Shihao Ji

Xiulong Yang, Sheng-Min Shih, Yinlin Fu et al.

Diffusion Denoising Probability Models (DDPM) and Vision Transformer (ViT) have demonstrated significant progress in generative tasks and discriminative tasks, respectively, and thus far these models have largely been developed in their own domains. In this paper, we establish a direct connection between DDPM and ViT by integrating the ViT architecture into DDPM, and introduce a new generative model called Generative ViT (GenViT). The modeling flexibility of ViT enables us to further extend GenViT to hybrid discriminative-generative modeling, and introduce a Hybrid ViT (HybViT). Our work is among the first to explore a single ViT for image generation and classification jointly. We conduct a series of experiments to analyze the performance of proposed models and demonstrate their superiority over prior state-of-the-arts in both generative and discriminative tasks. Our code and pre-trained models can be found in https://github.com/sndnyang/Diffusion_ViT .

Qing Su, Shihao Ji

Current stereo matching techniques are challenged by restricted searching space, occluded regions and sheer size. While single image depth estimation is spared from these challenges and can achieve satisfactory results with the extracted monocular cues, the lack of stereoscopic relationship renders the monocular prediction less reliable on its own, especially in highly dynamic or cluttered environments. To address these issues in both scenarios, we present an optic-chiasm-inspired self-supervised binocular depth estimation method, wherein a vision transformer (ViT) with gated positional cross-attention (GPCA) layers is designed to enable feature-sensitive pattern retrieval between views while retaining the extensive context information aggregated through self-attentions. Monocular cues from a single view are thereafter conditionally rectified by a blending layer with the retrieved pattern pairs. This crossover design is biologically analogous to the optic-chasma structure in the human visual system and hence the name, ChiTransformer. Our experiments show that this architecture yields substantial improvements over state-of-the-art self-supervised stereo approaches by 11%, and can be used on both rectilinear and non-rectilinear (e.g., fisheye) images. Project is available at https://github.com/ISL-CV/ChiTransformer.

Xiulong Yang, Qing Su, Shihao Ji

Can we train a hybrid discriminative-generative model within a single network? This question has recently been answered in the affirmative, introducing the field of Joint Energy-based Model (JEM), which achieves high classification accuracy and image generation quality simultaneously. Despite recent advances, there remain two performance gaps: the accuracy gap to the standard softmax classifier, and the generation quality gap to state-of-the-art generative models. In this paper, we introduce a variety of training techniques to bridge the accuracy gap and the generation quality gap of JEM. 1) We incorporate a recently proposed sharpness-aware minimization (SAM) framework to train JEM, which promotes the energy landscape smoothness and the generalizability of JEM. 2) We exclude data augmentation from the maximum likelihood estimate pipeline of JEM, and mitigate the negative impact of data augmentation to image generation quality. Extensive experiments on multiple datasets demonstrate that our SADA-JEM achieves state-of-the-art performances and outperforms JEM in image classification, image generation, calibration, out-of-distribution detection and adversarial robustness by a notable margin. Our code is available at https://github.com/sndnyang/SADAJEM.

Hui Ye, Rajshekhar Sunderraman, Shihao Ji

The eXtreme Multi-label text Classification(XMC) refers to training a classifier that assigns a text sample with relevant labels from an extremely large-scale label set (e.g., millions of labels). We propose MatchXML, an efficient text-label matching framework for XMC. We observe that the label embeddings generated from the sparse Term Frequency-Inverse Document Frequency(TF-IDF) features have several limitations. We thus propose label2vec to effectively train the semantic dense label embeddings by the Skip-gram model. The dense label embeddings are then used to build a Hierarchical Label Tree by clustering. In fine-tuning the pre-trained encoder Transformer, we formulate the multi-label text classification as a text-label matching problem in a bipartite graph. We then extract the dense text representations from the fine-tuned Transformer. Besides the fine-tuned dense text embeddings, we also extract the static dense sentence embeddings from a pre-trained Sentence Transformer. Finally, a linear ranker is trained by utilizing the sparse TF-IDF features, the fine-tuned dense text representations and static dense sentence features. Experimental results demonstrate that MatchXML achieves state-of-the-art accuracy on five out of six datasets. As for the speed, MatchXML outperforms the competing methods on all the six datasets. Our source code is publicly available at https://github.com/huiyegit/MatchXML.

Qing Su, Anton Netchaev, Hai Li et al.

Current self-supervised learning (SSL) methods (e.g., SimCLR, DINO, VICReg,MOCOv3) target primarily on representations at instance level and do not generalize well to dense prediction tasks, such as object detection and segmentation.Towards aligning SSL with dense predictions, this paper demonstrates for the first time the underlying mean-shift clustering process of Vision Transformers (ViT), which aligns well with natural image semantics (e.g., a world of objects and stuffs). By employing transformer for joint embedding and clustering, we propose a two-level feature clustering SSL method, coined Feature-Level Self-supervised Learning (FLSL). We present the formal definition of the FLSL problem and construct the objectives from the mean-shift and k-means perspectives. We show that FLSL promotes remarkable semantic cluster representations and learns an embedding scheme amenable to intra-view and inter-view feature clustering. Experiments show that FLSL yields significant improvements in dense prediction tasks, achieving 44.9 (+2.8)% AP and 46.5% AP in object detection, as well as 40.8 (+2.3)% AP and 42.1% AP in instance segmentation on MS-COCO, using Mask R-CNN with ViT-S/16 and ViT-S/8 as backbone, respectively. FLSL consistently outperforms existing SSL methods across additional benchmarks, including UAV17 object detection on UAVDT, and video instance segmentation on DAVIS 2017.We conclude by presenting visualization and various ablation studies to better understand the success of FLSL. The source code is available at https://github.com/ISL-CV/FLSL.

Xiulong Yang, Shihao Ji

Energy-based models (EBMs) exhibit a variety of desirable properties in predictive tasks, such as generality, simplicity and compositionality. However, training EBMs on high-dimensional datasets remains unstable and expensive. In this paper, we present a Manifold EBM (M-EBM) to boost the overall performance of unconditional EBM and Joint Energy-based Model (JEM). Despite its simplicity, M-EBM significantly improves unconditional EBMs in training stability and speed on a host of benchmark datasets, such as CIFAR10, CIFAR100, CelebA-HQ, and ImageNet 32x32. Once class labels are available, label-incorporated M-EBM (M-JEM) further surpasses M-EBM in image generation quality with an over 40% FID improvement, while enjoying improved accuracy. The code can be found at https://github.com/sndnyang/mebm.

