but the model gets a good prior for predicting the same object. In this way, we naturally achieve implicit instance association in parallel with segmentation and elegantly take advantage of temporal clues in videos. To show the effectiveness of our method InsPro, we evaluate it on two popular VIS benchmarks, i.e., YouTube-VIS 2019 and YouTube-VIS 2021. Without bells-and-whistles, our InsPro with ResNet-50 backbone achieves 43.2 AP and 37.6 AP on these two benchmarks respectively, outperforming all other online VIS methods.