Performance collapse is an intractable issue of Differentiable Architecture Search (DAS), where severe performance degradation of DAS happens when it trains on different search spaces or datasets. We theoretically analyze the issue from the information bottleneck (IB) perspective, and disclose that a solution to overcome this problem is to seek the bifurcation point of IB tradeoff between compression and prediction of the supernet. To this end, we propose a simple yet highly effective method, namely, Batch Entropy-decay Regularization (BER), to guide the learning of DAS, which restricts compression in DAS by imposing a penalty on the architecture parameters. Comprehensive theoretical analyses demonstrate that BER is able to completely resolve DAS's performance collapse issue. Compared with a number of state-of-the-art DAS variants, BER shows its overwhelmingly better performance on 7 search spaces (i.e., NAS-Bench-201, DARTS, S1-S4, MobileNet-like) and 5 popular datasets (i.e., CIFAR-10, CIFAR-100, ImageNet1k, PASCAL VOC 2007, and MS COCO 2017).

With the rapid advancement of large language models (LLMs) in both academia and industry, their growing size and complexity have introduced significant challenges in terms of computational cost and deployment efficiency. To address these issues, a wide range of inference optimization techniques-including but not limited to model compression-have been proposed to accelerate LLM inference while preserving model performance. This survey provides a comprehensive overview of LLM inference acceleration strategies, analyzing them from multiple perspectives, including foundational principles, algorithmic techniques, real-world applications, and open research challenges. We begin by introducing core concepts underlying inference optimization and propose a new taxonomy that categorizes existing approaches, including quantization, pruning, distillation, efficient architectures, compilation, and hardware-aware methods. Following the lifecycle of LLM development and deployment, we examine how these techniques interact with model training, fine-tuning, and serving. Furthermore, we highlight key applications of efficient LLMs and discuss emerging trends and unresolved issues in the field. By synthesizing recent advances, this survey aims to provide actionable insights and practical guidance for researchers and practitioners working with scalable and efficient LLM systems.

Neural Architecture Search (NAS) has recently outperformed hand-designed networks in various artificial intelligence areas. However, previous works only target a pre-defined task. For a new task in few-shot learning (FSL) scenarios, the architecture is either searched from scratch, which is neither efficient nor flexible, or borrowed architecture from the ones obtained on other tasks, which may lead to sub-optimal. Can we select the best neural architectures without involving any training and eliminate a significant portion of the search cost for new tasks in FSL? In this work, we provide an affirmative answer by proposing a novel information bottleneck (IB) theory driven Few-shot Neural Architecture Search (dubbed, IBFS) framework to address this issue. We first derive that the global convergence of Model-agnostic meta-learning (MAML) can be guaranteed by only considering the first-order loss landscape. Moreover, motivated by the observation that IB provides a unified view toward understanding machine learning models, we propose a novel Zero-Cost method tailored for FSL to rank and select architectures based on their expressivity obtained by IB mechanisms. Extensive experiments show that IBFS achieves state-of-the-art performance in FSL without training, which demonstrates the effectiveness of our IBFS.

Mixed precision quantization (MPQ) is an effective quantization approach to achieve accuracy-complexity trade-off of neural network, through assigning different bit-widths to network activations and weights in each layer. The typical way of existing MPQ methods is to optimize quantization policies (i.e., bit-width allocation) in a gradient descent manner, termed as Differentiable MPQ (DMPQ). At the end of the search, the bit-width associated to the quantization parameters which has the largest value will be selected to form the final mixed precision quantization policy, with the implicit assumption that the values of quantization parameters reflect the operation contribution to the accuracy improvement. While much has been discussed about the MPQ’s improvement, the bit-width selection process has received little attention. We study this problem and argue that the magnitude of quantization parameters does not necessarily reflect the actual contribution of the bit-width to the task performance. Then, we propose a Shapley-based MPQ (SMPQ) method, which measures the bit-width operation’s direct contribution on the MPQ task. To reduce computation cost, a Monte Carlo sampling-based approximation strategy is proposed for Shapley computation. Extensive experiments on mainstream benchmarks demonstrate that our SMPQ consistently achieves state-of-the-art performance than gradient-based competitors.

Video anomaly detection (VAD) is vital for public safety, yet current approaches struggle with limited generalization, low interpretability, and high resource demands. To address these challenges, we propose HoloTrace, an edge-cloud collaborative VAD system that integrates large language models (LLMs) to construct and update a novel bidirectional causal knowledge graph. At the edge, HoloTrace leverages LLM-based cross-modal understanding and employs Hidden Markov Model (HMM) for bidirectional event reasoning, obtaining anomaly boundaries with low computational overhead. On the cloud side, LLMs are leveraged to dynamically update the Bi-CKG graph with key frames sent from the edge, in order to update causal relationships between events. Additionally, we introduce SVAD, a new large-scale VAD dataset comprising 632 real-world surveillance videos across 10 anomaly types and diverse scenes, with manually labeled frame-level annotations. Experimental results demonstrate that HoloTrace not only achieves the highest accuracy but also enhances interpretability and efficiency, paving the way for more generalizable and explainable video anomaly detection systems.