Shihao Ji, Mingyu Li, Zihui Song

The quadratic complexity of self-attention remains a bottleneck for Large Language Models (LLMs) processing ultra-long contexts. The Naive Bayes Cognitive Engine (NBCE) parallelizes long-context inference by chunking documents and routing to the lowest-entropy chunk at each decoding step. This hard-selection strategy causes semantic fragmentation during cross-chunk reasoning, as abrupt routing changes between adjacent tokens disrupt the model's contextual grounding. We present Soft-NBCE, a lightweight extension that replaces discrete chunk selection with soft entropy-weighted chunk fusion. A temperature-scaled Softmax over predictive entropies assigns continuous weights to all chunks, enabling log-space aggregation across chunk-conditioned distributions. To partially compensate for the conditional independence assumption introduced by chunking, we propose Consistency Distillation, a LoRA-based self-distillation that constrains the chunked logit distribution toward a full-context teacher via KL-divergence. On LongBench multi-hop benchmarks, Soft-NBCE with Consistency Distillation improves consistently over NBCE-style baselines (MuSiQue F1: 0.310 vs.\ 0.275 for Vanilla NBCE; HotpotQA F1: 0.479 vs.\ 0.427) while maintaining retrieval accuracy (NIAH-32K: 0.909) at O(L^2/n) peak memory.

Shihao Ji, Haotao Tan, Zihui Song et al.

Large Language Models (LLMs) are increasingly used with formal interactive theorem provers such as Lean 4. Scaling these systems with reinforcement learning or search methods requires process reward models (PRMs) that can evaluate intermediate reasoning steps. Existing reward-model designs expose a practical trade-off. Value-head models provide continuous scores but modify the generative model interface, while generative reward models preserve textual rationales but are poorly matched to continuous floating-point regression because numeric values are split across tokens. We introduce Expected Value Alignment (EVA), a reward-modeling procedure that keeps the surface output discrete while extracting continuous scores from the model's token distribution. The model emits integer scores in a structured JSON format, and EVA computes a continuous score as the expectation over the logits of the corresponding anchor tokens. Training combines the causal language modeling objective with an auxiliary mean squared error loss on these expected values. We instantiate EVA in \textit{Leibniz}, a reward model for Lean 4 formal verification, and evaluate it against zero-shot and reward-modeling baselines. The evaluation demonstrates that continuous logit-based scoring significantly reduces discretization artifacts while retaining the interpretability of generative critiques.

Yang Li, Xin Ma, Raj Sunderraman et al.

Neuroimaging-based prediction methods for intelligence and cognitive abilities have seen a rapid development in literature. Among different neuroimaging modalities, prediction based on functional connectivity (FC) has shown great promise. Most literature has focused on prediction using static FC, but there are limited investigations on the merits of such analysis compared to prediction based on dynamic FC or region level functional magnetic resonance imaging (fMRI) times series that encode temporal variability. To account for the temporal dynamics in fMRI data, we propose a deep neural network involving bi-directional long short-term memory (bi-LSTM) approach that also incorporates feature selection mechanism. The proposed pipeline is implemented via an efficient GPU computation framework and applied to predict intelligence scores based on region level fMRI time series as well as dynamic FC. We compare the prediction performance for different intelligence measures based on static FC, dynamic FC, and region level time series acquired from the Adolescent Brain Cognitive Development (ABCD) study involving close to 7000 individuals. Our detailed analysis illustrates that static FC consistently has inferior prediction performance compared to region level time series or dynamic FC for unimodal rest and task fMRI experiments, and in almost all cases using a combination of task and rest features. In addition, the proposed bi-LSTM pipeline based on region level time series identifies several shared and differential important brain regions across task and rest fMRI experiments that drive intelligence prediction. A test-retest analysis of the selected features shows strong reliability across cross-validation folds. Given the large sample size from ABCD study, our results provide strong evidence that superior prediction of intelligence can be achieved by accounting for temporal variations in fMRI.

Yang Ye, Xiulong Yang, Shihao Ji

Processing large point clouds is a challenging task. Therefore, the data is often downsampled to a smaller size such that it can be stored, transmitted and processed more efficiently without incurring significant performance degradation. Traditional task-agnostic sampling methods, such as farthest point sampling (FPS), do not consider downstream tasks when sampling point clouds, and thus non-informative points to the tasks are often sampled. This paper explores a task-oriented sampling for 3D point clouds, and aims to sample a subset of points that are tailored specifically to a downstream task of interest. Similar to FPS, we assume that point to be sampled next should depend heavily on the points that have already been sampled. We thus formulate point cloud sampling as a sequential generation process, and develop an attention-based point cloud sampling network (APSNet) to tackle this problem. At each time step, APSNet attends to all the points in a cloud by utilizing the history of previously sampled points, and samples the most informative one. Both supervised learning and knowledge distillation-based self-supervised learning of APSNet are proposed. Moreover, joint training of APSNet over multiple sample sizes is investigated, leading to a single APSNet that can generate arbitrary length of samples with prominent performances. Extensive experiments demonstrate the superior performance of APSNet against state-of-the-arts in various downstream tasks, including 3D point cloud classification, reconstruction, and registration.

Mingchen Li, Shihao Ji

Building query graphs from natural language questions is an important step in complex question answering over knowledge graph (Complex KGQA). In general, a question can be correctly answered if its query graph is built correctly and the right answer is then retrieved by issuing the query graph against the KG. Therefore, this paper focuses on query graph generation from natural language questions. Existing approaches for query graph generation ignore the semantic structure of a question, resulting in a large number of noisy query graph candidates that undermine prediction accuracies. In this paper, we define six semantic structures from common questions in KGQA and develop a novel Structure-BERT to predict the semantic structure of a question. By doing so, we can first filter out noisy candidate query graphs, and then rank the remaining candidates with a BERT-based ranking model. Extensive experiments on two popular benchmarks MetaQA and WebQuestionsSP (WSP) demonstrate the effectiveness of our method as compared to state-of-the-arts.

Kaiyang Li, Shihao Ji, Zhipeng Cai et al.

Approximate subgraph matching (ASM) is a task that determines the approximate presence of a given query graph in a large target graph. Being an NP-hard problem, ASM is critical in graph analysis with a myriad of applications ranging from database systems and network science to biochemistry and privacy. Existing techniques often employ heuristic search strategies, which cannot fully utilize the graph information, leading to sub-optimal solutions. This paper proposes a Reinforcement Learning based Approximate Subgraph Matching (RL-ASM) algorithm that exploits graph transformers to effectively extract graph representations and RL-based policies for ASM. Our model is built upon the branch-and-bound algorithm that selects one pair of nodes from the two input graphs at a time for potential matches. Instead of using heuristics, we exploit a Graph Transformer architecture to extract feature representations that encode the full graph information. To enhance the training of the RL policy, we use supervised signals to guide our agent in an imitation learning stage. Subsequently, the policy is fine-tuned with the Proximal Policy Optimization (PPO) that optimizes the accumulative long-term rewards over episodes. Extensive experiments on both synthetic and real-world datasets demonstrate that our RL-ASM outperforms existing methods in terms of effectiveness and efficiency. Our source code is available at https://github.com/KaiyangLi1992/RL-ASM.

