Wei Wen

Zifei Wang, Wei Wen, Qiang Ji et al.

LLM agents complete complex tasks by composing multiple skills, and skill retrieval is a front-end stage for agents. Skill retrieval differs fundamentally from traditional document retrieval at the supervision level: top-K joint correctness depends not only on the semantic relevance of each individual query-skill pair, but also on whether the skills retrieved together can collaborate to fulfill the task under the given query. Such "skill compatibility" cannot be derived from independent relevance alone. Yet existing LLM-based data synthesis pipelines can produce a direct supervision signal for "which skills should not be jointly retrieved under this query" -- namely the LLM's own rejection decisions -- and this signal is routinely discarded as low-quality data. To address this gap, we propose Reject-as-Resource Retriever (R3) and construct R3-Skill, a bilingual (Chinese-English) skill retrieval benchmark targeting realistic agent skill routing. R3-Skill spans four language directions, features query phrasings close to real user requests, and is verified through multi-expert cross-checking. On R3-Skill, we build a two-stage retrieval system (R3-Embedding + R3-Reranker) with skill compatibility as an explicit training signal. Gradient analysis shows that the "push-away" signal is diluted by bilateral balancing in the bi-encoder but acts as lossless graded ranking supervision in the cross-encoder -- motivating its placement at the cross-encoder stage, as confirmed by ablations on two datasets. The R3-Embedding + R3-Reranker pipeline attains Hit@1 = 0.7714, NDCG@10 = 0.8327 and Set-Compat = 0.3525 on R3-Skill. The dataset, training code and model weights are released as open source for agent skill routing.

Xiujun Shu, Wei Wen, Haoqian Wu et al.

Text-based person retrieval aims to find the query person based on a textual description. The key is to learn a common latent space mapping between visual-textual modalities. To achieve this goal, existing works employ segmentation to obtain explicitly cross-modal alignments or utilize attention to explore salient alignments. These methods have two shortcomings: 1) Labeling cross-modal alignments are time-consuming. 2) Attention methods can explore salient cross-modal alignments but may ignore some subtle and valuable pairs. To relieve these issues, we introduce an Implicit Visual-Textual (IVT) framework for text-based person retrieval. Different from previous models, IVT utilizes a single network to learn representation for both modalities, which contributes to the visual-textual interaction. To explore the fine-grained alignment, we further propose two implicit semantic alignment paradigms: multi-level alignment (MLA) and bidirectional mask modeling (BMM). The MLA module explores finer matching at sentence, phrase, and word levels, while the BMM module aims to mine \textbf{more} semantic alignments between visual and textual modalities. Extensive experiments are carried out to evaluate the proposed IVT on public datasets, i.e., CUHK-PEDES, RSTPReID, and ICFG-PEDES. Even without explicit body part alignment, our approach still achieves state-of-the-art performance. Code is available at: https://github.com/TencentYoutuResearch/PersonRetrieval-IVT.

Tunhou Zhang, Dehua Cheng, Yuchen He et al.

The rise of deep neural networks offers new opportunities in optimizing recommender systems. However, optimizing recommender systems using deep neural networks requires delicate architecture fabrication. We propose NASRec, a paradigm that trains a single supernet and efficiently produces abundant models/sub-architectures by weight sharing. To overcome the data multi-modality and architecture heterogeneity challenges in the recommendation domain, NASRec establishes a large supernet (i.e., search space) to search the full architectures. The supernet incorporates versatile choice of operators and dense connectivity to minimize human efforts for finding priors. The scale and heterogeneity in NASRec impose several challenges, such as training inefficiency, operator-imbalance, and degraded rank correlation. We tackle these challenges by proposing single-operator any-connection sampling, operator-balancing interaction modules, and post-training fine-tuning. Our crafted models, NASRecNet, show promising results on three Click-Through Rates (CTR) prediction benchmarks, indicating that NASRec outperforms both manually designed models and existing NAS methods with state-of-the-art performance. Our work is publicly available at https://github.com/facebookresearch/NasRec.

Hanjun Li, Xiujun Shu, Sunan He et al.

Temporal sentence grounding (TSG) aims to locate a specific moment from an untrimmed video with a given natural language query. Recently, weakly supervised methods still have a large performance gap compared to fully supervised ones, while the latter requires laborious timestamp annotations. In this study, we aim to reduce the annotation cost yet keep competitive performance for TSG task compared to fully supervised ones. To achieve this goal, we investigate a recently proposed glance-supervised temporal sentence grounding task, which requires only single frame annotation (referred to as glance annotation) for each query. Under this setup, we propose a Dynamic Gaussian prior based Grounding framework with Glance annotation (D3G), which consists of a Semantic Alignment Group Contrastive Learning module (SA-GCL) and a Dynamic Gaussian prior Adjustment module (DGA). Specifically, SA-GCL samples reliable positive moments from a 2D temporal map via jointly leveraging Gaussian prior and semantic consistency, which contributes to aligning the positive sentence-moment pairs in the joint embedding space. Moreover, to alleviate the annotation bias resulting from glance annotation and model complex queries consisting of multiple events, we propose the DGA module, which adjusts the distribution dynamically to approximate the ground truth of target moments. Extensive experiments on three challenging benchmarks verify the effectiveness of the proposed D3G. It outperforms the state-of-the-art weakly supervised methods by a large margin and narrows the performance gap compared to fully supervised methods. Code is available at https://github.com/solicucu/D3G.

Jiyang Jiang, Dadong Wang, Yang Song et al.

Cerebral small vessel disease (CSVD) is a major vascular contributor to cognitive impairment in ageing, including dementias. Imaging remains the most promising method for in vivo studies of CSVD. To replace the subjective and laborious visual rating approaches, emerging studies have applied state-of-the-art artificial intelligence to extract imaging biomarkers of CSVD from MRI scans. We aimed to summarise published computer-aided methods to examine three imaging biomarkers of CSVD, namely cerebral microbleeds (CMB), dilated perivascular spaces (PVS), and lacunes of presumed vascular origin. Seventy-one classical image processing, classical machine learning, and deep learning studies were identified. CMB and PVS have been better studied, compared to lacunes. While good performance metrics have been achieved in local test datasets, there have not been generalisable pipelines validated in different research or clinical cohorts. Transfer learning and weak supervision techniques have been applied to accommodate the limitations in training data. Future studies could consider pooling data from multiple sources to increase diversity, and validating the performance of the methods using both image processing metrics and associations with clinical measures.

Junru Lu, Jiarui Qin, Lingfeng Qiao et al.

We introduce Youtu-LLM, a lightweight yet powerful language model that harmonizes high computational efficiency with native agentic intelligence. Unlike typical small models that rely on distillation, Youtu-LLM (1.96B) is pre-trained from scratch to systematically cultivate reasoning and planning capabilities. The key technical advancements are as follows: (1) Compact Architecture with Long-Context Support: Built on a dense Multi-Latent Attention (MLA) architecture with a novel STEM-oriented vocabulary, Youtu-LLM supports a 128k context window. This design enables robust long-context reasoning and state tracking within a minimal memory footprint, making it ideal for long-horizon agent and reasoning tasks. (2) Principled "Commonsense-STEM-Agent" Curriculum: We curated a massive corpus of approximately 11T tokens and implemented a multi-stage training strategy. By progressively shifting the pre-training data distribution from general commonsense to complex STEM and agentic tasks, we ensure the model acquires deep cognitive abilities rather than superficial alignment. (3) Scalable Agentic Mid-training: Specifically for the agentic mid-training, we employ diverse data construction schemes to synthesize rich and varied trajectories across math, coding, and tool-use domains. This high-quality data enables the model to internalize planning and reflection behaviors effectively. Extensive evaluations show that Youtu-LLM sets a new state-of-the-art for sub-2B LLMs. On general benchmarks, it achieves competitive performance against larger models, while on agent-specific tasks, it significantly surpasses existing SOTA baselines, demonstrating that lightweight models can possess strong intrinsic agentic capabilities.

