In large-scale recommendation systems like LinkedIn’s, the retrieval stage is critical for narrowing billions of potential candidates to a manageable subset for ranking. LinkedIn's feed now serves suggested content based on the topical interests of members, where 2000 candidates are retrieved from several million candidates with a latency budget of a few milliseconds and inbound QPS of several thousand per second. This paper presents a novel retrieval approach that fine tunes a large causal language model (Meta’s LLaMA 3) as a dual encoder to generate high quality embeddings for both users (members) and content (items), using only textual input. We describe the end to end pipeline, including prompt design for embedding generation, techniques for fine tuning at LinkedIn scale, and infrastructure for low latency, cost effective online serving. We share our findings on how quantizing numerical features in the prompt enables the information getting encoded in the embedding facilitating greater alignment between the retrieval and ranking layer. The system was evaluated using offline metrics and an online A/B test, which showed substantial improvements in member engagement. We observed significant gains among newer members, who often lack strong network connections, indicating that high-quality suggested content aids retention. This work demonstrates how generative language models can be effectively adapted for real time, high throughput retrieval in industrial applications.

Integrated communication and sensing, which can make full use of the limited spectrum resources to perform communication and sensing tasks simultaneously, is an up-and-coming technology in wireless communication networks. In this work, we investigate the performance of an uncrewed aerial vehicle (UAV) -assisted secure integrated sensing, communication, and computing system, where the UAV sends radar signals to locate and suppress a potential eavesdropper while providing offload services to ground users (GUs). Considering the constraints of UAV maximum speed, transmit power, propulsion energy, data transmission, and computation time, the total energy consumption of GUs is minimized by jointly optimizing user offloading ratio, user scheduling strategy, transmit beamforming, and UAV trajectory. An efficient iterative optimization algorithm is proposed to solve the nonconvex optimization problem caused by tightly coupled dependent variables. In particular, the original optimization problem is decomposed into four suboptimization problems, and the nonconvex subproblems are transformed into approximately convex forms via successive convex approximation. Then, all subproblems are solved successively by using the block coordinate descent technique. Numerical results demonstrate the convergence and validate the effectiveness of the proposed algorithm.

This paper introduces LiNR, LinkedIn's large-scale, GPU-based retrieval system. LiNR supports a billion-sized index on GPU models. We discuss our experiences and challenges in creating scalable, differentiable search indexes using TensorFlow and PyTorch at production scale. In LiNR, both items and model weights are integrated into the model binary. Viewing index construction as a form of model training, we describe scaling our system for large indexes, incorporating full scans and efficient filtering. A key focus is on enabling attribute-based pre-filtering for exhaustive GPU searches, addressing the common challenge of post-filtering in KNN searches that often reduces system quality. We further provide multi-embedding retrieval algorithms and strategies for tackling cold start issues in retrieval. Our advancements in supporting larger indexes through quantization are also discussed. We believe LiNR represents one of the industry's first Live-updated model-based retrieval indexes. Applied to out-of-network post recommendations on LinkedIn Feed, LiNR has contributed to a 3% relative increase in professional daily active users. We envisage LiNR as a step towards integrating retrieval and ranking into a single GPU model, simplifying complex infrastructures and enabling end-to-end optimization of the entire differentiable infrastructure through gradient descent.

Multi-view feature learning has garnered much attention recently since many real world data are comprised of different representations or views. How to explore the consensus structure and eliminate the inconsistency noise in different views remains a challenging problem in multi-view feature learning. In this paper, we propose a multi-way deep autoencoder for multi-view feature learning to explore the deep consensus structure and reconcile the efficiency of encoding process meanwhile. Through a multi-way encoding process, we embed the original data feature views to nonnegative representations of multiple levels which are structured hierarchically. Along the structure of embedded representations, we recover the diversity and important information layer by layer in the decoding process. The experiments on two image datasets show the superior performance of our method.

To alleviate the deployment of deep object detectors with large model capacity and complex computation, an effective model compression pipeline is designed in this paper. Firstly, attributed to the refined soft filter pruning, 3D filters of each convolution layer are regularized and then auto-pruned to achieve an overall more compact backbone. Afterwards, the branch layers of the object detector are simplified with the usage of simple residual blocks and fixed channel deletion. Experimental results reveal the superior effectiveness and generality of proposed pipeline for compressing detection models. Notably, on the generic detection dataset, proposed pipeline reduces more than 67% model size on RefineDet with negligible mAP loss. Moreover, the pipeline decreases more than 73% model size on the PyramidBox face detector with little loss of hard AP on WIDER FACE.

Multi-view feature learning has garnered much attention recently since many real-world data are comprised of different representations or views. As an important and challenging problem in multi-view feature learning, the key issue is how to explore complementary information and eliminate the inconsistency noise. Motivated by this observation, we propose to encode the multiple feature views to an underlying consensus representation. To enhance the robustness against the noise of disagreement among different feature views and exploit the consistency of the shared structure, we impose the nuclear norm constraint on the consensus representation and enforce the encoded representation to be nonnegative low rank structured. The superior experimental results on three image benchmarks show the effectiveness of our method.

This paper presents an on-line admission control methodology for periodic and aperiodic service processes with end-to-end real-time constraints. Both types of service process requests dynamically join and leave a system at run time. During the admission test, the schedulability of a periodic task is determined by using its fixed task capacity. Aperiodic tasks are admitted using the available capacity after admitted periodic tasks. At run time, the earliest deadline first (EDF) scheduling is used to schedule the mixed periodic and aperiodic workloads. Simulation results show that the proposed algorithm may achieve up to 90% in system utilization, while incurring a low admission overhead for each service request.

Currently, most search engines are text-based and their structures are centralized. These kinds of engine are sufficient for searching text information in Internet. However, while searching audio resource, an efficient content-based audio search engine is required. In this paper, we demonstrate an audio search engine ASEKS based on keyword spotting technology in the peer-to-peer (P2P) network. The indexing sub-model spots information in local audio files and generates indices for later query; and the P2P networks distributes the query and gathers the results. ASEKS supports scalability and avoids the bottleneck of network load that usually exists in centralized architecture. The average accuracy of the keyword spotting sub-model is 88.4% in detection rate on the 5.267 false alarm per keyword per hour