Qing Su, Shihao Ji

Distillation-based self-supervised learning typically leads to more compressed representations due to its radical clustering process and the implementation of a sharper target distribution. To overcome this limitation and preserve more information from input, we introduce UDI, conceptualized as Unsqueezed Distillation-based self-supervised learning (SSL). UDI enriches the learned representation by encouraging multimodal prediction distilled from a consolidated profile of local predictions that are derived via stratified sampling. Our evaluations show that UDI not only promotes semantically meaningful representations at instance level, delivering superior or competitive results to state-of-the-art SSL methods in image classification, but also effectively preserves the nuisance of input, which yields significant improvement in dense prediction tasks, including object detection and segmentation. Additionally, UDI performs competitively in low-shot image classification, improving the scalability of joint-embedding pipelines. Various visualizations and ablation studies are presented to further elucidate the mechanisms behind UDI. Our source code is available at https://github.com/ISL-CV/udi.

Yuan Zhuang, Yuexin Bian, Sihong He et al.

Scaling critic capacity is a promising direction for enhancing off-policy reinforcement learning (RL). However, larger critics are prone to overfitting and unstable in replay-buffer-based bootstrap training. This paper leverages Low-Rank Adaptation (LoRA) as a structural-sparsity regularizer for off-policy critics. Our approach freezes randomly initialized base matrices and solely optimizes low-rank adapters, thereby constraining critic updates to a low-dimensional subspace. Built on top of SimbaV2, we further develop a LoRA formulation, compatible with SimbaV2, that preserves its hyperspherical normalization geometry under frozen-backbone training. We evaluate our method with SAC and FastTD3 on DeepMind Control locomotion and IsaacLab robotics benchmarks. LoRA consistently achieves lower critic loss during training and stronger policy performance. Extensive experiments demonstrate that adaptive low-rank updates provide a simple, scalable, and effective structural regularization for critic learning in off-policy RL.

Kaiyang Li, Shaobo Han, Qing Su et al.

Low-Rank Adaptation (LoRA) has become the de facto parameter-efficient fine-tuning (PEFT) method for large language models (LLMs) by constraining weight updates to low-rank matrices. Recent works such as Tied-LoRA, VeRA, and VB-LoRA push efficiency further by introducing additional constraints to reduce the trainable parameter space. In this paper, we show that the parameter space reduction strategies employed by these LoRA variants can be formulated within a unified framework, Uni-LoRA, where the LoRA parameter space, flattened as a high-dimensional vector space $R^D$, can be reconstructed through a projection from a subspace R^d, with $d \ll D$. We demonstrate that the fundamental difference among various LoRA methods lies in the choice of the projection matrix, $P \in R^{D \times d}$.Most existing LoRA variants rely on layer-wise or structure-specific projections that limit cross-layer parameter sharing, thereby compromising parameter efficiency. In light of this, we introduce an efficient and theoretically grounded projection matrix that is isometric, enabling global parameter sharing and reducing computation overhead. Furthermore, under the unified view of Uni-LoRA, this design requires only a single trainable vector to reconstruct LoRA parameters for the entire LLM - making Uni-LoRA both a unified framework and a "one-vector-only" solution. Extensive experiments on GLUE, mathematical reasoning, and instruction tuning benchmarks demonstrate that Uni-LoRA achieves state-of-the-art parameter efficiency while outperforming or matching prior approaches in predictive performance. Our code is available at https://github.com/KaiyangLi1992/Uni-LoRA.

Kaiyang Li, Shaobo Han, Qing Su et al.

Model merging aims to combine multiple task-specific expert models into a single model without joint retraining, offering a practical alternative to multi-task learning when data access or computational budget is limited. Existing methods, however, face two key limitations: (1) they overlook the valuable inductive bias of strong anchor models and estimate the merged weights from scratch, and (2) they rely on a shared hyperparameter setting across different modules of the network, lacking a global optimization strategy. This paper introduces Bayesian Model Merging (BMM), a plug-and-play bi-level optimization framework, where the inner level formulates the model merging as an activation-based Bayesian regression under a strong prior induced by an anchor model, yielding an efficient closed-form solution; and the outer level leverages a Bayesian optimization procedure to search module-specific hyperparameters globally based on a small validation set. Furthermore, we reveal a key alignment between activation statistics and task vectors, enabling us to derive a data-free variant of BMM that estimates the Gram matrix for regression without any auxiliary data. Across extensive benchmarks, including up to 20-task merging in vision and 5-task merging in language, BMM consistently outperforms all plug-and-play anchor baselines (e.g., TA, WUDI-Merging, and TSV). In particular, on the ViT-L/14 benchmark for 8-task merging, a single merged model reaches 95.1, closely matching the average performance of eight task-specific experts (95.8).

Xiang Li, Shihao Ji

Adversarial patch attack is a family of attack algorithms that perturb a part of image to fool a deep neural network model. Existing patch attacks mostly consider injecting adversarial patches at input-agnostic locations: either a predefined location or a random location. This attack setup may be sufficient for attack but has considerable limitations when using it for adversarial training. Thus, robust models trained with existing patch attacks cannot effectively defend other adversarial attacks. In this paper, we first propose an end-to-end patch attack algorithm, Generative Dynamic Patch Attack (GDPA), which generates both patch pattern and patch location adversarially for each input image. We show that GDPA is a generic attack framework that can produce dynamic/static and visible/invisible patches with a few configuration changes. Secondly, GDPA can be readily integrated for adversarial training to improve model robustness to various adversarial attacks. Extensive experiments on VGGFace, Traffic Sign and ImageNet show that GDPA achieves higher attack success rates than state-of-the-art patch attacks, while adversarially trained model with GDPA demonstrates superior robustness to adversarial patch attacks than competing methods. Our source code can be found at https://github.com/lxuniverse/gdpa.

Xiulong Yang, Shihao Ji

Joint Energy-based Model (JEM) is a recently proposed hybrid model that retains strong discriminative power of modern CNN classifiers, while generating samples rivaling the quality of GAN-based approaches. In this paper, we propose a variety of new training procedures and architecture features to improve JEM's accuracy, training stability, and speed altogether. 1) We propose a proximal SGLD to generate samples in the proximity of samples from the previous step, which improves the stability. 2) We further treat the approximate maximum likelihood learning of EBM as a multi-step differential game, and extend the YOPO framework to cut out redundant calculations during backpropagation, which accelerates the training substantially. 3) Rather than initializing SGLD chain from random noise, we introduce a new informative initialization that samples from a distribution estimated from training data. 4) This informative initialization allows us to enable batch normalization in JEM, which further releases the power of modern CNN architectures for hybrid modeling. Code: https://github.com/sndnyang/JEMPP

Yang Li, Shihao Ji

Training deep neural networks with an $L_0$ regularization is one of the prominent approaches for network pruning or sparsification. The method prunes the network during training by encouraging weights to become exactly zero. However, recent work of Gale et al. reveals that although this method yields high compression rates on smaller datasets, it performs inconsistently on large-scale learning tasks, such as ResNet50 on ImageNet. We analyze this phenomenon through the lens of variational inference and find that it is likely due to the independent modeling of binary gates, the mean-field approximation, which is known in Bayesian statistics for its poor performance due to the crude approximation. To mitigate this deficiency, we propose a dependency modeling of binary gates, which can be modeled effectively as a multi-layer perceptron (MLP). We term our algorithm Dep-$L_0$ as it prunes networks via a dependency-enabled $L_0$ regularization. Extensive experiments on CIFAR10, CIFAR100 and ImageNet with VGG16, ResNet50, ResNet56 show that our Dep-$L_0$ outperforms the original $L_0$-HC algorithm of Louizos et al. by a significant margin, especially on ImageNet. Compared with the state-of-the-arts network sparsification algorithms, our dependency modeling makes the $L_0$-based sparsification once again very competitive on large-scale learning tasks. Our source code is available at https://github.com/leo-yangli/dep-l0.

