Serving deep learning based recommendation models (DLRM) at scale is challenging. Existing approaches rely on dedicated ANN indexing and filtering services on CPUs, suffering from non-negligible costs and missing co-design opportunities. Such inefficiency makes them difficult to support complex model architectures, such as learned similarities and multi-task retrieval. In this paper, we present SilverTorch, a model-based serving system that brings all components into one unified model. It unifies model serving by replacing standalone indexing and filtering services with model layers. We propose a model-based GPU Bloom index for feature filtering and a fused Int8 ANN kernel for nearest neighbor search. Through co-design of the ANN search and feature filtering, we reduce GPU memory usage and eliminate computation. Benefiting from this design, we scale up retrieval by introducing an OverArch scoring layer and a multi-task retrieval with a Value Model to aggregate scores. These advancements improve the retrieval accuracy and enable future studies for serving more complex models. Our evaluation on industry-scale datasets shows that SilverTorch achieves up to 23.7× higher throughput compared to the state-of-the-art approaches. We also demonstrate that SilverTorch's solution is 13.35× more cost-efficient than CPU-based solution while improving accuracy via serving more complex models.

With the rise of Extended Reality (XR) technology, there is a growing need for real-time light field reconstruction from sparse view inputs. Existing methods can be classified into offline techniques, which can generate high-quality novel views but at the cost of long inference/training time, and online methods, which either lack generalizability or produce unsatisfactory results. However, we have observed that the intrinsic sparse manifold of Multi-plane Images (MPI) enables a significant acceleration of light field reconstruction while maintaining rendering quality. Based on this insight, we introduce RealLiFe, a novel light field optimization method, which leverages the proposed Hierarchical Sparse Gradient Descent (HSGD) to produce high-quality light fields from sparse input images in real time. Technically, the coarse MPI of a scene is first generated using a 3D CNN, and it is further optimized leveraging only the scene content aligned sparse MPI gradients in a few iterations. Extensive experiments demonstrate that our method achieves comparable visual quality while being 100x faster on average than state-of-the-art offline methods and delivers better performance (about 2 dB higher in PSNR) compared to other online approaches.

Neural rendering algorithms have revolutionized computer graphics, yet their impact on real-time rendering under arbitrary lighting conditions remains limited due to strict latency constraints in practical applications. The key challenge lies in formulating a compact yet expressive material representation. To address this, we propose TransGI, a novel neural rendering method for real-time, high-fidelity global illumination. It comprises an object-centric neural transfer model for material representation and a radiance-sharing lighting system for efficient illumination. Traditional BSDF representations and spatial neural material representations lack expressiveness, requiring thousands of ray evaluations to converge to noise-free colors. Conversely, real-time methods trade quality for efficiency by supporting only diffuse materials. In contrast, our object-centric neural transfer model achieves compactness and expressiveness through an MLP-based decoder and vertex-attached latent features, supporting glossy effects with low memory overhead. For dynamic, varying lighting conditions, we introduce local light probes capturing scene radiance, coupled with an across-probe radiance-sharing strategy for efficient probe generation. We implemented our method in a real-time rendering engine, combining compute shaders and CUDA-based neural networks. Experimental results demonstrate that our method achieves real-time performance of less than 10 ms to render a frame and significantly improved rendering quality compared to baseline methods.

The benefits demonstrated in smart home and energy-saving applications have sparked considerable interest in non-contact technologies for detecting indoor human presence. In opti-mizing energy utilization, precise remote assessment of individual count is of paramount importance.However, the presence of multiple human targets in indoor environments poses a challenge for quantity estimation, especially when using a simple one-transmitter-two-receiver 24GHz radar system. Due to limitations in bandwidth and the number of receiving antennas, the system exhibits poor resolution, rendering traditional distance-angle estimation algorithms ineffective in distinguishing targets.To address the indistinguishability issue of multiple targets in the distance-angle domain, we propose a target counting algorithm based on the micro-Doppler domain. Leveraging an atomic norm minimization method, this approach separates target respiratory frequencies and determines the number of targets. Through simulations and practical experiments, our approach demonstrates superior performance in target counting compared to traditional Fourier transform and compressive sensing methods.