Chenchen Zhu, Saksham Suri, Cijo Jose et al.

Running AI models on smart edge devices can unlock versatile user experiences, but presents challenges due to limited compute and the need to handle multiple tasks simultaneously. This requires a vision encoder with small size but powerful and versatile representations. We present our method, Efficient Universal Perception Encoder (EUPE), which offers both inference efficiency and universally good representations for diverse downstream tasks. We achieve this by distilling from multiple domain-expert foundation vision encoders. Unlike previous agglomerative methods that directly scale down from multiple teachers to an efficient encoder, we demonstrate the importance of first scaling up to a large proxy teacher and then scaling down from this single teacher. Experiments show that EUPE achieves on-par or better performance than individual domain experts of the same size on diverse task domains and also outperforms previous agglomerative encoders. We will release the full family of EUPE models and the code to foster future research.

Tunhou Zhang, Wei Wen, Igor Fedorov et al.

Search efficiency and serving efficiency are two major axes in building feature interactions and expediting the model development process in recommender systems. On large-scale benchmarks, searching for the optimal feature interaction design requires extensive cost due to the sequential workflow on the large volume of data. In addition, fusing interactions of various sources, orders, and mathematical operations introduces potential conflicts and additional redundancy toward recommender models, leading to sub-optimal trade-offs in performance and serving cost. In this paper, we present DistDNAS as a neat solution to brew swift and efficient feature interaction design. DistDNAS proposes a supernet to incorporate interaction modules of varying orders and types as a search space. To optimize search efficiency, DistDNAS distributes the search and aggregates the choice of optimal interaction modules on varying data dates, achieving over 25x speed-up and reducing search cost from 2 days to 2 hours. To optimize serving efficiency, DistDNAS introduces a differentiable cost-aware loss to penalize the selection of redundant interaction modules, enhancing the efficiency of discovered feature interactions in serving. We extensively evaluate the best models crafted by DistDNAS on a 1TB Criteo Terabyte dataset. Experimental evaluations demonstrate 0.001 AUC improvement and 60% FLOPs saving over current state-of-the-art CTR models.

Xiujun Shu, Wei Wen, Liangsheng Xu et al.

Video temporal character grouping locates appearing moments of major characters within a video according to their identities. To this end, recent works have evolved from unsupervised clustering to graph-based supervised clustering. However, graph methods are built upon the premise of fixed affinity graphs, bringing many inexact connections. Besides, they extract multi-modal features with kinds of models, which are unfriendly to deployment. In this paper, we present a unified and dynamic graph (UniDG) framework for temporal character grouping. This is accomplished firstly by a unified representation network that learns representations of multiple modalities within the same space and still preserves the modality's uniqueness simultaneously. Secondly, we present a dynamic graph clustering where the neighbors of different quantities are dynamically constructed for each node via a cyclic matching strategy, leading to a more reliable affinity graph. Thirdly, a progressive association method is introduced to exploit spatial and temporal contexts among different modalities, allowing multi-modal clustering results to be well fused. As current datasets only provide pre-extracted features, we evaluate our UniDG method on a collected dataset named MTCG, which contains each character's appearing clips of face and body and speaking voice tracks. We also evaluate our key components on existing clustering and retrieval datasets to verify the generalization ability. Experimental results manifest that our method can achieve promising results and outperform several state-of-the-art approaches.

Yufan Cao, Tunhou Zhang, Wei Wen et al.

Weight-sharing Neural Architecture Search (WS-NAS) provides an efficient mechanism for developing end-to-end deep recommender models. However, in complex search spaces, distinguishing between superior and inferior architectures (or paths) is challenging. This challenge is compounded by the limited coverage of the supernet and the co-adaptation of subnet weights, which restricts the exploration and exploitation capabilities inherent to weight-sharing mechanisms. To address these challenges, we introduce Farthest Greedy Path Sampling (FGPS), a new path sampling strategy that balances path quality and diversity. FGPS enhances path diversity to facilitate more comprehensive supernet exploration, while emphasizing path quality to ensure the effective identification and utilization of promising architectures. By incorporating FGPS into a Two-shot NAS (TS-NAS) framework, we derive high-performance architectures. Evaluations on three Click-Through Rate (CTR) prediction benchmarks demonstrate that our approach consistently achieves superior results, outperforming both manually designed and most NAS-based models.

Wenxuan Zhang, Lemeng Wu, Changsheng Zhao et al.

Diffusion Large Language Models (dLLMs) introduce a new paradigm for language generation, which in turn presents new challenges for aligning them with human preferences. In this work, we aim to improve the policy optimization for dLLMs by reducing the cost of the trajectory probability calculation, thereby enabling scaled-up offline policy training. We prove that: (i) under reference policy regularization, the probability ratio of the newly unmasked tokens is an unbiased estimate of that of intermediate diffusion states, and (ii) the probability of the full trajectory can be effectively estimated with a single forward pass of a re-masked final state. By integrating these two trajectory reduction strategies into a policy optimization objective, we propose Trajectory Reduction Policy Optimization (dTRPO). We evaluate dTRPO on 7B dLLMs across instruction-following and reasoning benchmarks. Results show that it substantially improves the core performance of state-of-the-art dLLMs, achieving gains of up to 9.6% on STEM tasks, up to 4.3% on coding tasks, and up to 3.0% on instruction-following tasks. Moreover, dTRPO exhibits strong training efficiency due to its offline, single-forward nature, and achieves improved generation efficiency through high-quality outputs.

Wei Wen, Kuang-Hung Liu, Igor Fedorov et al.

Neural Architecture Search (NAS) has demonstrated its efficacy in computer vision and potential for ranking systems. However, prior work focused on academic problems, which are evaluated at small scale under well-controlled fixed baselines. In industry system, such as ranking system in Meta, it is unclear whether NAS algorithms from the literature can outperform production baselines because of: (1) scale - Meta ranking systems serve billions of users, (2) strong baselines - the baselines are production models optimized by hundreds to thousands of world-class engineers for years since the rise of deep learning, (3) dynamic baselines - engineers may have established new and stronger baselines during NAS search, and (4) efficiency - the search pipeline must yield results quickly in alignment with the productionization life cycle. In this paper, we present Rankitect, a NAS software framework for ranking systems at Meta. Rankitect seeks to build brand new architectures by composing low level building blocks from scratch. Rankitect implements and improves state-of-the-art (SOTA) NAS methods for comprehensive and fair comparison under the same search space, including sampling-based NAS, one-shot NAS, and Differentiable NAS (DNAS). We evaluate Rankitect by comparing to multiple production ranking models at Meta. We find that Rankitect can discover new models from scratch achieving competitive tradeoff between Normalized Entropy loss and FLOPs. When utilizing search space designed by engineers, Rankitect can generate better models than engineers, achieving positive offline evaluation and online A/B test at Meta scale.

