Yang Zhong

Yang Zhong, Hongyu Yu, Mao Su et al.

Using the message-passing mechanism in machine learning (ML) instead of self-consistent iterations to directly build the mapping from structures to electronic Hamiltonian matrices will greatly improve the efficiency of density functional theory (DFT) calculations. In this work, we proposed a general analytic Hamiltonian representation in an E(3) equivariant framework, which can fit the ab initio Hamiltonian of molecules and solids by a complete data-driven method and are equivariant under rotation, space inversion, and time reversal operations. Our model reached state-of-the-art precision in the benchmark test and accurately predicted the electronic Hamiltonian matrices and related properties of various periodic and aperiodic systems, showing high transferability and generalization ability. This framework provides a general transferable model that can be used to accelerate the electronic structure calculations on different large systems with the same network weights trained on small structures.

Mohamed Elaraby, Yang Zhong, Diane Litman

We propose a simple approach for the abstractive summarization of long legal opinions that considers the argument structure of the document. Legal opinions often contain complex and nuanced argumentation, making it challenging to generate a concise summary that accurately captures the main points of the legal opinion. Our approach involves using argument role information to generate multiple candidate summaries, then reranking these candidates based on alignment with the document's argument structure. We demonstrate the effectiveness of our approach on a dataset of long legal opinions and show that it outperforms several strong baselines.

Yang Zhong, Diane Litman

Though many algorithms can be used to automatically summarize legal case decisions, most fail to incorporate domain knowledge about how important sentences in a legal decision relate to a representation of its document structure. For example, analysis of a legal case summarization dataset demonstrates that sentences serving different types of argumentative roles in the decision appear in different sections of the document. In this work, we propose an unsupervised graph-based ranking model that uses a reweighting algorithm to exploit properties of the document structure of legal case decisions. We also explore the impact of using different methods to compute the document structure. Results on the Canadian Legal Case Law dataset show that our proposed method outperforms several strong baselines.

Hongyu Yu, Yang Zhong, Liangliang Hong et al.

The development of machine learning interatomic potentials has immensely contributed to the accuracy of simulations of molecules and crystals. However, creating interatomic potentials for magnetic systems that account for both magnetic moments and structural degrees of freedom remains a challenge. This work introduces SpinGNN, a spin-dependent interatomic potential approach that employs the graph neural network (GNN) to describe magnetic systems. SpinGNN consists of two types of edge GNNs: Heisenberg edge GNN (HEGNN) and spin-distance edge GNN (SEGNN). HEGNN is tailored to capture Heisenberg-type spin-lattice interactions, while SEGNN accurately models multi-body and high-order spin-lattice coupling. The effectiveness of SpinGNN is demonstrated by its exceptional precision in fitting a high-order spin Hamiltonian and two complex spin-lattice Hamiltonians with great precision. Furthermore, it successfully models the subtle spin-lattice coupling in BiFeO3 and performs large-scale spin-lattice dynamics simulations, predicting its antiferromagnetic ground state, magnetic phase transition, and domain wall energy landscape with high accuracy. Our study broadens the scope of graph neural network potentials to magnetic systems, serving as a foundation for carrying out large-scale spin-lattice dynamic simulations of such systems.

Mert İnan, Yang Zhong, Sabit Hassan et al.

End-to-end sign language generation models do not accurately represent the prosody in sign language. A lack of temporal and spatial variations leads to poor-quality generated presentations that confuse human interpreters. In this paper, we aim to improve the prosody in generated sign languages by modeling intensification in a data-driven manner. We present different strategies grounded in linguistics of sign language that inform how intensity modifiers can be represented in gloss annotations. To employ our strategies, we first annotate a subset of the benchmark PHOENIX-14T, a German Sign Language dataset, with different levels of intensification. We then use a supervised intensity tagger to extend the annotated dataset and obtain labels for the remaining portion of it. This enhanced dataset is then used to train state-of-the-art transformer models for sign language generation. We find that our efforts in intensification modeling yield better results when evaluated with automatic metrics. Human evaluation also indicates a higher preference of the videos generated using our model.

Hongyu Yu, Boyu Liu, Yang Zhong et al.