Xiulong Yang, Hui Ye, Yang Ye et al.

Joint Energy-based Model (JEM) of Grathwohl et al. shows that a standard softmax classifier can be reinterpreted as an energy-based model (EBM) for the joint distribution p(x,y); the resulting model can be optimized to improve calibration, robustness, and out-of-distribution detection, while generating samples rivaling the quality of recent GAN-based approaches. However, the softmax classifier that JEM exploits is inherently discriminative and its latent feature space is not well formulated as probabilistic distributions, which may hinder its potential for image generation and incur training instability. We hypothesize that generative classifiers, such as Linear Discriminant Analysis (LDA), might be more suitable for image generation since generative classifiers model the data generation process explicitly. This paper therefore investigates an LDA classifier for image classification and generation. In particular, the Max-Mahalanobis Classifier (MMC), a special case of LDA, fits our goal very well. We show that our Generative MMC (GMMC) can be trained discriminatively, generatively, or jointly for image classification and generation. Extensive experiments on multiple datasets show that GMMC achieves state-of-the-art discriminative and generative performances, while outperforming JEM in calibration, adversarial robustness, and out-of-distribution detection by a significant margin. Our source code is available at https://github.com/sndnyang/GMMC.

Kaiyang Li, Guangchun Luo, Yang Ye et al.

Recently, the surge in popularity of Internet of Things (IoT), mobile devices, social media, etc. has opened up a large source for graph data. Graph embedding has been proved extremely useful to learn low-dimensional feature representations from graph structured data. These feature representations can be used for a variety of prediction tasks from node classification to link prediction. However, existing graph embedding methods do not consider users' privacy to prevent inference attacks. That is, adversaries can infer users' sensitive information by analyzing node representations learned from graph embedding algorithms. In this paper, we propose Adversarial Privacy Graph Embedding (APGE), a graph adversarial training framework that integrates the disentangling and purging mechanisms to remove users' private information from learned node representations. The proposed method preserves the structural information and utility attributes of a graph while concealing users' private attributes from inference attacks. Extensive experiments on real-world graph datasets demonstrate the superior performance of APGE compared to the state-of-the-arts. Our source code can be found at https://github.com/uJ62JHD/Privacy-Preserving-Social-Network-Embedding.

Xiang Li, Shihao Ji

Explaining the prediction of deep neural networks (DNNs) and semantic image compression are two active research areas of deep learning with a numerous of applications in decision-critical systems, such as surveillance cameras, drones and self-driving cars, where interpretable decision is critical and storage/network bandwidth is limited. In this paper, we propose a novel end-to-end Neural Image Compression and Explanation (NICE) framework that learns to (1) explain the predictions of convolutional neural networks (CNNs), and (2) subsequently compress the input images for efficient storage or transmission. Specifically, NICE generates a sparse mask over an input image by attaching a stochastic binary gate to each pixel of the image, whose parameters are learned through the interaction with the CNN classifier to be explained. The generated mask is able to capture the saliency of each pixel measured by its influence to the final prediction of CNN; it can also be used to produce a mixed-resolution image, where important pixels maintain their original high resolution and insignificant background pixels are subsampled to a low resolution. The produced images achieve a high compression rate (e.g., about 0.6x of original image file size), while retaining a similar classification accuracy. Extensive experiments across multiple image classification benchmarks demonstrate the superior performance of NICE compared to the state-of-the-art methods in terms of explanation quality and semantic image compression rate. Our code is available at: https://github.com/lxuniverse/NICE.

Yang Li, Shihao Ji

We consider network sparsification as an $L_0$-norm regularized binary optimization problem, where each unit of a neural network (e.g., weight, neuron, or channel, etc.) is attached with a stochastic binary gate, whose parameters are jointly optimized with original network parameters. The Augment-Reinforce-Merge (ARM), a recently proposed unbiased gradient estimator, is investigated for this binary optimization problem. Compared to the hard concrete gradient estimator from Louizos et al., ARM demonstrates superior performance of pruning network architectures while retaining almost the same accuracies of baseline methods. Similar to the hard concrete estimator, ARM also enables conditional computation during model training but with improved effectiveness due to the exact binary stochasticity. Thanks to the flexibility of ARM, many smooth or non-smooth parametric functions, such as scaled sigmoid or hard sigmoid, can be used to parameterize this binary optimization problem and the unbiasness of the ARM estimator is retained, while the hard concrete estimator has to rely on the hard sigmoid function to achieve conditional computation and thus accelerated training. Extensive experiments on multiple public datasets demonstrate state-of-the-art pruning rates with almost the same accuracies of baseline methods. The resulting algorithm $L_0$-ARM sparsifies the Wide-ResNet models on CIFAR-10 and CIFAR-100 while the hard concrete estimator cannot. The code is public available at https://github.com/leo-yangli/l0-arm.

Xiang Li, Shihao Ji

Deep neural networks (DNNs) have been enormously successful across a variety of prediction tasks. However, recent research shows that DNNs are particularly vulnerable to adversarial attacks, which poses a serious threat to their applications in security-sensitive systems. In this paper, we propose a simple yet effective defense algorithm Defense-VAE that uses variational autoencoder (VAE) to purge adversarial perturbations from contaminated images. The proposed method is generic and can defend white-box and black-box attacks without the need of retraining the original CNN classifiers, and can further strengthen the defense by retraining CNN or end-to-end finetuning the whole pipeline. In addition, the proposed method is very efficient compared to the optimization-based alternatives, such as Defense-GAN, since no iterative optimization is needed for online prediction. Extensive experiments on MNIST, Fashion-MNIST, CelebA and CIFAR-10 demonstrate the superior defense accuracy of Defense-VAE compared to Defense-GAN, while being 50x faster than the latter. This makes Defense-VAE widely deployable in real-time security-sensitive systems. Our source code can be found at https://github.com/lxuniverse/defense-vae.