Xiujun Shu, Wei Wen, Taian Guo et al.

This technical report presents the 3rd winning solution for MTVG, a new task introduced in the 4-th Person in Context (PIC) Challenge at ACM MM 2022. MTVG aims at localizing the temporal boundary of the step in an untrimmed video based on a textual description. The biggest challenge of this task is the fi ne-grained video-text semantics of make-up steps. However, current methods mainly extract video features using action-based pre-trained models. As actions are more coarse-grained than make-up steps, action-based features are not sufficient to provide fi ne-grained cues. To address this issue,we propose to achieve fi ne-grained representation via exploiting feature diversities. Specifically, we proposed a series of methods from feature extraction, network optimization, to model ensemble. As a result, we achieved 3rd place in the MTVG competition.

Shuming Liu, Mingchen Zhuge, Changsheng Zhao et al.

Chain-of-thought (CoT) reasoning has emerged as a powerful tool for multimodal large language models on video understanding tasks. However, its necessity and advantages over direct answering remain underexplored. In this paper, we first demonstrate that for RL-trained video models, direct answering often matches or even surpasses CoT performance, despite CoT producing step-by-step analyses at a higher computational cost. Motivated by this, we propose VideoAuto-R1, a video understanding framework that adopts a reason-when-necessary strategy. During training, our approach follows a Thinking Once, Answering Twice paradigm: the model first generates an initial answer, then performs reasoning, and finally outputs a reviewed answer. Both answers are supervised via verifiable rewards. During inference, the model uses the confidence score of the initial answer to determine whether to proceed with reasoning. Across video QA and grounding benchmarks, VideoAuto-R1 achieves state-of-the-art accuracy with significantly improved efficiency, reducing the average response length by ~3.3x, e.g., from 149 to just 44 tokens. Moreover, we observe a low rate of thinking-mode activation on perception-oriented tasks, but a higher rate on reasoning-intensive tasks. This suggests that explicit language-based reasoning is generally beneficial but not always necessary.

Patrick Huber, Ernie Chang, Wei Wen et al.

Efficient on-device language models around 1 billion parameters are essential for powering low-latency AI applications on mobile and wearable devices. However, achieving strong performance in this model class, while supporting long context windows and practical deployment remains a significant challenge. We introduce MobileLLM-Pro, a 1-billion-parameter language model optimized for on-device deployment. MobileLLM-Pro achieves state-of-the-art results across 11 standard benchmarks, significantly outperforming both Gemma 3-1B and Llama 3.2-1B, while supporting context windows of up to 128,000 tokens and showing only minor performance regressions at 4-bit quantization. These improvements are enabled by four core innovations: (1) implicit positional distillation, a novel technique that effectively instills long-context capabilities through knowledge distillation; (2) a specialist model merging framework that fuses multiple domain experts into a compact model without parameter growth; (3) simulation-driven data mixing using utility estimation; and (4) 4-bit quantization-aware training with self-distillation. We release our model weights and code to support future research in efficient on-device language models.

Tianjun Wei, Wei Wen, Ruizhi Qiao et al.

Evaluating large language models (LLMs) in diverse and challenging scenarios is essential to align them with human preferences. To mitigate the prohibitive costs associated with human evaluations, utilizing a powerful LLM as a judge has emerged as a favored approach. Nevertheless, this methodology encounters several challenges, including substantial expenses, concerns regarding privacy and security, and reproducibility. In this paper, we propose a straightforward, replicable, and accurate automated evaluation method by leveraging a lightweight LLM as the judge, named RocketEval. Initially, we identify that the performance disparity between lightweight and powerful LLMs in evaluation tasks primarily stems from their ability to conduct comprehensive analyses, which is not easily enhanced through techniques such as chain-of-thought reasoning. By reframing the evaluation task as a multi-faceted Q&A using an instance-specific checklist, we demonstrate that the limited judgment accuracy of lightweight LLMs is largely attributes to high uncertainty and positional bias. To address these challenges, we introduce an automated evaluation process grounded in checklist grading, which is designed to accommodate a variety of scenarios and questions. This process encompasses the creation of checklists, the grading of these checklists by lightweight LLMs, and the reweighting of checklist items to align with the supervised annotations. Our experiments carried out on the automated evaluation benchmarks, MT-Bench and WildBench datasets, reveal that RocketEval, when using Gemma-2-2B as the judge, achieves a high correlation (0.965) with human preferences, which is comparable to GPT-4o. Moreover, RocketEval provides a cost reduction exceeding 50-fold for large-scale evaluation and comparison scenarios. Our code is available at https://github.com/Joinn99/RocketEval-ICLR .

Hang Yin, Kuang-Hung Liu, Mengying Sun et al.

Web-scale ranking systems at Meta serving billions of users is complex. Improving ranking models is essential but engineering heavy. Automated Machine Learning (AutoML) can release engineers from labor intensive work of tuning ranking models; however, it is unknown if AutoML is efficient enough to meet tight production timeline in real-world and, at the same time, bring additional improvements to the strong baselines. Moreover, to achieve higher ranking performance, there is an ever-increasing demand to scale up ranking models to even larger capacity, which imposes more challenges on the efficiency. The large scale of models and tight production schedule requires AutoML to outperform human baselines by only using a small number of model evaluation trials (around 100). We presents a sampling-based AutoML method, focusing on neural architecture search and hyperparameter optimization, addressing these challenges in Meta-scale production when building large capacity models. Our approach efficiently handles large-scale data demands. It leverages a lightweight predictor-based searcher and reinforcement learning to explore vast search spaces, significantly reducing the number of model evaluations. Through experiments in large capacity modeling for CTR and CVR applications, we show that our method achieves outstanding Return on Investment (ROI) versus human tuned baselines, with up to 0.09% Normalized Entropy (NE) loss reduction or $25\%$ Query per Second (QPS) increase by only sampling one hundred models on average from a curated search space. The proposed AutoML method has already made real-world impact where a discovered Instagram CTR model with up to -0.36% NE gain (over existing production baseline) was selected for large-scale online A/B test and show statistically significant gain. These production results proved AutoML efficacy and accelerated its adoption in ranking systems at Meta.

Tao Wang, Wei Wen, Jingzhi Zhai et al.

Point cloud segmentation is crucial for robotic visual perception and environmental understanding, enabling applications such as robotic navigation and 3D reconstruction. However, handling the sparse and unordered nature of point cloud data presents challenges for efficient and accurate segmentation. Inspired by the Mamba model's success in natural language processing, we propose the Serialized Point Cloud Mamba Segmentation Model (Serialized Point Mamba), which leverages a state-space model to dynamically compress sequences, reduce memory usage, and enhance computational efficiency. Serialized Point Mamba integrates local-global modeling capabilities with linear complexity, achieving state-of-the-art performance on both indoor and outdoor datasets. This approach includes novel techniques such as staged point cloud sequence learning, grid pooling, and Conditional Positional Encoding, facilitating effective segmentation across diverse point cloud tasks. Our method achieved 76.8 mIoU on Scannet and 70.3 mIoU on S3DIS. In Scannetv2 instance segmentation, it recorded 40.0 mAP. It also had the lowest latency and reasonable memory use, making it the SOTA among point semantic segmentation models based on mamba.