This study introduces time-reversal E(3)-equivariant neural network and SpinGNN++ framework for constructing a comprehensive interatomic potential for magnetic systems, encompassing spin-orbit coupling and noncollinear magnetic moments. SpinGNN++ integrates multitask spin equivariant neural network with explicit spin-lattice terms, including Heisenberg, Dzyaloshinskii-Moriya, Kitaev, single-ion anisotropy, and biquadratic interactions, and employs time-reversal equivariant neural network to learn high-order spin-lattice interactions using time-reversal E(3)-equivariant convolutions. To validate SpinGNN++, a complex magnetic model dataset is introduced as a benchmark and employed to demonstrate its capabilities. SpinGNN++ provides accurate descriptions of the complex spin-lattice coupling in monolayer CrI$_3$ and CrTe$_2$, achieving sub-meV errors. Importantly, it facilitates large-scale parallel spin-lattice dynamics, thereby enabling the exploration of associated properties, including the magnetic ground state and phase transition. Remarkably, SpinGNN++ identifies a new ferrimagnetic state as the ground magnetic state for monolayer CrTe2, thereby enriching its phase diagram and providing deeper insights into the distinct magnetic signals observed in various experiments.

Zhexiong Liu, Mohamed Elaraby, Yang Zhong et al.

This paper presents an overview of the ImageArg shared task, the first multimodal Argument Mining shared task co-located with the 10th Workshop on Argument Mining at EMNLP 2023. The shared task comprises two classification subtasks - (1) Subtask-A: Argument Stance Classification; (2) Subtask-B: Image Persuasiveness Classification. The former determines the stance of a tweet containing an image and a piece of text toward a controversial topic (e.g., gun control and abortion). The latter determines whether the image makes the tweet text more persuasive. The shared task received 31 submissions for Subtask-A and 21 submissions for Subtask-B from 9 different teams across 6 countries. The top submission in Subtask-A achieved an F1-score of 0.8647 while the best submission in Subtask-B achieved an F1-score of 0.5561.

Kangan Qian, ChuChu Xie, Yang Zhong et al.

Vision-Language-Action (VLA) models drive next-generation autonomous systems, but training them requires scalable, high-quality annotations from complex environments. Current cloud pipelines rely on generic vision-language models (VLMs) that lack geometric reasoning and domain semantics due to their 2D image-text pretraining. To address this mismatch, we propose XEmbodied, a cloud-side foundation model that endows VLMs with intrinsic 3D geometric awareness and interaction with physical cues (e.g., occupancy grids, 3D boxes). Instead of treating geometry as auxiliary input, XEmbodied integrates geometric representations via a structured 3D Adapter and distills physical signals into context tokens using an Efficient Image-Embodied Adapter. Through progressive domain curriculum and reinforcement learning post-training, XEmbodied preserves general capabilities while demonstrating robust performance across 18 public benchmarks. It significantly improves spatial reasoning, traffic semantics, embodied affordance, and out-of-distribution generalization for large-scale scenario mining and embodied VQA.

Kangan Qian, Sicong Jiang, Yang Zhong et al. · tsinghua

Vision-Language Models (VLMs) show promise for autonomous driving, yet their struggle with hallucinations, inefficient reasoning, and limited real-world validation hinders accurate perception and robust step-by-step reasoning. To overcome this, we introduce \textbf{AgentThink}, a pioneering unified framework that integrates Chain-of-Thought (CoT) reasoning with dynamic, agent-style tool invocation for autonomous driving tasks. AgentThink's core innovations include: \textbf{(i) Structured Data Generation}, which establishes an autonomous driving tool library to automatically construct structured, self-verified reasoning data explicitly incorporating tool usage for diverse driving scenarios; \textbf{(ii) A Two-stage Training Pipeline}, employing Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO) to equip VLMs with the capability for autonomous tool invocation; and \textbf{(iii) Agent-style Tool-Usage Evaluation}, introducing a novel multi-tool assessment protocol to rigorously evaluate the model's tool invocation and utilization. Experiments on the DriveLMM-o1 benchmark demonstrate that AgentThink significantly boosts overall reasoning scores by \textbf{53.91%} and enhances answer accuracy by \textbf{33.54%}, while markedly improving reasoning quality and consistency. Furthermore, ablation studies and robust zero-shot/few-shot generalization experiments across various benchmarks underscore its powerful capabilities. These findings highlight a promising trajectory for developing trustworthy and tool-aware autonomous driving models. Code is available at https://github.com/curryqka/AgentThink.