Hui Ye, Rajshekhar Sunderraman, Shihao Ji

Unmanned Aerial Vehicles (UAVs), equipped with cameras, are employed in numerous applications, including aerial photography, surveillance, and agriculture. In these applications, robust object detection and tracking are essential for the effective deployment of UAVs. However, existing benchmarks for UAV applications are mainly designed for traditional 2D perception tasks, restricting the development of real-world applications that require a 3D understanding of the environment. Furthermore, despite recent advancements in single-UAV perception, limited views of a single UAV platform significantly constrain its perception capabilities over long distances or in occluded areas. To address these challenges, we introduce UAV3D, a benchmark designed to advance research in both 3D and collaborative 3D perception tasks with UAVs. UAV3D comprises 1,000 scenes, each of which has 20 frames with fully annotated 3D bounding boxes on vehicles. We provide the benchmark for four 3D perception tasks: single-UAV 3D object detection, single-UAV object tracking, collaborative-UAV 3D object detection, and collaborative-UAV object tracking. Our dataset and code are available at https://huiyegit.github.io/UAV3D_Benchmark/.

Shihao Ji, Zihui Song, Fucheng Zhong et al.

Recent advancements in large language models (LLMs) have demonstrated their impressive abilities in various reasoning and decision-making tasks. However, the quality and coherence of the reasoning process can still benefit from enhanced introspection and self-reflection. In this paper, we introduce Multiplex CoT (Chain of Thought), a method that enables LLMs to simulate a form of self-review while reasoning, by initiating double Chain of Thought (CoT) thinking. Multiplex CoT leverages the power of iterative reasoning, where the model generates an initial chain of thought and subsequently critiques and refines this reasoning with a second round of thought generation. This recursive approach allows for more coherent, logical, and robust answers, improving the overall decision-making process. We demonstrate how this method can be effectively implemented using simple prompt engineering in existing LLM architectures, achieving an effect similar to that of the Learning-Refinement Model (LRM) without the need for additional training. Additionally, we present a practical guide for implementing the method in Google Colab, enabling easy integration into real-world applications.

Zihui Song, Shihao Ji, Hongxi Li et al.

We investigate the hypothesis that real-data scaling laws are governed by progressive coverage of a latent predictive contribution spectrum rather than by token-frequency tails alone. We work with a suffix-automaton representation of text corpora and define a data-intrinsic global-KL predictive contribution spectrum, in which each state contributes according to its empirical mass times its KL deviation from a global next-token baseline. Across 12 real corpora, the tail slope of this spectrum is already strongly correlated with the empirical data-scaling exponent of a fixed small GPT learner. We then go beyond slope correlation and define, for each training size N, an effective truncation rank K(N) by matching the observed excess loss to the residual tail mass of the prepared 1000k global-KL spectrum. Empirically, log K is close to linear in log N, with pooled R^2 about 0.96 for the raw spectrum and R^2 about 0.90 for the smoothed spectrum. These findings provide strong empirical support for a simple mechanism picture: training scale advances an effective frontier through a predictive state spectrum, and the residual tail mass of that spectrum tracks the remaining excess loss.

Khaled Mohammed Saifuddin, Shihao Ji, Esra Akbas

Recent advancements in Graph Contrastive Learning (GCL) have demonstrated remarkable effectiveness in improving graph representations. However, relying on predefined augmentations (e.g., node dropping, edge perturbation, attribute masking) may result in the loss of task-relevant information and a lack of adaptability to diverse input data. Furthermore, the selection of negative samples remains rarely explored. In this paper, we introduce HyperGCL, a novel multimodal GCL framework from a hypergraph perspective. HyperGCL constructs three distinct hypergraph views by jointly utilizing the input graph's structure and attributes, enabling a comprehensive integration of multiple modalities in contrastive learning. A learnable adaptive topology augmentation technique enhances these views by preserving important relations and filtering out noise. View-specific encoders capture essential characteristics from each view, while a network-aware contrastive loss leverages the underlying topology to define positive and negative samples effectively. Extensive experiments on benchmark datasets demonstrate that HyperGCL achieves state-of-the-art node classification performance.

Lumen AI, Tengzhou No. 1 Middle School, Shihao Ji et al.

This paper proposes an innovative Multi-Modal Transformer framework (MMF-Trans) designed to significantly improve the prediction accuracy of the Chinese stock market by integrating multi-source heterogeneous information including macroeconomy, micro-market, financial text, and event knowledge. The framework consists of four core modules: (1) A four-channel parallel encoder that processes technical indicators, financial text, macro data, and event knowledge graph respectively for independent feature extraction of multi-modal data; (2) A dynamic gated cross-modal fusion mechanism that adaptively learns the importance of different modalities through differentiable weight allocation for effective information integration; (3) A time-aligned mixed-frequency processing layer that uses an innovative position encoding method to effectively fuse data of different time frequencies and solves the time alignment problem of heterogeneous data; (4) A graph attention-based event impact quantification module that captures the dynamic impact of events on the market through event knowledge graph and quantifies the event impact coefficient. We introduce a hybrid-frequency Transformer and Event2Vec algorithm to effectively fuse data of different frequencies and quantify the event impact. Experimental results show that in the prediction task of CSI 300 constituent stocks, the root mean square error (RMSE) of the MMF-Trans framework is reduced by 23.7% compared to the baseline model, the event response prediction accuracy is improved by 41.2%, and the Sharpe ratio is improved by 32.6%.

Yang Ye, Shihao Ji

As point cloud provides a natural and flexible representation usable in myriad applications (e.g., robotics and self-driving cars), the ability to synthesize point clouds for analysis becomes crucial. Recently, Xie et al. propose a generative model for unordered point sets in the form of an energy-based model (EBM). Despite the model achieving an impressive performance for point cloud generation, one separate model needs to be trained for each category to capture the complex point set distributions. Besides, their method is unable to classify point clouds directly and requires additional fine-tuning for classification. One interesting question is: Can we train a single network for a hybrid generative and discriminative model of point clouds? A similar question has recently been answered in the affirmative for images, introducing the framework of Joint Energy-based Model (JEM), which achieves high performance in image classification and generation simultaneously. This paper proposes GDPNet, the first hybrid Generative and Discriminative PointNet that extends JEM for point cloud classification and generation. Our GDPNet retains strong discriminative power of modern PointNet classifiers, while generating point cloud samples rivaling state-of-the-art generative approaches.

Shihao Ji, Zihui Song

The remarkable zero-shot reasoning capabilities of large-scale Visual Language Models (VLMs) on static images have yet to be fully translated to the video domain. Conventional video understanding models often rely on extensive, task-specific training on annotated datasets, a process that is both costly and limited in scalability. This paper introduces a novel, training-free framework for video understanding that circumvents end-to-end training by synergistically combining the rich semantic priors of pre-trained VLMs with classic machine learning algorithms for pattern discovery. Our core idea is to reframe video understanding as a self-supervised spatio-temporal clustering problem within a high-dimensional semantic feature space. The proposed pipeline first transforms a video stream into a semantic feature trajectory using the frozen visual encoder of a pre-trained VLM. Subsequently, we employ Kernel Temporal Segmentation (KTS), a robust machine learning technique, to partition the continuous feature stream into discrete, semantically coherent event segments. These segments are then subjected to unsupervised density-based clustering to identify recurring macroscopic scenes and themes throughout the video. By selecting representative keyframes from each discovered cluster and leveraging the VLM's generative capabilities for textual description, our framework automatically produces a structured, multi-modal summary of the video content. This approach provides an effective, interpretable, and model-agnostic pathway for zero-shot, automated structural analysis of video content.