The success of computer vision tasks is mainly attributed to the architectural design of neural networks. This highlights the need to automatically design high-performance architectures via Neural Architecture Search (NAS). To accelerate the search process, training-free NAS is proposed, which aims to search high-performance architectures at initialization via zero-cost proxies (ZCPs). However, existing zero-cost proxies heavily rely on manual design, which is often labor-intensive and requires extensive expert knowledge. In addition, these crafted proxies often suffer from poor correlation with final model performance and high computational complexity, severely limiting NAS efficiency in real-world applications. To address those issues, this paper proposes a novel Large Language Models (LLMs)-driven $\underline{A}$utomatic $\underline{P}$roxy $\underline{D}$iscovery ($\textbf{APD}$) framework, which revolutionizes the design paradigm of ZCPs by leveraging LLMs to automatically discover optimal ZCPs for Training-Free NAS. Moreover, we utilize actor-critic based reinforcement learning to optimize prompts, enabling to generate better ZCPs in the next generation. We conduct extensive experiments on mainstream NAS benchmarks, demonstrating APD excels in both performance and efficiency. Besides, we firmly believe that our APD will dramatically benefit the deep learning community through providing novel paradigm of design algorithms via LLMs.

Recently, visual sensors have been deployed almost everywhere, generating substantial volumes of surveillance video data in smart cities. Anomaly detection aims to detect anomalous events for smart surveillance video analytics. However, exsiting methods rely on manually designed architectures to learn and extract feature representations, which is inefficient and labor intensive. To address this bottleneck, in this work, we propose a Neural Architecture Search (NAS) method (AdNAS), which aims to automatically find optimal architecture for anomaly detection in surveillance video. First, we design search space for anomaly detection task, which contains various convolutional operators and a weight-sharing supernet. Moreover, we we present a differentiable strategy to facilitate the deployment of latency-aware architectures across a range of edge platforms in an end-to-end manner. Extensive experiments show that our AdNAS achieves competitive performance in several benchmarks, especially achieving highest mean Average Precision (mAP) on the fire dataset-1 for anomaly detection task.

Deep learning has emerged in many practical applications, such as vascular segmentation, emotion recognition, and target detection. Moreover, convolutional neural networks (CNNs), representative techniques of deep learning, have been used to solve face recognition. However, the current design of CNNs for face recognition is highly dependent on domain knowledge and needs a large amount of trial and error. In this paper, we propose a new method based on the brain storm optimization algorithm to tackle the network structure and training parameters selection problem of CNN for face detection. In our method, an efficient mixed-length encoding strategy is designed to represent the CNN network information and the model training parameters. Specifically, convolutional layers, pooling layers, and fully connected layers can be designed automatically, and then an optimal set of selection results can be obtained. A series of experimental results show that the architecture by our method achieves higher accuracy (98.7%) compared to the state-of-the-art results on the FDDB dataset, as well as competitive results (accuracy = 94.3%, 92.5%, and 84.9%) on the WIDER FACE dataset.

Performance predictors are used to reduce architecture evaluation costs in neural architecture search, which however suffers from a large amount of budget consumption in annotating substantial architectures trained from scratch. Hence, how to leverage existing annotated architectures to train a generalized predictor to find the optimal architecture on unseen target search spaces becomes a new research topic. To solve this issue, we propose a Single-Domain Generalized Predictor (SDGP), which aims to make the predictor only trained on a single source search space but perform well on target search spaces. In meta-learning, we firstly adopt feature extractor in learning the domain-invariant features of the architectures. Then, a neural predictor is trained to map the architectures to the accuracy of the candidate architectures over the target domain simulated on the source search space. Moreover, a novel multi-head attention driven regularizer is designed to regulate the predictor to further improve the generalization ability of the predictor for the feature extractor. A series of experimental results have shown that the proposed predictor outperforms the state-of-the-art predictors in generalization and achieves significant performance gains in finding the optimal architectures with test error 2.40% on CIFAR-10 and 23.20% on ImageNet1k within 0.01 GPU days.

Real-time video analytics on edge devices has gained increasing attention across a wide range of business areas. However, edge devices usually have limited computing resources. Consequently, conventional approaches to video analytics either deploy simplified models on the edge (resulting in low accuracy) or transmit video content to the cloud (resulting in high latency and network overheads) to enable deep learning inference (e.g., object detection). In this paper, we introduce ParaLoupe, a novel real-time video analytics system that parallelizes deep learning inference in the edge cluster with task-oriented mini models. These mini models do not attain State-of-the-Art accuracy individually, but collectively can achieve much better accuracy-latency tradeoff than State-of-the-Art models. To achieve this, ParaLoupe crops multiple single-object patches from a given video frame. These single-object patches are then sent to multiple edge devices for parallel inference with specifically designed mini models. A patch-based task scheduling algorithm is further proposed to leverage the computing resources of the edge cluster to meet the service-level objectives. Our experimental results on real-world datasets show that ParaLoupe significantly outperforms baseline methods, achieving up to 14.1× inference speedup with accuracy on par with state-of-the-art models, or improving accuracy up to 45.1% under the same latency constraints.