Sicong Jiang, Zilin Huang, Kangan Qian et al.

The rapid progress of multimodal large language models (MLLM) has paved the way for Vision-Language-Action (VLA) paradigms, which integrate visual perception, natural language understanding, and control within a single policy. Researchers in autonomous driving are actively adapting these methods to the vehicle domain. Such models promise autonomous vehicles that can interpret high-level instructions, reason about complex traffic scenes, and make their own decisions. However, the literature remains fragmented and is rapidly expanding. This survey offers the first comprehensive overview of VLA for Autonomous Driving (VLA4AD). We (i) formalize the architectural building blocks shared across recent work, (ii) trace the evolution from early explainer to reasoning-centric VLA models, and (iii) compare over 20 representative models according to VLA's progress in the autonomous driving domain. We also consolidate existing datasets and benchmarks, highlighting protocols that jointly measure driving safety, accuracy, and explanation quality. Finally, we detail open challenges - robustness, real-time efficiency, and formal verification - and outline future directions of VLA4AD. This survey provides a concise yet complete reference for advancing interpretable socially aligned autonomous vehicles. Github repo is available at \href{https://github.com/JohnsonJiang1996/Awesome-VLA4AD}{SicongJiang/Awesome-VLA4AD}.

Yang Zhong, Xiwen Li, Xingao Gong et al.

Machine-learning electronic Hamiltonians achieve orders-of-magnitude speedups over density-functional theory, yet current models omit long-range Coulomb interactions that govern physics in polar crystals and heterostructures. We derive closed-form long-range Hamiltonian matrix elements in a nonorthogonal atomic-orbital basis through variational decomposition of the electrostatic energy, deriving a variationally consistent mapping from the electron density matrix to effective atomic charges. We implement this framework in HamGNN-LR, a dual-channel architecture combining E(3)-equivariant message passing with reciprocal-space Ewald summation. Benchmarks demonstrate that physics-based long-range corrections are essential: purely data-driven attention mechanisms fail to capture macroscopic electrostatic potentials. Benchmarks on polar ZnO slabs, CdSe/ZnS heterostructures, and GaN/AlN superlattices show two- to threefold error reductions and robust transferability to systems far beyond training sizes, eliminating the characteristic staircase artifacts that plague short-range models in the presence of built-in electric fields.

Yang Zhong, Diane Litman

We propose an approach for the structure controllable summarization of long legal opinions that considers the argument structure of the document. Our approach involves using predicted argument role information to guide the model in generating coherent summaries that follow a provided structure pattern. We demonstrate the effectiveness of our approach on a dataset of legal opinions and show that it outperforms several strong baselines with respect to ROUGE, BERTScore, and structure similarity.

Yang Zhong, Hongyu Yu, Jihui Yang et al.

While density functional theory (DFT) serves as a prevalent computational approach in electronic structure calculations, its computational demands and scalability limitations persist. Recently, leveraging neural networks to parameterize the Kohn-Sham DFT Hamiltonian has emerged as a promising avenue for accelerating electronic structure computations. Despite advancements, challenges such as the necessity for computing extensive DFT training data to explore each new system and the complexity of establishing accurate ML models for multi-elemental materials still exist. Addressing these hurdles, this study introduces a universal electronic Hamiltonian model trained on Hamiltonian matrices obtained from first-principles DFT calculations of nearly all crystal structures on the Materials Project. We demonstrate its generality in predicting electronic structures across the whole periodic table, including complex multi-elemental systems, solid-state electrolytes, Moiré twisted bilayer heterostructure, and metal-organic frameworks (MOFs). Moreover, we utilize the universal model to conduct high-throughput calculations of electronic structures for crystals in GeNOME datasets, identifying 3,940 crystals with direct band gaps and 5,109 crystals with flat bands. By offering a reliable efficient framework for computing electronic properties, this universal Hamiltonian model lays the groundwork for advancements in diverse fields, such as easily providing a huge data set of electronic structures and also making the materials design across the whole periodic table possible.