Qinsi Wang, Hancheng Ye, Jinhee Kim et al.

Think about how human handles complex reading tasks: marking key points, inferring their relationships, and structuring information to guide understanding and responses. Likewise, can a large language model benefit from text structure to enhance text-processing performance? To explore it, in this work, we first introduce Structure of Thought (SoT), a prompting technique that explicitly guides models to construct intermediate text structures, consistently boosting performance across eight tasks and three model families. Building upon this insight, we present T2S-Bench, the first benchmark designed to evaluate and improve text-to-structure capabilities of models. T2S-Bench includes 1.8K samples across 6 scientific domains and 32 structural types, rigorously constructed to ensure accuracy, fairness, and quality. Evaluation on 45 mainstream models reveals substantial improvement potential: the average accuracy on the multi-hop reasoning task is only 52.1%, and even the most advanced model achieves 58.1% node accuracy in end-to-end extraction. Furthermore, on Qwen2.5-7B-Instruct, SoT alone yields an average +5.7% improvement across eight diverse text-processing tasks, and fine-tuning on T2S-Bench further increases this gain to +8.6%. These results highlight the value of explicit text structuring and the complementary contributions of SoT and T2S-Bench. Dataset and eval code have been released at https://t2s-bench.github.io/T2S-Bench-Page/.

Changsheng Zhao, Ernie Chang, Zechun Liu et al.

The paradigm shift in large language models (LLMs) from instinctive responses to chain-of-thought (CoT) reasoning has fueled two prevailing assumptions: (1) reasoning capabilities only emerge in sufficiently large models, and (2) such capabilities require training on massive datasets. While the first assumption has already been challenged by recent sub-billion-parameter reasoning models such as Qwen3-0.6B and DeepSeek distilled variants, the second remains largely unquestioned. In this work, we revisit the necessity of scaling to extremely large corpora (>10T tokens) for reasoning emergence. By carefully curating and resampling open-source datasets that we identify as beneficial under our designed metrics, we demonstrate that strong reasoning abilities can emerge with far less data. Specifically, we show that only ~2T tokens of high-quality data are sufficient, and pre-training with 4.2T tokens on the dataset resampled from these ~2T tokens, followed by a established post-training procedure, enables the development of MobileLLM-R1, a series of sub-billion-parameter reasoning models that substantially outperform prior models trained on fully open-sourced data. For example, MobileLLM-R1-950M achieves an AIME score of 15.5, compared to just 0.6 for OLMo-2-1.48B and 0.3 for SmolLM-2-1.7B. Remarkably, despite being trained on only 11.7% of the tokens compared to Qwen3's proprietary 36T-token corpus for pretraining, MobileLLM-R1-950M matches or surpasses Qwen3-0.6B across multiple reasoning benchmarks. To facilitate further research in this direction, we have released the complete training recipe, data sources, data mixing ratio, and model checkpoints, together with the key insights obtained throughout this study.

Hua Zheng, Kuang-Hung Liu, Igor Fedorov et al.

Neural Architecture Search (NAS) has become a widely used tool for automating neural network design. While one-shot NAS methods have successfully reduced computational requirements, they often require extensive training. On the other hand, zero-shot NAS utilizes training-free proxies to evaluate a candidate architecture's test performance but has two limitations: (1) inability to use the information gained as a network improves with training and (2) unreliable performance, particularly in complex domains like RecSys, due to the multi-modal data inputs and complex architecture configurations. To synthesize the benefits of both methods, we introduce a "sub-one-shot" paradigm that serves as a bridge between zero-shot and one-shot NAS. In sub-one-shot NAS, the supernet is trained using only a small subset of the training data, a phase we refer to as "warm-up." Within this framework, we present SiGeo, a proxy founded on a novel theoretical framework that connects the supernet warm-up with the efficacy of the proxy. Extensive experiments have shown that SiGeo, with the benefit of warm-up, consistently outperforms state-of-the-art NAS proxies on various established NAS benchmarks. When a supernet is warmed up, it can achieve comparable performance to weight-sharing one-shot NAS methods, but with a significant reduction ($\sim 60$\%) in computational costs.

Wei Wen, Quanyu Zhu, Weiwei Chu et al.

Scaling up deep learning models has been proven effective to improve intelligence of machine learning (ML) models, especially for industry recommendation models and large language models. The co-design of large distributed ML systems and algorithms (to maximize training performance) plays a pivotal role for its success. As it scales, the number of co-design hyper-parameters grows rapidly which brings challenges to feasibly find the optimal setup for system performance maximization. In this paper, we propose CubicML which uses ML to automatically optimize training performance of large distributed ML systems. In CubicML, we use an ML model as a proxy to predict the training performance for search efficiency and performance modeling flexibility. We proved that CubicML can effectively optimize training speed of in-house ads recommendation models with 73 billion parameters and large language models up to 405 billion parameters at Meta.

Xiujun Shu, Yusheng Tao, Ruizhi Qiao et al.

Person search in media has seen increasing potential in Internet applications, such as video clipping and character collection. This task is common but overlooked by previous person search works which focus on surveillance scenes. The media scenarios have some different challenges from surveillance scenes. For example, a person may change his clothes frequently. To alleviate this issue, this paper proposes a Unified Detector and Graph Network (UDGNet) for person search in media. UDGNet is the first person search framework to detect and re-identify the human body and head simultaneously. Specifically, it first builds two branches based on a unified network to detect the human body and head, then the detected body and head are used for re-identification. This dual-task approach can significantly enhance discriminative learning. To tackle the cloth-changing issue, UDGNet builds two graphs to explore reliable links among cloth-changing samples and utilizes a graph network to learn better embeddings. This design effectively enhances the robustness of person search to cloth-changing challenges. Besides, we demonstrate that UDGNet can be implemented with both anchor-based and anchor-free person search frameworks and further achieve performance improvement. This paper also contributes a large-scale dataset for Person Search in Media (PSM), which provides both body and head annotations. It is by far the largest dataset for person search in media. Experiments show that UDGNet improves the anchor-free model AlignPS by 12.1% in mAP. Meanwhile, it shows good generalization across surveillance and longterm scenarios. The dataset and code will be available at: https://github.com/shuxjweb/PSM.git.

Jian Cheng, Ziyang Liu, Hao Guan et al.

Chronological age of healthy people is able to be predicted accurately using deep neural networks from neuroimaging data, and the predicted brain age could serve as a biomarker for detecting aging-related diseases. In this paper, a novel 3D convolutional network, called two-stage-age-network (TSAN), is proposed to estimate brain age from T1-weighted MRI data. Compared with existing methods, TSAN has the following improvements. First, TSAN uses a two-stage cascade network architecture, where the first-stage network estimates a rough brain age, then the second-stage network estimates the brain age more accurately from the discretized brain age by the first-stage network. Second, to our knowledge, TSAN is the first work to apply novel ranking losses in brain age estimation, together with the traditional mean square error (MSE) loss. Third, densely connected paths are used to combine feature maps with different scales. The experiments with $6586$ MRIs showed that TSAN could provide accurate brain age estimation, yielding mean absolute error (MAE) of $2.428$ and Pearson's correlation coefficient (PCC) of $0.985$, between the estimated and chronological ages. Furthermore, using the brain age gap between brain age and chronological age as a biomarker, Alzheimer's disease (AD) and Mild Cognitive Impairment (MCI) can be distinguished from healthy control (HC) subjects by support vector machine (SVM). Classification AUC in AD/HC and MCI/HC was $0.904$ and $0.823$, respectively. It showed that brain age gap is an effective biomarker associated with risk of dementia, and has potential for early-stage dementia risk screening. The codes and trained models have been released on GitHub: https://github.com/Milan-BUAA/TSAN-brain-age-estimation.