Shihao Ji, Zihui Song

The Mixture of Experts (MoE) architecture enables the scaling of Large Language Models (LLMs) to trillions of parameters by activating a sparse subset of weights for each input, maintaining constant computational cost during inference. Concurrently, Low-Rank Adaptation (LoRA) has emerged as a dominant technique for parameter-efficiently fine-tuning LLMs on specialized tasks. In this work, we unify these two paradigms into a novel, end-to-end trainable framework named L-MoE: a Lightweight Mixture of LoRA Experts. L-MoE redefines MoE experts not as dense feed-forward networks, but as a collection of task-specialized, low-rank adapters. A lightweight gating network, trained jointly with the experts, learns to dynamically compose these LoRA adapters by computing a weighted average of their parameters for each input token. This composition is fully differentiable, allowing gradients from a standard auto-regressive language modeling objective to flow back through the entire architecture, simultaneously refining both the expert adapters and the routing strategy. This approach creates a highly parameter-efficient MoE model that is modular by design, allows for dynamic skill composition, and is trainable from end-to-end. We present the formal mathematical framework for L-MoE, detailing the differentiable routing mechanism and the joint optimization objective, thereby providing a new path toward building more efficient, scalable, and specialized language models.

Shihao Ji, Zihui Song, Jiajie Huang

Large Language Models (LLMs) hallucinate, generating factually incorrect yet confident assertions. We argue this stems from the Transformer's Softmax function, which creates "Artificial Certainty" by collapsing ambiguous attention scores into a single probability distribution, discarding uncertainty information at each layer. To fix this, we introduce the Credal Transformer, which replaces standard attention with a Credal Attention Mechanism (CAM) based on evidential theory. CAM produces a "credal set" (a set of distributions) instead of a single attention vector, with the set's size directly measuring model uncertainty. We implement this by re-conceptualizing attention scores as evidence masses for a Dirichlet distribution: sufficient evidence recovers standard attention, while insufficient evidence yields a diffuse distribution, representing ambiguity. Empirically, the Credal Transformer identifies out-of-distribution inputs, quantifies ambiguity, and significantly reduces confident errors on unanswerable questions by abstaining. Our contribution is a new architecture to mitigate hallucinations and a design paradigm that integrates uncertainty quantification directly into the model, providing a foundation for more reliable AI.

Yuan Zhuang, Yi Shen, Yuexin Bian et al.

Recent studies have shown that combining parameter-efficient fine-tuning (PEFT) with mixture-of-experts (MoE) is an effective strategy for adapting large language models (LLMs) to the downstream tasks. However, most existing approaches rely on conventional TopK routing, which requires careful hyperparameter tuning and assigns a fixed number of experts to each token. In this work, we propose LD-MoLE, a Learnable Dynamic routing mechanism for Mixture of LoRA Experts that enables adaptive, token-dependent, and layer-wise expert allocation. Our method replaces the non-differentiable TopK selection with a differentiable routing function and a closed-form solution. Moreover, our design allows the model to adaptively determine the number of experts to activate for each token at different layers. In addition, we introduce an analytical sparsity control objective to regularize the number of activated experts. Extensive experiments on the Qwen3-1.7B and Llama-3.2-3B models show that LD-MoLE achieves the highest average scores compared to state-of-the-art baselines, across a diverse set of benchmarks. Our method not only achieves superior performance, but also demonstrates the ability to learn token-dependent and layer-wise expert allocation.

Shihao Ji, Zihui Song

Continual or Lifelong Learning aims to develop models capable of acquiring new knowledge from a sequence of tasks without catastrophically forgetting what has been learned before. Existing approaches often rely on storing samples from previous tasks (experience replay) or employing complex regularization terms to protect learned weights. However, these methods face challenges related to data privacy, storage limitations, and performance degradation when tasks are dissimilar. To address these challenges, we introduce MyGO (Memory Yielding Generative Offline-consolidation), a novel lifelong learning framework inspired by the biological wake-sleep cycle. During the "wake" phase, the system rapidly learns a new task and trains a compact generative model (Generative Memory, G-mem) to capture its data distribution. During the "sleep" phase, the system enters an offline state, using all learned G-mem models to generate pseudo-data ("dreams") and consolidate new and old knowledge into a core feature extractor via knowledge distillation. This approach obviates the need to store any raw data, retaining only compact generative models, which offers significant advantages in privacy and storage efficiency. We evaluate MyGO on computer vision (Split-MNIST) and natural language processing (Split-AG News) benchmarks, comparing it against a sequential fine-tuning baseline. The results demonstrate that MyGO significantly mitigates catastrophic forgetting and maintains high average accuracy across tasks, proving the framework's effectiveness and domain-generality.

Lumen AI, Zaozhuang No. 28 Middle School, Shihao Ji et al.

This report details Lumen Labs' novel approach to processing Social Networking Service (SNS) data. We leverage knowledge distillation, specifically a simple distillation method inspired by DeepSeek-R1's CoT acquisition, combined with prompt hacking, to extract valuable training data from the Grok model. This data is then used to fine-tune a Phi-3-mini model, augmented with a mask-like mechanism specifically designed for handling the nuances of SNS data. Our method demonstrates state-of-the-art (SOTA) performance on several SNS data processing tasks, outperforming existing models like Grok, Phi-3, and GPT-4. We provide a comprehensive analysis of our approach, including mathematical formulations, engineering details, ablation studies, and comparative evaluations.

Lumen AI, Tengzhou No. 1 Middle School, Shihao Ji et al.

Large language models (LLMs) face significant token efficiency bottlenecks in code generation and logical reasoning tasks, a challenge that directly impacts inference cost and model interpretability. This paper proposes a formal framework based on symbolic compression,integrating combinatory logic, information-theoretic optimal encoding, and context-aware inference techniques to achieve a step-change improvement in token efficiency while preserving semantic integrity. We establish a mathematical framework within a functional programming paradigm, derive the quantitative relationship between symbolic density and model interpretability, and propose a differentiable compression factor metric to evaluate encoding efficiency. Furthermore, we leverage parameter-efficient fine-tuning (PEFT) techniques to achieve a low-cost application of the GAEL language. Experimental results show that this method achieves a 78.3% token compression rate in code generation tasks while improving logical traceability by 62% through structural explicitness. This research provides new theoretical tools for efficient inference in LLMs and opens a symbolic path for modelinterpretability research.

Lumen AI, Tengzhou No. 1 Middle School, Shihao Ji et al.