Tianyou Jiang, Yang Zhong

You Look Only Once (YOLO) models have been widely used for building real-time object detectors across various domains. With the increasing frequency of new YOLO versions being released, key questions arise. Are the newer versions always better than their previous versions? What are the core innovations in each YOLO version and how do these changes translate into real-world performance gains? In this paper, we summarize the key innovations from YOLOv1 to YOLOv11, introduce a comprehensive benchmark called ODverse33, which includes 33 datasets spanning 11 diverse domains (Autonomous driving, Agricultural, Underwater, Medical, Videogame, Industrial, Aerial, Wildlife, Retail, Microscopic, and Security), and explore the practical impact of model improvements in real-world, multi-domain applications through extensive experimental results. We hope this study can provide some guidance to the extensive users of object detection models and give some references for future real-time object detector development.

Yang Zhong, Mohamed Elaraby, Diane Litman et al.

This paper introduces ReflectSumm, a novel summarization dataset specifically designed for summarizing students' reflective writing. The goal of ReflectSumm is to facilitate developing and evaluating novel summarization techniques tailored to real-world scenarios with little training data, %practical tasks with potential implications in the opinion summarization domain in general and the educational domain in particular. The dataset encompasses a diverse range of summarization tasks and includes comprehensive metadata, enabling the exploration of various research questions and supporting different applications. To showcase its utility, we conducted extensive evaluations using multiple state-of-the-art baselines. The results provide benchmarks for facilitating further research in this area.

Siwen Jiao, Kangan Qian, Hao Ye et al.

Autonomous driving faces significant challenges in achieving human-like iterative decision-making, which continuously generates, evaluates, and refines trajectory proposals. Current generation-evaluation frameworks isolate trajectory generation from quality assessment, preventing iterative refinement essential for planning, while reinforcement learning methods collapse multi-dimensional preferences into scalar rewards, obscuring critical trade-offs and yielding scalarization bias.To overcome these issues, we present EvaDrive, a novel multi-objective reinforcement learning framework that establishes genuine closed-loop co-evolution between trajectory generation and evaluation via adversarial optimization. EvaDrive frames trajectory planning as a multi-round adversarial game. In this game, a hierarchical generator continuously proposes candidate paths by combining autoregressive intent modeling for temporal causality with diffusion-based refinement for spatial flexibility. These proposals are then rigorously assessed by a trainable multi-objective critic that explicitly preserves diverse preference structures without collapsing them into a single scalarization bias.This adversarial interplay, guided by a Pareto frontier selection mechanism, enables iterative multi-round refinement, effectively escaping local optima while preserving trajectory diversity.Extensive experiments on NAVSIM and Bench2Drive benchmarks demonstrate SOTA performance, achieving 94.9 PDMS on NAVSIM v1 (surpassing DiffusionDrive by 6.8, DriveSuprim by 5.0, and TrajHF by 0.9) and 64.96 Driving Score on Bench2Drive. EvaDrive generates diverse driving styles via dynamic weighting without external preference data, introducing a closed-loop adversarial framework for human-like iterative decision-making, offering a novel scalarization-free trajectory optimization approach.

Yang Zhong, Diane Litman

Detecting factual inconsistency for long document summarization remains challenging, given the complex structure of the source article and long summary length. In this work, we study factual inconsistency errors and connect them with a line of discourse analysis. We find that errors are more common in complex sentences and are associated with several discourse features. We propose a framework that decomposes long texts into discourse-inspired chunks and utilizes discourse information to better aggregate sentence-level scores predicted by natural language inference models. Our approach shows improved performance on top of different model baselines over several evaluation benchmarks, covering rich domains of texts, focusing on long document summarization. This underscores the significance of incorporating discourse features in developing models for scoring summaries for long document factual inconsistency.

Yang Zhong, Yifan Yao, Tong Luo et al.

Food analysis is becoming a hot topic in health area, in which fine-grained food recognition task plays an important role. In this paper, we describe the details of our solution to the LargeFineFoodAI-ICCV Workshop-Recognition challenge held on Kaggle. We find a proper combination of Arcface loss[1] and Circle loss[9] can bring improvement to the performance. With Arcface and the combined loss, model was trained with carefully tuned configurations and ensembled to get the final results. Our solution won the 3rd place in the competition.

Yang Zhong, Zhiming Wang, Zhaoyang Li et al.