Yuhui Xu, Yuxi Li, Shuai Zhang et al.

To accelerate DNNs inference, low-rank approximation has been widely adopted because of its solid theoretical rationale and efficient implementations. Several previous works attempted to directly approximate a pre-trained model by low-rank decomposition; however, small approximation errors in parameters can ripple over a large prediction loss. Apparently, it is not optimal to separate low-rank approximation from training. Unlike previous works, this paper integrates low rank approximation and regularization into the training process. We propose Trained Rank Pruning (TRP), which alternates between low rank approximation and training. TRP maintains the capacity of the original network while imposing low-rank constraints during training. A nuclear regularization optimized by stochastic sub-gradient descent is utilized to further promote low rank in TRP. Networks trained with TRP has a low-rank structure in nature, and is approximated with negligible performance loss, thus eliminating fine-tuning after low rank approximation. The proposed method is comprehensively evaluated on CIFAR-10 and ImageNet, outperforming previous compression counterparts using low rank approximation. Our code is available at: https://github.com/yuhuixu1993/Trained-Rank-Pruning.

Wei Wen, Feng Yan, Yiran Chen et al.

Depth is a key component of Deep Neural Networks (DNNs), however, designing depth is heuristic and requires many human efforts. We propose AutoGrow to automate depth discovery in DNNs: starting from a shallow seed architecture, AutoGrow grows new layers if the growth improves the accuracy; otherwise, stops growing and thus discovers the depth. We propose robust growing and stopping policies to generalize to different network architectures and datasets. Our experiments show that by applying the same policy to different network architectures, AutoGrow can always discover near-optimal depth on various datasets of MNIST, FashionMNIST, SVHN, CIFAR10, CIFAR100 and ImageNet. For example, in terms of accuracy-computation trade-off, AutoGrow discovers a better depth combination in ResNets than human experts. Our AutoGrow is efficient. It discovers depth within similar time of training a single DNN. Our code is available at https://github.com/wenwei202/autogrow.

Yuhui Xu, Yuxi Li, Shuai Zhang et al.

The performance of Deep Neural Networks (DNNs) keeps elevating in recent years with increasing network depth and width. To enable DNNs on edge devices like mobile phones, researchers proposed several network compression methods including pruning, quantization and factorization. Among the factorization-based approaches, low-rank approximation has been widely adopted because of its solid theoretical rationale and efficient implementations. Several previous works attempted to directly approximate a pre-trained model by low-rank decomposition; however, small approximation errors in parameters can ripple a large prediction loss. As a result, performance usually drops significantly and a sophisticated fine-tuning is required to recover accuracy. We argue that it is not optimal to separate low-rank approximation from training. Unlike previous works, this paper integrates low rank approximation and regularization into the training. We propose Trained Rank Pruning (TRP), which iterates low rank approximation and training. TRP maintains the capacity of original network while imposes low-rank constraints during training. A stochastic sub-gradient descent optimized nuclear regularization is utilized to further encourage low rank in TRP. The TRP trained network has low-rank structure in nature, and can be approximated with negligible performance loss, eliminating fine-tuning after low rank approximation. The methods are comprehensively evaluated on CIFAR-10 and ImageNet, outperforming previous compression methods using low rank approximation. Code is available: https://github.com/yuhuixu1993/Trained-Rank-Pruning

Tingting Yu, Wei Wen, Xue Han et al.

Concurrent programs are difficult to test due to their inherent non-determinism. To address this problem, testing often requires the exploration of thread schedules of a program; this can be time-consuming when applied to real-world programs. Software defect prediction has been used to help developers find faults and prioritize their testing efforts. Prior studies have used machine learning to build such predicting models based on designed features that encode the characteristics of programs. However, research has focused on sequential programs; to date, no work has considered defect prediction for concurrent programs, with program characteristics distinguished from sequential programs. In this paper, we present ConPredictor, an approach to predict defects specific to concurrent programs by combining both static and dynamic program metrics. Specifically, we propose a set of novel static code metrics based on the unique properties of concurrent programs. We also leverage additional guidance from dynamic metrics constructed based on mutation analysis. Our evaluation on four large open source projects shows that ConPredictor improved both within-project defect prediction and cross-project defect prediction compared to traditional features.

Wei Wen, Yuxiong He, Samyam Rajbhandari et al.

Model compression is significant for the wide adoption of Recurrent Neural Networks (RNNs) in both user devices possessing limited resources and business clusters requiring quick responses to large-scale service requests. This work aims to learn structurally-sparse Long Short-Term Memory (LSTM) by reducing the sizes of basic structures within LSTM units, including input updates, gates, hidden states, cell states and outputs. Independently reducing the sizes of basic structures can result in inconsistent dimensions among them, and consequently, end up with invalid LSTM units. To overcome the problem, we propose Intrinsic Sparse Structures (ISS) in LSTMs. Removing a component of ISS will simultaneously decrease the sizes of all basic structures by one and thereby always maintain the dimension consistency. By learning ISS within LSTM units, the obtained LSTMs remain regular while having much smaller basic structures. Based on group Lasso regularization, our method achieves 10.59x speedup without losing any perplexity of a language modeling of Penn TreeBank dataset. It is also successfully evaluated through a compact model with only 2.69M weights for machine Question Answering of SQuAD dataset. Our approach is successfully extended to non- LSTM RNNs, like Recurrent Highway Networks (RHNs). Our source code is publicly available at https://github.com/wenwei202/iss-rnns

Wei Wen, Cong Xu, Feng Yan et al.

High network communication cost for synchronizing gradients and parameters is the well-known bottleneck of distributed training. In this work, we propose TernGrad that uses ternary gradients to accelerate distributed deep learning in data parallelism. Our approach requires only three numerical levels {-1,0,1}, which can aggressively reduce the communication time. We mathematically prove the convergence of TernGrad under the assumption of a bound on gradients. Guided by the bound, we propose layer-wise ternarizing and gradient clipping to improve its convergence. Our experiments show that applying TernGrad on AlexNet does not incur any accuracy loss and can even improve accuracy. The accuracy loss of GoogLeNet induced by TernGrad is less than 2% on average. Finally, a performance model is proposed to study the scalability of TernGrad. Experiments show significant speed gains for various deep neural networks. Our source code is available.

Wei Wen, Cong Xu, Chunpeng Wu et al.