In this paper, we present a model integrating joint sparsity and sinusoidal fitting to address the challenge of static human detection in indoor environments with interference. The motivation behind this model lies in leveraging the frequency domain sparsity induced by the micro-Doppler effect caused by human breathing. The tasks of interference suppression and human detection are formulated as an optimization problem, fitting the signals using sinusoidal functions with sparse coefficients. To effectively address this challenge, we introduce an iterative algorithm based on Improved Sparse Matching Pursuit (ISMP) within the compressed sensing framework, aiming to capture the micro-Doppler information of the target, fit the low-dimensional structure of human respiration in the phase domain, and simultaneously suppress interference micro-motion features. Compared to conventional methods using max average power and standard deviation , our model effectively suppresses interference and enhances target signals in real single-channel radar experiments, even amidst mixed interference with human targets.

Traditional supervised learning methods have achieved great success in automatic target recognition (ATR). Unfortunately, if the model is trained exclusively on new class samples, the model will forget all knowledge about the old class samples. This phenomenon is called catastrophic forgetting. Incremental learning method can prevent catastrophic forgetting by keeping a small number of old class samples as exemplars and training them together with new class samples. However, the ratio of old to new class samples is still seriously out of balance. In this paper, an data augmentation method of old class samples is proposed by using random interpolation (RI) between two exemplars to generate fake samples in the process of incremental learning. In this method, the distribution of the original classes is partially restored while also creating a numerical balance between the old and new class samples. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset demonstrate the effectiveness of this method in Incremental SAR ATR.

Millimeter-wave radar has found widespread applications in perception domains such as gesture recognition. In comparison to single-hand gestures, radar echo signals from multi-hand gestures involve the superposition of multiple signals, presenting a more challenging recognition task. This paper introduces a multi-hand gesture signal separation method based on the enhanced Multiple Signal Classification (MUSIC) algorithm, capable of simultaneously distinguishing the motion states of the left and right hands. Specifically, this paper utilizes Minimum Variance Distortionless Response (MVDR) to extract joint angle-range information for gesture features. The enhanced MUSIC algorithm is then applied to reconstruct single-hand gesture Range-Angle Map (RAM) by constructing a new noise subspace. The effectiveness of the proposed algorithm is validated through simulations and radar signal experiments, demonstrating its capability to separate multi-hand gesture signals from the foundation of multi-hand gesture RAM to obtain single hand RAM.

To enable the synthetic aperture radar (SAR) automatic target recognition (ATR) system to continuously adapt to new recognition scenarios, it is necessary to equip the system with the ability to quickly update models. However, when these models learn new tasks, the knowledge of old tasks is quickly forgotten, a phenomenon known as catastrophic forgetting. The reason for catastrophic forgetting is that the model does not use the features of old tasks sufficiently. In this letter, an exemplar-free class incremental learning based on maximizing mutual information (CIL-MMI) is proposed to solve this problem. To effectively use the extracted features, CIL-MMI actively clusters features to maximize the mutual information (MI) between features and corresponding labels. The proposed method successfully avoids the distribution overlap caused by the small interclass differences and large intraclass variances inherent in SAR images. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset indicate that the proposed method outperforms state-of-the-art approaches, demonstrating improvements of 5.41%, 1.93%, and 2.47% at incremental steps 1, 2, and 3, respectively.

The availability of a large amount of labeled data has promoted the success of deep learning in Synthetic aperture radar (SAR) target recognition tasks, but these labeled data are invariably obtained manually, which requires huge labor costs. With the development of the SAR field, SAR target categories are still increasing, so there is a large amount of data to be marked. In this paper, we use labeled data to discover novel categories, thereby reducing the cumbersome labeling process. We innovatively propose to apply AutoMix to the labeled dataset to expand the neighborhood distribution and motivate the discrete sample space continuously, which greatly improves the generalization ability of the model. Three stage learning is used to solve the problem of novel class discovery. Firstly, self-supervised learning is used to learn common features from labeled data and unlabeled data. Secondly, AutoMix is used to labeled data which is used to train the feature extractor by supervised learning. Finally, the knowledge of labeled data is transferred to unlabeled images to generate pairs of pseudo-labels for clustering. We experimentally proved that our model can effectively discover novel categories in unlabeled data.