This paper introduces the 3rd place solution to the ICCV LargeFineFoodAI Retrieval Competition on Kaggle. Four basic models are independently trained with the weighted sum of ArcFace and Circle loss, then TTA and Ensemble are successively applied to improve feature representation ability. In addition, a new reranking method for retrieval is proposed based on diffusion and k-reciprocal reranking. Finally, our method scored 0.81219 and 0.81191 mAP@100 on the public and private leaderboard, respectively.

Bohan Yang, Gang Liu, Yang Zhong et al.

Proton arc therapy (PAT) is an emerging and promising modality in radiotherapy, offering improved dose distribution and treatment robustness over intensity-modulated proton therapy. Yet, identifying the optimal energy layer (EL) sequence remains challenging due to the intensive computational demand and prolonged treatment delivery time. This study proposes an unsupervised deep learning model for fast EL pre-selection that minimizes EL switch (ELS) time while maintaining high plan quality. We introduce a novel data representation method, spot-count representation, which encodes the number of proton spots intersecting the target and organs at risk (OAR) in a matrix structured by sorted gantry angles and energy layers. This representation serves as the input of an U-Net style architecture, SPArc_dl, which is trained using a tri-objective function: maximizing spot-counts on target, minimizing spot-counts on OAR, and reducing ELS time. The model is evaluated on 35 nasopharyngeal cancer cases, and its performance is compared to SPArc_particle_swarm (SPArc_ps). SPArc_dl produces EL pre-selection that significantly improves both plan quality and delivery efficiency. Compared to SPArc_ps, it enhances the conformity index by 0.1 (p<0.01), reduces the homogeneity index by 0.71 (p<0.01), lowers the brainstem mean dose by 0.25 (p<0.01), and shortens the ELS time by 37.2% (p < 0.01). The results unintentionally reveal employing unchanged ELS is more time-wise efficient than descended ELS. SPArc_dl's inference time is within 1 second. However, SPArc_dl plan demonstrates limitation in robustness. The proposed spot-count representation lays a foundation for incorporating unsupervised deep learning approaches into EL pre-selection task. SPArc_dl is a fast tool for generating high-quality PAT plans by strategically pre-selecting EL to reduce delivery time while maintaining excellent dosimetric performance.

Yang Zhong, Jiangang Hao, Michael Fauss et al.

The rapid advancement of large language models (LLMs) has enabled the generation of coherent essays, making AI-assisted writing increasingly common in educational and professional settings. Using large-scale empirical data, we examine and benchmark the characteristics and quality of essays generated by popular LLMs and discuss their implications for two key components of writing assessments: automated scoring and academic integrity. Our findings highlight limitations in existing automated scoring systems, such as e-rater, when applied to essays generated or heavily influenced by AI, and identify areas for improvement, including the development of new features to capture deeper thinking and recalibrating feature weights. Despite growing concerns that the increasing variety of LLMs may undermine the feasibility of detecting AI-generated essays, our results show that detectors trained on essays generated from one model can often identify texts from others with high accuracy, suggesting that effective detection could remain manageable in practice.

Yang Zhong, Weiping Dou, Andrew Cohen et al.

To extend the antenna design on printed circuit boards (PCBs) for more engineers of interest, we propose a simple method that models PCB antennas with a few basic components. By taking two separate steps to decide their geometric dimensions and positions, antenna prototypes can be facilitated with no experience required. Random sampling statistics relate to the quality of dimensions are used in selecting among dimension candidates. A novel image-based classifier using a convolutional neural network (CNN) is introduced to further determine the positions of these fixed-dimension components. Two examples from wearable products have been chosen to examine the entire workflow. Their final designs are realistic and their performance metrics are not inferior to the ones designed by experienced engineers.

Yang Zhong, Peter Renner, Weiping Dou et al.

To facilitate the antenna design with the aid of computer, one of the practices in consumer electronic industry is to model and optimize antenna performances with a simplified antenna geometric scheme. Traditional antenna modeling requires profound prior knowledge of electromagnetics in order to achieve a good design which satisfies the performance specifications from both antenna and product designs. The ease of handling multidimensional optimization problems and the less dependence on domain knowledge and experience are the key to achieve the popularity of simulation driven antenna design and optimization for the industry. In this paper, we introduce a flexible geometric scheme with the concept of mesh network that can form any arbitrary shape by connecting different nodes. For such problems with high dimensional parameters, we propose a machine learning based generative method to assist the searching of optimal solutions. It consists of discriminators and generators. The discriminators are used to predict the performance of geometric models, and the generators to create new candidates that will pass the discriminators. Moreover, an evolutionary criterion approach is proposed for further improving the efficiency of our method. Finally, not only optimal solutions can be found, but also the well trained generators can be used to automate future antenna design and optimization. For a dual resonance antenna design with wide bandwidth, our proposed method is in par with Trust Region Framework and much better than the other mature machine learning algorithms including the widely used Genetic Algorithm and Particle Swarm Optimization. When there is no wide bandwidth requirement, it is better than Trust Region Framework.