Very large-scale Deep Neural Networks (DNNs) have achieved remarkable successes in a large variety of computer vision tasks. However, the high computation intensity of DNNs makes it challenging to deploy these models on resource-limited systems. Some studies used low-rank approaches that approximate the filters by low-rank basis to accelerate the testing. Those works directly decomposed the pre-trained DNNs by Low-Rank Approximations (LRA). How to train DNNs toward lower-rank space for more efficient DNNs, however, remains as an open area. To solve the issue, in this work, we propose Force Regularization, which uses attractive forces to enforce filters so as to coordinate more weight information into lower-rank space. We mathematically and empirically verify that after applying our technique, standard LRA methods can reconstruct filters using much lower basis and thus result in faster DNNs. The effectiveness of our approach is comprehensively evaluated in ResNets, AlexNet, and GoogLeNet. In AlexNet, for example, Force Regularization gains 2x speedup on modern GPU without accuracy loss and 4.05x speedup on CPU by paying small accuracy degradation. Moreover, Force Regularization better initializes the low-rank DNNs such that the fine-tuning can converge faster toward higher accuracy. The obtained lower-rank DNNs can be further sparsified, proving that Force Regularization can be integrated with state-of-the-art sparsity-based acceleration methods. Source code is available in https://github.com/wenwei202/caffe

Wei Wen, Chunpeng Wu, Yandan Wang et al.

High demand for computation resources severely hinders deployment of large-scale Deep Neural Networks (DNN) in resource constrained devices. In this work, we propose a Structured Sparsity Learning (SSL) method to regularize the structures (i.e., filters, channels, filter shapes, and layer depth) of DNNs. SSL can: (1) learn a compact structure from a bigger DNN to reduce computation cost; (2) obtain a hardware-friendly structured sparsity of DNN to efficiently accelerate the DNNs evaluation. Experimental results show that SSL achieves on average 5.1x and 3.1x speedups of convolutional layer computation of AlexNet against CPU and GPU, respectively, with off-the-shelf libraries. These speedups are about twice speedups of non-structured sparsity; (3) regularize the DNN structure to improve classification accuracy. The results show that for CIFAR-10, regularization on layer depth can reduce 20 layers of a Deep Residual Network (ResNet) to 18 layers while improve the accuracy from 91.25% to 92.60%, which is still slightly higher than that of original ResNet with 32 layers. For AlexNet, structure regularization by SSL also reduces the error by around ~1%. Open source code is in https://github.com/wenwei202/caffe/tree/scnn

Jongsoo Park, Sheng Li, Wei Wen et al.

Phenomenally successful in practical inference problems, convolutional neural networks (CNN) are widely deployed in mobile devices, data centers, and even supercomputers. The number of parameters needed in CNNs, however, are often large and undesirable. Consequently, various methods have been developed to prune a CNN once it is trained. Nevertheless, the resulting CNNs offer limited benefits. While pruning the fully connected layers reduces a CNN's size considerably, it does not improve inference speed noticeably as the compute heavy parts lie in convolutions. Pruning CNNs in a way that increase inference speed often imposes specific sparsity structures, thus limiting the achievable sparsity levels. We present a method to realize simultaneously size economy and speed improvement while pruning CNNs. Paramount to our success is an efficient general sparse-with-dense matrix multiplication implementation that is applicable to convolution of feature maps with kernels of arbitrary sparsity patterns. Complementing this, we developed a performance model that predicts sweet spots of sparsity levels for different layers and on different computer architectures. Together, these two allow us to demonstrate 3.1--7.3$\times$ convolution speedups over dense convolution in AlexNet, on Intel Atom, Xeon, and Xeon Phi processors, spanning the spectrum from mobile devices to supercomputers. We also open source our project at https://github.com/IntelLabs/SkimCaffe.

Wei Wen, Sihang Deng, Tianjun Wei et al.

Query optimization is a crucial component for the efficacy of Retrieval-Augmented Generation (RAG) systems. While reinforcement learning (RL)-based agentic and reasoning methods have recently emerged as a promising direction on query optimization, most existing approaches focus on the expansion and abstraction of a single query. However, complex user queries are prevalent in real-world scenarios, often requiring multiple parallel and sequential search strategies to handle disambiguation and decomposition. Directly applying RL to these complex cases introduces significant hurdles. Determining the optimal number of sub-queries and effectively re-ranking and merging retrieved documents vastly expands the search space and complicates reward design, frequently leading to training instability. To address these challenges, we propose a novel RL framework called Adaptive Complex Query Optimization (ACQO). Our framework is designed to adaptively determine when and how to expand the search process. It features two core components: an Adaptive Query Reformulation (AQR) module that dynamically decides when to decompose a query into multiple sub-queries, and a Rank-Score Fusion (RSF) module that ensures robust result aggregation and provides stable reward signals for the learning agent. To mitigate training instabilities, we adopt a Curriculum Reinforcement Learning (CRL) approach, which stabilizes the training process by progressively introducing more challenging queries through a two-stage strategy. Our comprehensive experiments demonstrate that ACQO achieves state-of-the-art performance on three complex query benchmarks, significantly outperforming established baselines. The framework also showcases improved computational efficiency and broad compatibility with different retrieval architectures, establishing it as a powerful and generalizable solution for next-generation RAG systems.

Junjie Fei, Jun Chen, Zechun Liu et al.

Adapting Multimodal Large Language Models (MLLMs) for hour-long videos is bottlenecked by context limits. Dense visual streams saturate token budgets and exacerbate the lost-in-the-middle phenomenon. Existing heuristics, like sparse sampling or uniform pooling, blindly sacrifice fidelity by discarding decisive moments and wasting bandwidth on irrelevant backgrounds. We propose Tempo, an efficient query-aware framework compressing long videos for downstream understanding. Tempo leverages a Small Vision-Language Model (SVLM) as a local temporal compressor, casting token reduction as an early cross-modal distillation process to generate compact, intent-aligned representations in a single forward pass. To enforce strict budgets without breaking causality, we introduce Adaptive Token Allocation (ATA). Exploiting the SVLM's zero-shot relevance prior and semantic front-loading, ATA acts as a training-free $O(1)$ dynamic router. It allocates dense bandwidth to query-critical segments while compressing redundancies into minimal temporal anchors to maintain the global storyline. Extensive experiments show our 6B architecture achieves state-of-the-art performance with aggressive dynamic compression (0.5-16 tokens/frame). On the extreme-long LVBench (4101s), Tempo scores 52.3 under a strict 8K visual budget, outperforming GPT-4o and Gemini 1.5 Pro. Scaling to 2048 frames reaches 53.7. Crucially, Tempo compresses hour-long videos substantially below theoretical limits, proving true long-form video understanding relies on intent-driven efficiency rather than greedily padded context windows.

Keyu Chen, Zhifeng Shen, Daohai Yu et al.

The increasing scale and complexity of large language models (LLMs) pose significant inference latency challenges, primarily due to their autoregressive decoding paradigm characterized by the sequential nature of next-token prediction. By re-examining the outputs of autoregressive models, we observed that some segments exhibit parallelizable structures, which we term intrinsic parallelism. Decoding each parallelizable branch simultaneously (i.e. parallel decoding) can significantly improve the overall inference speed of LLMs. In this paper, we propose an Adaptive Serial-Parallel Decoding (ASPD), which addresses two core challenges: automated construction of parallelizable data and efficient parallel decoding mechanism. More specifically, we introduce a non-invasive pipeline that automatically extracts and validates parallelizable structures from the responses of autoregressive models. To empower efficient adaptive serial-parallel decoding, we implement a Hybrid Decoding Engine which enables seamless transitions between serial and parallel decoding modes while maintaining a reusable KV cache, maximizing computational efficiency. Extensive evaluations across General Tasks, Retrieval-Augmented Generation, Mathematical Reasoning, demonstrate that ASPD achieves unprecedented performance in both effectiveness and efficiency. Notably, on Vicuna Bench, our method achieves up to 3.19x speedup (1.85x on average) while maintaining response quality within 1% difference compared to autoregressive models, realizing significant acceleration without compromising generation quality. Our framework sets a groundbreaking benchmark for efficient LLM parallel inference, paving the way for its deployment in latency-sensitive applications such as AI-powered customer service bots and answer retrieval engines.