Addressing the resource waste caused by fixed computation paradigms in deep learning models under dynamic scenarios, this paper proposes a Transformer$^{-1}$ architecture based on the principle of deep adaptivity. This architecture achieves dynamic matching between input features and computational resources by establishing a joint optimization model for complexity and computation. Our core contributions include: (1) designing a two-layer control mechanism, composed of a complexity predictor and a reinforcement learning policy network, enabling end-to-end optimization of computation paths; (2) deriving a lower bound theory for dynamic computation, proving the system's theoretical reach to optimal efficiency; and (3) proposing a layer folding technique and a CUDA Graph pre-compilation scheme, overcoming the engineering bottlenecks of dynamic architectures. In the ImageNet-1K benchmark test, our method reduces FLOPs by 42.7\% and peak memory usage by 34.1\% compared to the standard Transformer, while maintaining comparable accuracy ($\pm$0.3\%). Furthermore, we conducted practical deployment on the Jetson AGX Xavier platform, verifying the effectiveness and practical value of this method in resource-constrained environments. To further validate the generality of the method, we also conducted experiments on several natural language processing tasks and achieved significant improvements in resource efficiency.

Hui Ye, Xiulong Yang, Martin Takac et al.

The goal of text-to-image synthesis is to generate a visually realistic image that matches a given text description. In practice, the captions annotated by humans for the same image have large variance in terms of contents and the choice of words. The linguistic discrepancy between the captions of the identical image leads to the synthetic images deviating from the ground truth. To address this issue, we propose a contrastive learning approach to improve the quality and enhance the semantic consistency of synthetic images. In the pretraining stage, we utilize the contrastive learning approach to learn the consistent textual representations for the captions corresponding to the same image. Furthermore, in the following stage of GAN training, we employ the contrastive learning method to enhance the consistency between the generated images from the captions related to the same image. We evaluate our approach over two popular text-to-image synthesis models, AttnGAN and DM-GAN, on datasets CUB and COCO, respectively. Experimental results have shown that our approach can effectively improve the quality of synthetic images in terms of three metrics: IS, FID and R-precision. Especially, on the challenging COCO dataset, our approach boosts the FID signifcantly by 29.60% over AttnGAN and by 21.96% over DM-GAN.

Krishanu Sarker, Sharbani Pandit, Anupam Sarker et al.

Effective and reliable screening of patients via Computer-Aided Diagnosis can play a crucial part in the battle against COVID-19. Most of the existing works focus on developing sophisticated methods yielding high detection performance, yet not addressing the issue of predictive uncertainty. In this work, we introduce uncertainty estimation to detect confusing cases for expert referral to address the unreliability of state-of-the-art (SOTA) DNNs on COVID-19 detection. To the best of our knowledge, we are the first to address this issue on the COVID-19 detection problem. In this work, we investigate a number of SOTA uncertainty estimation methods on publicly available COVID dataset and present our experimental findings. In collaboration with medical professionals, we further validate the results to ensure the viability of the best performing method in clinical practice.

Krishanu Sarker, Xiulong Yang, Yang Li et al.

The success of Deep Neural Networks (DNNs) highly depends on data quality. Moreover, predictive uncertainty makes high performing DNNs risky for real-world deployment. In this paper, we aim to address these two issues by proposing a unified filtering framework leveraging underlying data density, that can effectively denoise training data as well as avoid predicting uncertain test data points. Our proposed framework leverages underlying data distribution to differentiate between noise and clean data samples without requiring any modification to existing DNN architectures or loss functions. Extensive experiments on multiple image classification datasets and multiple CNN architectures demonstrate that our simple yet effective framework can outperform the state-of-the-art techniques in denoising training data and abstaining uncertain test data.

Xiulong Yang, Shihao Ji

Adversarial Training (AT) and Virtual Adversarial Training (VAT) are the regularization techniques that train Deep Neural Networks (DNNs) with adversarial examples generated by adding small but worst-case perturbations to input examples. In this paper, we propose xAT and xVAT, new adversarial training algorithms, that generate \textbf{multiplicative} perturbations to input examples for robust training of DNNs. Such perturbations are much more perceptible and interpretable than their \textbf{additive} counterparts exploited by AT and VAT. Furthermore, the multiplicative perturbations can be generated transductively or inductively while the standard AT and VAT only support a transductive implementation. We conduct a series of experiments that analyze the behavior of the multiplicative perturbations and demonstrate that xAT and xVAT match or outperform state-of-the-art classification accuracies across multiple established benchmarks while being about 30\% faster than their additive counterparts. Furthermore, the resulting DNNs also demonstrate distinct weight distributions.

Yang Ye, Shihao Ji

Graph Neural Networks (GNNs) have proved to be an effective representation learning framework for graph-structured data, and have achieved state-of-the-art performance on many practical predictive tasks, such as node classification, link prediction and graph classification. Among the variants of GNNs, Graph Attention Networks (GATs) learn to assign dense attention coefficients over all neighbors of a node for feature aggregation, and improve the performance of many graph learning tasks. However, real-world graphs are often very large and noisy, and GATs are prone to overfitting if not regularized properly. Even worse, the local aggregation mechanism of GATs may fail on disassortative graphs, where nodes within local neighborhood provide more noise than useful information for feature aggregation. In this paper, we propose Sparse Graph Attention Networks (SGATs) that learn sparse attention coefficients under an $L_0$-norm regularization, and the learned sparse attentions are then used for all GNN layers, resulting in an edge-sparsified graph. By doing so, we can identify noisy/task-irrelevant edges, and thus perform feature aggregation on most informative neighbors. Extensive experiments on synthetic and real-world graph learning benchmarks demonstrate the superior performance of SGATs. In particular, SGATs can remove about 50\%-80\% edges from large assortative graphs, while retaining similar classification accuracies. On disassortative graphs, SGATs prune majority of noisy edges and outperform GATs in classification accuracies by significant margins. Furthermore, the removed edges can be interpreted intuitively and quantitatively. To the best of our knowledge, this is the first graph learning algorithm that shows significant redundancies in graphs and edge-sparsified graphs can achieve similar or sometimes higher predictive performances than original graphs.

Azim Ahmadzadeh, Sushant S. Mahajan, Dustin J. Kempton et al.

We use a well-known deep neural network framework, called Mask R-CNN, for identification of solar filaments in full-disk H-alpha images from Big Bear Solar Observatory (BBSO). The image data, collected from BBSO's archive, are integrated with the spatiotemporal metadata of filaments retrieved from the Heliophysics Events Knowledgebase (HEK) system. This integrated data is then treated as the ground-truth in the training process of the model. The available spatial metadata are the output of a currently running filament-detection module developed and maintained by the Feature Finding Team; an international consortium selected by NASA. Despite the known challenges in the identification and characterization of filaments by the existing module, which in turn are inherited into any other module that intends to learn from such outputs, Mask R-CNN shows promising results. Trained and validated on two years worth of BBSO data, this model is then tested on the three following years. Our case-by-case and overall analyses show that Mask R-CNN can clearly compete with the existing module and in some cases even perform better. Several cases of false positives and false negatives, that are correctly segmented by this model are also shown. The overall advantages of using the proposed model are two-fold: First, deep neural networks' performance generally improves as more annotated data, or better annotations are provided. Second, such a model can be scaled up to detect other solar events, as well as a single multi-purpose module. The results presented in this study introduce a proof of concept in benefits of employing deep neural networks for detection of solar events, and in particular, filaments.