Yang Zhong, Hongyu Yu, Xingao Gong et al.

Message-passing neural networks (MPNN) have shown extremely high efficiency and accuracy in predicting the physical properties of molecules and crystals, and are expected to become the next-generation material simulation tool after the density functional theory (DFT). However, there is currently a lack of a general MPNN framework for directly predicting the tensor properties of the crystals. In this work, a general framework for the prediction of tensor properties was proposed: the tensor property of a crystal can be decomposed into the average of the tensor contributions of all the atoms in the crystal, and the tensor contribution of each atom can be expanded as the sum of the tensor projections in the directions of the edges connecting the atoms. On this basis, the edge-based expansions of force vectors, Born effective charges (BECs), dielectric (DL) and piezoelectric (PZ) tensors were proposed. These expansions are rotationally equivariant, while the coefficients in these tensor expansions are rotationally invariant scalars which are similar to physical quantities such as formation energy and band gap. The advantage of this tensor prediction framework is that it does not require the network itself to be equivariant. Therefore, in this work, we directly designed the edge-based tensor prediction graph neural network (ETGNN) model on the basis of the invariant graph neural network to predict tensors. The validity and high precision of this tensor prediction framework were shown by the tests of ETGNN on the extended systems, random perturbed structures and JARVIS-DFT datasets. This tensor prediction framework is general for nearly all the GNNs and can achieve higher accuracy with more advanced GNNs in the future.

Chao Jiang, Mounica Maddela, Wuwei Lan et al.

The success of a text simplification system heavily depends on the quality and quantity of complex-simple sentence pairs in the training corpus, which are extracted by aligning sentences between parallel articles. To evaluate and improve sentence alignment quality, we create two manually annotated sentence-aligned datasets from two commonly used text simplification corpora, Newsela and Wikipedia. We propose a novel neural CRF alignment model which not only leverages the sequential nature of sentences in parallel documents but also utilizes a neural sentence pair model to capture semantic similarity. Experiments demonstrate that our proposed approach outperforms all the previous work on monolingual sentence alignment task by more than 5 points in F1. We apply our CRF aligner to construct two new text simplification datasets, Newsela-Auto and Wiki-Auto, which are much larger and of better quality compared to the existing datasets. A Transformer-based seq2seq model trained on our datasets establishes a new state-of-the-art for text simplification in both automatic and human evaluation.

Yang Zhong, Chao Jiang, Wei Xu et al.

This paper presents a data-driven study focusing on analyzing and predicting sentence deletion -- a prevalent but understudied phenomenon in document simplification -- on a large English text simplification corpus. We inspect various document and discourse factors associated with sentence deletion, using a new manually annotated sentence alignment corpus we collected. We reveal that professional editors utilize different strategies to meet readability standards of elementary and middle schools. To predict whether a sentence will be deleted during simplification to a certain level, we harness automatically aligned data to train a classification model. Evaluated on our manually annotated data, our best models reached F1 scores of 65.2 and 59.7 for this task at the levels of elementary and middle school, respectively. We find that discourse level factors contribute to the challenging task of predicting sentence deletion for simplification.

Yang Zhong, Atsuto Maki

This paper is about regularizing deep convolutional networks (CNNs) based on an adaptive framework for transfer learning with limited training data in the target domain. Recent advances of CNN regularization in this context are commonly due to the use of additional regularization objectives. They guide the training away from the target task using some forms of concrete tasks. Unlike those related approaches, we suggest that an objective without a concrete goal can still serve well as a regularized. In particular, we demonstrate Pseudo-task Regularization (PtR) which dynamically regularizes a network by simply attempting to regress image representations to pseudo-regression targets during fine-tuning. That is, a CNN is efficiently regularized without additional resources of data or prior domain expertise. In sum, the proposed PtR provides: a) an alternative for network regularization without dependence on the design of concrete regularization objectives or extra annotations; b) a dynamically adjusted and maintained strength of regularization effect by balancing the gradient norms between objectives on-line. Through numerous experiments, surprisingly, the improvements on classification accuracy by PtR are shown greater or on a par to the recent state-of-the-art methods.