Sixun Dong, Juhua Hu, Steven Li et al.

Most vision-language models (VLMs) apply a large language model (LLM) as the decoder, where the response tokens are generated sequentially through autoregression. Therefore, the number of output tokens can be the bottleneck of the end-to-end latency. However, different models may require vastly different numbers of output tokens to achieve comparable performance. In this work, we conduct a comprehensive analysis of the latency across different components of VLMs on simulated data. The experiment shows that a large model with fewer output tokens can be more efficient than a small model with a long output sequence. The empirical study on diverse real-world benchmarks confirms the observation that a large model can achieve better or comparable performance as a small model with significantly fewer output tokens. To leverage the efficiency of large models, we propose a multi-agent inference framework that keeps large models with short responses but transfers the key reasoning tokens from the small model when necessary. The comparison on benchmark tasks demonstrates that by reusing the reasoning tokens from small models, it can help approach the performance of a large model with its own reasoning, which confirms the effectiveness of our proposal.

Zhichen Zeng, Xiaolong Liu, Mengyue Hang et al.

Click-through rate (CTR) prediction, which predicts the probability of a user clicking an ad, is a fundamental task in recommender systems. The emergence of heterogeneous information, such as user profile and behavior sequences, depicts user interests from different aspects. A mutually beneficial integration of heterogeneous information is the cornerstone towards the success of CTR prediction. However, most of the existing methods suffer from two fundamental limitations, including (1) insufficient inter-mode interaction due to the unidirectional information flow between modes, and (2) aggressive information aggregation caused by early summarization, resulting in excessive information loss. To address the above limitations, we propose a novel module named InterFormer to learn heterogeneous information interaction in an interleaving style. To achieve better interaction learning, InterFormer enables bidirectional information flow for mutually beneficial learning across different modes. To avoid aggressive information aggregation, we retain complete information in each data mode and use a separate bridging arch for effective information selection and summarization. Our proposed InterFormer achieves state-of-the-art performance on three public datasets and a large-scale industrial dataset.

Yuhui Xu, Yuxi Li, Shuai Zhang et al.

To enable DNNs on edge devices like mobile phones, low-rank approximation has been widely adopted because of its solid theoretical rationale and efficient implementations. Several previous works attempted to directly approximate a pretrained model by low-rank decomposition; however, small approximation errors in parameters can ripple over a large prediction loss. As a result, performance usually drops significantly and a sophisticated effort on fine-tuning is required to recover accuracy. Apparently, it is not optimal to separate low-rank approximation from training. Unlike previous works, this paper integrates low rank approximation and regularization into the training process. We propose Trained Rank Pruning (TRP), which alternates between low rank approximation and training. TRP maintains the capacity of the original network while imposing low-rank constraints during training. A nuclear regularization optimized by stochastic sub-gradient descent is utilized to further promote low rank in TRP. The TRP trained network inherently has a low-rank structure, and is approximated with negligible performance loss, thus eliminating the fine-tuning process after low rank decomposition. The proposed method is comprehensively evaluated on CIFAR-10 and ImageNet, outperforming previous compression methods using low rank approximation.

Huanrui Yang, Minxue Tang, Wei Wen et al.

Modern deep neural networks (DNNs) often require high memory consumption and large computational loads. In order to deploy DNN algorithms efficiently on edge or mobile devices, a series of DNN compression algorithms have been explored, including factorization methods. Factorization methods approximate the weight matrix of a DNN layer with the multiplication of two or multiple low-rank matrices. However, it is hard to measure the ranks of DNN layers during the training process. Previous works mainly induce low-rank through implicit approximations or via costly singular value decomposition (SVD) process on every training step. The former approach usually induces a high accuracy loss while the latter has a low efficiency. In this work, we propose SVD training, the first method to explicitly achieve low-rank DNNs during training without applying SVD on every step. SVD training first decomposes each layer into the form of its full-rank SVD, then performs training directly on the decomposed weights. We add orthogonality regularization to the singular vectors, which ensure the valid form of SVD and avoid gradient vanishing/exploding. Low-rank is encouraged by applying sparsity-inducing regularizers on the singular values of each layer. Singular value pruning is applied at the end to explicitly reach a low-rank model. We empirically show that SVD training can significantly reduce the rank of DNN layers and achieve higher reduction on computation load under the same accuracy, comparing to not only previous factorization methods but also state-of-the-art filter pruning methods.

Wei Wen, Hanxiao Liu, Hai Li et al.

Neural Architecture Search methods are effective but often use complex algorithms to come up with the best architecture. We propose an approach with three basic steps that is conceptually much simpler. First we train N random architectures to generate N (architecture, validation accuracy) pairs and use them to train a regression model that predicts accuracy based on the architecture. Next, we use this regression model to predict the validation accuracies of a large number of random architectures. Finally, we train the top-K predicted architectures and deploy the model with the best validation result. While this approach seems simple, it is more than 20 times as sample efficient as Regularized Evolution on the NASBench-101 benchmark and can compete on ImageNet with more complex approaches based on weight sharing, such as ProxylessNAS.

Chunpeng Wu, Wei Wen, Yiran Chen et al.

Generative adversarial network (GAN) has greatly improved the quality of unsupervised image generation. Previous GAN-based methods often require a large amount of high-quality training data while producing a small number (e.g., tens) of classes. This work aims to scale up GANs to thousands of classes meanwhile reducing the use of high-quality data in training. We propose an image generation method based on conditional transferring features, which can capture pixel-level semantic changes when transforming low-quality images into high-quality ones. Moreover, self-supervision learning is integrated into our GAN architecture to provide more label-free semantic supervisory information observed from the training data. As such, training our GAN architecture requires much fewer high-quality images with a small number of additional low-quality images. The experiments on CIFAR-10 and STL-10 show that even removing 30% high-quality images from the training set, our method can still outperform previous ones. The scalability on object classes has been experimentally validated: our method with 30% fewer high-quality images obtains the best quality in generating 1,000 ImageNet classes, as well as generating all 3,755 classes of CASIA-HWDB1.0 Chinese handwriting characters.

Huanrui Yang, Wei Wen, Hai Li

In seeking for sparse and efficient neural network models, many previous works investigated on enforcing L1 or L0 regularizers to encourage weight sparsity during training. The L0 regularizer measures the parameter sparsity directly and is invariant to the scaling of parameter values, but it cannot provide useful gradients, and therefore requires complex optimization techniques. The L1 regularizer is almost everywhere differentiable and can be easily optimized with gradient descent. Yet it is not scale-invariant, causing the same shrinking rate to all parameters, which is inefficient in increasing sparsity. Inspired by the Hoyer measure (the ratio between L1 and L2 norms) used in traditional compressed sensing problems, we present DeepHoyer, a set of sparsity-inducing regularizers that are both differentiable almost everywhere and scale-invariant. Our experiments show that enforcing DeepHoyer regularizers can produce even sparser neural network models than previous works, under the same accuracy level. We also show that DeepHoyer can be applied to both element-wise and structural pruning.