Yang Li, Shihao Ji

Neural plasticity is an important functionality of human brain, in which number of neurons and synapses can shrink or expand in response to stimuli throughout the span of life. We model this dynamic learning process as an $L_0$-norm regularized binary optimization problem, in which each unit of a neural network (e.g., weight, neuron or channel, etc.) is attached with a stochastic binary gate, whose parameters determine the level of activity of a unit in the network. At the beginning, only a small portion of binary gates (therefore the corresponding neurons) are activated, while the remaining neurons are in a hibernation mode. As the learning proceeds, some neurons might be activated or deactivated if doing so can be justified by the cost-benefit tradeoff measured by the $L_0$-norm regularized objective. As the training gets mature, the probability of transition between activation and deactivation will diminish until a final hardening stage. We demonstrate that all of these learning dynamics can be modulated by a single parameter $k$ seamlessly. Our neural plasticity network (NPN) can prune or expand a network depending on the initial capacity of network provided by the user; it also unifies dropout (when $k=0$), traditional training of DNNs (when $k=\infty$) and interpolates between these two. To the best of our knowledge, this is the first learning framework that unifies network sparsification and network expansion in an end-to-end training pipeline. Extensive experiments on synthetic dataset and multiple image classification benchmarks demonstrate the superior performance of NPN. We show that both network sparsification and network expansion can yield compact models of similar architectures, while retaining competitive accuracies of the original networks.

Krishanu Sarker, Mohamed Masoud, Saeid Belkasim et al.

Human activity recognition based on video streams has received numerous attentions in recent years. Due to lack of depth information, RGB video based activity recognition performs poorly compared to RGB-D video based solutions. On the other hand, acquiring depth information, inertia etc. is costly and requires special equipment, whereas RGB video streams are available in ordinary cameras. Hence, our goal is to investigate whether similar or even higher accuracy can be achieved with RGB-only modality. In this regard, we propose a novel framework that couples skeleton data extracted from RGB video and deep Bidirectional Long Short Term Memory (BLSTM) model for activity recognition. A big challenge of training such a deep network is the limited training data, and exploring RGB-only stream significantly exaggerates the difficulty. We therefore propose a set of algorithmic techniques to train this model effectively, e.g., data augmentation, dynamic frame dropout and gradient injection. The experiments demonstrate that our RGB-only solution surpasses the state-of-the-art approaches that all exploit RGB-D video streams by a notable margin. This makes our solution widely deployable with ordinary cameras.

Shihao Ji, Nadathur Satish, Sheng Li et al.

Word2vec is a widely used algorithm for extracting low-dimensional vector representations of words. State-of-the-art algorithms including those by Mikolov et al. have been parallelized for multi-core CPU architectures, but are based on vector-vector operations with "Hogwild" updates that are memory-bandwidth intensive and do not efficiently use computational resources. In this paper, we propose "HogBatch" by improving reuse of various data structures in the algorithm through the use of minibatching and negative sample sharing, hence allowing us to express the problem using matrix multiply operations. We also explore different techniques to distribute word2vec computation across nodes in a compute cluster, and demonstrate good strong scalability up to 32 nodes. The new algorithm is particularly suitable for modern multi-core/many-core architectures, especially Intel's latest Knights Landing processors, and allows us to scale up the computation near linearly across cores and nodes, and process hundreds of millions of words per second, which is the fastest word2vec implementation to the best of our knowledge.

Jiong Zhang, Parameswaran Raman, Shihao Ji et al.

Stochastic variational inference (SVI), the state-of-the-art algorithm for scaling variational inference to large-datasets, is inherently serial. Moreover, it requires the parameters to fit in the memory of a single processor; this is problematic when the number of parameters is in billions. In this paper, we propose extreme stochastic variational inference (ESVI), an asynchronous and lock-free algorithm to perform variational inference for mixture models on massive real world datasets. ESVI overcomes the limitations of SVI by requiring that each processor only access a subset of the data and a subset of the parameters, thus providing data and model parallelism simultaneously. We demonstrate the effectiveness of ESVI by running Latent Dirichlet Allocation (LDA) on UMBC-3B, a dataset that has a vocabulary of 3 million and a token size of 3 billion. In our experiments, we found that ESVI not only outperforms VI and SVI in wallclock-time, but also achieves a better quality solution. In addition, we propose a strategy to speed up computation and save memory when fitting large number of topics.

Shihao Ji, Nadathur Satish, Sheng Li et al.

Word2Vec is a widely used algorithm for extracting low-dimensional vector representations of words. It generated considerable excitement in the machine learning and natural language processing (NLP) communities recently due to its exceptional performance in many NLP applications such as named entity recognition, sentiment analysis, machine translation and question answering. State-of-the-art algorithms including those by Mikolov et al. have been parallelized for multi-core CPU architectures but are based on vector-vector operations that are memory-bandwidth intensive and do not efficiently use computational resources. In this paper, we improve reuse of various data structures in the algorithm through the use of minibatching, hence allowing us to express the problem using matrix multiply operations. We also explore different techniques to distribute word2vec computation across nodes in a compute cluster, and demonstrate good strong scalability up to 32 nodes. In combination, these techniques allow us to scale up the computation near linearly across cores and nodes, and process hundreds of millions of words per second, which is the fastest word2vec implementation to the best of our knowledge.

Shihao Ji, S. V. N. Vishwanathan, Nadathur Satish et al.

We propose BlackOut, an approximation algorithm to efficiently train massive recurrent neural network language models (RNNLMs) with million word vocabularies. BlackOut is motivated by using a discriminative loss, and we describe a new sampling strategy which significantly reduces computation while improving stability, sample efficiency, and rate of convergence. One way to understand BlackOut is to view it as an extension of the DropOut strategy to the output layer, wherein we use a discriminative training loss and a weighted sampling scheme. We also establish close connections between BlackOut, importance sampling, and noise contrastive estimation (NCE). Our experiments, on the recently released one billion word language modeling benchmark, demonstrate scalability and accuracy of BlackOut; we outperform the state-of-the art, and achieve the lowest perplexity scores on this dataset. Moreover, unlike other established methods which typically require GPUs or CPU clusters, we show that a carefully implemented version of BlackOut requires only 1-10 days on a single machine to train a RNNLM with a million word vocabulary and billions of parameters on one billion words. Although we describe BlackOut in the context of RNNLM training, it can be used to any networks with large softmax output layers.