Yang Zhong, Vladimir Li, Ryuzo Okada et al.

This paper presents an automatic network adaptation method that finds a ConvNet structure well-suited to a given target task, e.g., image classification, for efficiency as well as accuracy in transfer learning. We call the concept target-aware transfer learning. Given only small-scale labeled data, and starting from an ImageNet pre-trained network, we exploit a scheme of removing its potential redundancy for the target task through iterative operations of filter-wise pruning and network optimization. The basic motivation is that compact networks are on one hand more efficient and should also be more tolerant, being less complex, against the risk of overfitting which would hinder the generalization of learned representations in the context of transfer learning. Further, unlike existing methods involving network simplification, we also let the scheme identify redundant portions across the entire network, which automatically results in a network structure adapted to the task at hand. We achieve this with a few novel ideas: (i) cumulative sum of activation statistics for each layer, and (ii) a priority evaluation of pruning across multiple layers. Experimental results by the method on five datasets (Flower102, CUB200-2011, Dog120, MIT67, and Stanford40) show favorable accuracies over the related state-of-the-art techniques while enhancing the computational and storage efficiency of the transferred model.

Yang Zhong, Josephine Sullivan, Haibo Li

Predicting attributes from face images in the wild is a challenging computer vision problem. To automatically describe face attributes from face containing images, traditionally one needs to cascade three technical blocks --- face localization, facial descriptor construction, and attribute classification --- in a pipeline. As a typical classification problem, face attribute prediction has been addressed using deep learning. Current state-of-the-art performance was achieved by using two cascaded Convolutional Neural Networks (CNNs), which were specifically trained to learn face localization and attribute description. In this paper, we experiment with an alternative way of employing the power of deep representations from CNNs. Combining with conventional face localization techniques, we use off-the-shelf architectures trained for face recognition to build facial descriptors. Recognizing that the describable face attributes are diverse, our face descriptors are constructed from different levels of the CNNs for different attributes to best facilitate face attribute prediction. Experiments on two large datasets, LFWA and CelebA, show that our approach is entirely comparable to the state-of-the-art. Our findings not only demonstrate an efficient face attribute prediction approach, but also raise an important question: how to leverage the power of off-the-shelf CNN representations for novel tasks.

Yang Zhong, Josephine Sullivan, Haibo Li

Predicting facial attributes from faces in the wild is very challenging due to pose and lighting variations in the real world. The key to this problem is to build proper feature representations to cope with these unfavourable conditions. Given the success of Convolutional Neural Network (CNN) in image classification, the high-level CNN feature, as an intuitive and reasonable choice, has been widely utilized for this problem. In this paper, however, we consider the mid-level CNN features as an alternative to the high-level ones for attribute prediction. This is based on the observation that face attributes are different: some of them are locally oriented while others are globally defined. Our investigations reveal that the mid-level deep representations outperform the prediction accuracy achieved by the (fine-tuned) high-level abstractions. We empirically demonstrate that the midlevel representations achieve state-of-the-art prediction performance on CelebA and LFWA datasets. Our investigations also show that by utilizing the mid-level representations one can employ a single deep network to achieve both face recognition and attribute prediction.

Yang Zhong, Haibo Li

Different from face verification, face identification is much more demanding. To reach comparable performance, an identifier needs to be roughly N times better than a verifier. To expect a breakthrough in face identification, we need a fresh look at the fundamental building blocks of face recognition. In this paper we focus on the selection of a suitable signal representation and better matching strategy for face identification. We demonstrate how Gabor phase could be leveraged to improve the performance of face identification by using the Block Matching method. Compared to the existing approaches, the proposed method features much lower algorithmic complexity: face images are only filtered by a single-scale Gabor filter pair and the matching is performed between any pairs of face images at hand without involving any training process. Benchmark evaluations show that the proposed approach is totally comparable to and even better than state-of-the-art algorithms, which are typically based on more features extracted from a large set of Gabor faces and/or rely on heavy training processes.