Qing Yang, Wei Wen, Zuoguan Wang et al.

With the rapid scaling up of deep neural networks (DNNs), extensive research studies on network model compression such as weight pruning have been performed for improving deployment efficiency. This work aims to advance the compression beyond the weights to neuron activations. We propose the joint regularization technique which simultaneously regulates the distribution of weights and activations. By distinguishing and leveraging the significance difference among neuron responses and connections during learning, the jointly pruned network, namely \textit{JPnet}, optimizes the sparsity of activations and weights for improving execution efficiency. The derived deep sparsification of JPnet reveals more optimization space for the existing DNN accelerators dedicated for sparse matrix operations. We thoroughly evaluate the effectiveness of joint regularization through various network models with different activation functions and on different datasets. With $0.4\%$ degradation constraint on inference accuracy, a JPnet can save $72.3\% \sim 98.8\%$ of computation cost compared to the original dense models, with up to $5.2\times$ and $12.3\times$ reductions in activation and weight numbers, respectively.

Sangkug Lym, Esha Choukse, Siavash Zangeneh et al.

State-of-the-art convolutional neural networks (CNNs) used in vision applications have large models with numerous weights. Training these models is very compute- and memory-resource intensive. Much research has been done on pruning or compressing these models to reduce the cost of inference, but little work has addressed the costs of training. We focus precisely on accelerating training. We propose PruneTrain, a cost-efficient mechanism that gradually reduces the training cost during training. PruneTrain uses a structured group-lasso regularization approach that drives the training optimization toward both high accuracy and small weight values. Small weights can then be periodically removed by reconfiguring the network model to a smaller one. By using a structured-pruning approach and additional reconfiguration techniques we introduce, the pruned model can still be efficiently processed on a GPU accelerator. Overall, PruneTrain achieves a reduction of 39% in the end-to-end training time of ResNet50 for ImageNet by reducing computation cost by 40% in FLOPs, memory accesses by 37% for memory bandwidth bound layers, and the inter-accelerator communication by 55%.

Sangkug Lym, Armand Behroozi, Wei Wen et al.

Training convolutional neural networks (CNNs) requires intense computations and high memory bandwidth. We find that bandwidth today is over-provisioned because most memory accesses in CNN training can be eliminated by rearranging computation to better utilize on-chip buffers and avoid traffic resulting from large per-layer memory footprints. We introduce the MBS CNN training approach that significantly reduces memory traffic by partially serializing mini-batch processing across groups of layers. This optimizes reuse within on-chip buffers and balances both intra-layer and inter-layer reuse. We also introduce the WaveCore CNN training accelerator that effectively trains CNNs in the MBS approach with high functional-unit utilization. Combined, WaveCore and MBS reduce DRAM traffic by 75%, improve performance by 53%, and save 26% system energy for modern deep CNN training compared to conventional training mechanisms and accelerators.

Wei Wen, Yandan Wang, Feng Yan et al.

In Deep Learning, Stochastic Gradient Descent (SGD) is usually selected as a training method because of its efficiency; however, recently, a problem in SGD gains research interest: sharp minima in Deep Neural Networks (DNNs) have poor generalization; especially, large-batch SGD tends to converge to sharp minima. It becomes an open question whether escaping sharp minima can improve the generalization. To answer this question, we propose SmoothOut framework to smooth out sharp minima in DNNs and thereby improve generalization. In a nutshell, SmoothOut perturbs multiple copies of the DNN by noise injection and averages these copies. Injecting noises to SGD is widely used in the literature, but SmoothOut differs in lots of ways: (1) a de-noising process is applied before parameter updating; (2) noise strength is adapted to filter norm; (3) an alternative interpretation on the advantage of noise injection, from the perspective of sharpness and generalization; (4) usage of uniform noise instead of Gaussian noise. We prove that SmoothOut can eliminate sharp minima. Training multiple DNN copies is inefficient, we further propose an unbiased stochastic SmoothOut which only introduces the overhead of noise injecting and de-noising per batch. An adaptive variant of SmoothOut, AdaSmoothOut, is also proposed to improve generalization. In a variety of experiments, SmoothOut and AdaSmoothOut consistently improve generalization in both small-batch and large-batch training on the top of state-of-the-art solutions.

Chunpeng Wu, Wei Wen, Tariq Afzal et al.

Recently, DNN model compression based on network architecture design, e.g., SqueezeNet, attracted a lot attention. No accuracy drop on image classification is observed on these extremely compact networks, compared to well-known models. An emerging question, however, is whether these model compression techniques hurt DNN's learning ability other than classifying images on a single dataset. Our preliminary experiment shows that these compression methods could degrade domain adaptation (DA) ability, though the classification performance is preserved. Therefore, we propose a new compact network architecture and unsupervised DA method in this paper. The DNN is built on a new basic module Conv-M which provides more diverse feature extractors without significantly increasing parameters. The unified framework of our DA method will simultaneously learn invariance across domains, reduce divergence of feature representations, and adapt label prediction. Our DNN has 4.1M parameters, which is only 6.7% of AlexNet or 59% of GoogLeNet. Experiments show that our DNN obtains GoogLeNet-level accuracy both on classification and DA, and our DA method slightly outperforms previous competitive ones. Put all together, our DA strategy based on our DNN achieves state-of-the-art on sixteen of total eighteen DA tasks on popular Office-31 and Office-Caltech datasets.

Yandan Wang, Wei Wen, Beiye Liu et al.

Synapse crossbar is an elementary structure in Neuromorphic Computing Systems (NCS). However, the limited size of crossbars and heavy routing congestion impedes the NCS implementations of big neural networks. In this paper, we propose a two-step framework (namely, group scissor) to scale NCS designs to big neural networks. The first step is rank clipping, which integrates low-rank approximation into the training to reduce total crossbar area. The second step is group connection deletion, which structurally prunes connections to reduce routing congestion between crossbars. Tested on convolutional neural networks of LeNet on MNIST database and ConvNet on CIFAR-10 database, our experiments show significant reduction of crossbar area and routing area in NCS designs. Without accuracy loss, rank clipping reduces total crossbar area to 13.62\% and 51.81\% in the NCS designs of LeNet and ConvNet, respectively. Following rank clipping, group connection deletion further reduces the routing area of LeNet and ConvNet to 8.1\% and 52.06\%, respectively.

Yandan Wang, Wei Wen, Linghao Song et al.

Brain inspired neuromorphic computing has demonstrated remarkable advantages over traditional von Neumann architecture for its high energy efficiency and parallel data processing. However, the limited resolution of synaptic weights degrades system accuracy and thus impedes the use of neuromorphic systems. In this work, we propose three orthogonal methods to learn synapses with one-level precision, namely, distribution-aware quantization, quantization regularization and bias tuning, to make image classification accuracy comparable to the state-of-the-art. Experiments on both multi-layer perception and convolutional neural networks show that the accuracy drop can be well controlled within 0.19% (5.53%) for MNIST (CIFAR-10) database, compared to an ideal system without quantization.