arXiv subjects

Qi Zhao

Publications and source records attributed to Qi Zhao.

At least 217 records · Page 12Linked to original sources

arXiv2020

$n$-Reference Transfer Learning for Saliency Prediction

Benefiting from deep learning research and large-scale datasets, saliency prediction has achieved significant success in the past decade. However, it still remains challenging to predict saliency maps on images in new domains that lack sufficient data for data-hungry models. To solve this problem, we propose a few-shot transfer learning paradigm for saliency prediction, which enables efficient transfer of knowledge learned from the existing large-scale saliency datasets to a target domain with limited labeled examples. Specifically, very few target domain examples are used as the reference to train a model with a source domain dataset such that the training process can converge to a local minimum in favor of the target domain. Then, the learned model is further fine-tuned with the reference. The proposed framework is gradient-based and model-agnostic. We conduct comprehensive experiments and ablation study on various source domain and target domain pairs. The results show that the proposed framework achieves a significant performance improvement. The code is publicly available at \url{https://github.com/luoyan407/n-reference}.

cs.CV↗

arXiv2020

Universal and Operational Benchmarking of Quantum Memories

Quantum memory -- the capacity to store and faithfully recover unknown quantum states -- is essential for quantum-enhanced technology. There is thus a pressing need for operationally meaningful means to benchmark candidate memories across diverse physical platforms. Here we introduce a universal benchmark distinguished by its relevance across multiple key operational settings, exactly quantifying (1) the memory's robustness to noise, (2) the number of noiseless qubits needed for its synthesis, (3) its potential to speed up statistical sampling tasks, and (4) performance advantage in non-local games beyond classical limits. The measure is analytically computable for low-dimensional systems and can be efficiently bounded in experiment without tomography. We thus illustrate quantum memory as a meaningful resource, with our benchmark reflecting both its cost of creation and what it can accomplish. We demonstrate the benchmark on the five-qubit IBM Q hardware, and apply it to witness efficacy of error-suppression techniques and quantify non-Markovian noise. We thus present an experimentally accessible, practically meaningful, and universally relevant quantifier of a memory's capability to preserve quantum advantage.

quant-ph↗

arXiv2020

Active Learning for Skewed Data Sets

Consider a sequential active learning problem where, at each round, an agent selects a batch of unlabeled data points, queries their labels and updates a binary classifier. While there exists a rich body of work on active learning in this general form, in this paper, we focus on problems with two distinguishing characteristics: severe class imbalance (skew) and small amounts of initial training data. Both of these problems occur with surprising frequency in many web applications. For instance, detecting offensive or sensitive content in online communities (pornography, violence, and hate-speech) is receiving enormous attention from industry as well as research communities. Such problems have both the characteristics we describe -- a vast majority of content is not offensive, so the number of positive examples for such content is orders of magnitude smaller than the negative examples. Furthermore, there is usually only a small amount of initial training data available when building machine-learned models to solve such problems. To address both these issues, we propose a hybrid active learning algorithm (HAL) that balances exploiting the knowledge available through the currently labeled training examples with exploring the large amount of unlabeled data available. Through simulation results, we show that HAL makes significantly better choices for what points to label when compared to strong baselines like margin-sampling. Classifiers trained on the examples selected for labeling by HAL easily out-perform the baselines on target metrics (like area under the precision-recall curve) given the same budget for labeling examples. We believe HAL offers a simple, intuitive, and computationally tractable way to structure active learning for a wide range of machine learning applications.

cs.LG↗

arXiv2020

Video Storytelling: Textual Summaries for Events

Bridging vision and natural language is a longstanding goal in computer vision and multimedia research. While earlier works focus on generating a single-sentence description for visual content, recent works have studied paragraph generation. In this work, we introduce the problem of video storytelling, which aims at generating coherent and succinct stories for long videos. Video storytelling introduces new challenges, mainly due to the diversity of the story and the length and complexity of the video. We propose novel methods to address the challenges. First, we propose a context-aware framework for multimodal embedding learning, where we design a Residual Bidirectional Recurrent Neural Network to leverage contextual information from past and future. Second, we propose a Narrator model to discover the underlying storyline. The Narrator is formulated as a reinforcement learning agent which is trained by directly optimizing the textual metric of the generated story. We evaluate our method on the Video Story dataset, a new dataset that we have collected to enable the study. We compare our method with multiple state-of-the-art baselines, and show that our method achieves better performance, in terms of quantitative measures and user study.

cs.MM↗

arXiv2020

GradMix: Multi-source Transfer across Domains and Tasks

The computer vision community is witnessing an unprecedented rate of new tasks being proposed and addressed, thanks to the deep convolutional networks' capability to find complex mappings from X to Y. The advent of each task often accompanies the release of a large-scale annotated dataset, for supervised training of deep network. However, it is expensive and time-consuming to manually label sufficient amount of training data. Therefore, it is important to develop algorithms that can leverage off-the-shelf labeled dataset to learn useful knowledge for the target task. While previous works mostly focus on transfer learning from a single source, we study multi-source transfer across domains and tasks (MS-DTT), in a semi-supervised setting. We propose GradMix, a model-agnostic method applicable to any model trained with gradient-based learning rule, to transfer knowledge via gradient descent by weighting and mixing the gradients from all sources during training. GradMix follows a meta-learning objective, which assigns layer-wise weights to the source gradients, such that the combined gradient follows the direction that minimize the loss for a small set of samples from the target dataset. In addition, we propose to adaptively adjust the learning rate for each mini-batch based on its importance to the target task, and a pseudo-labeling method to leverage the unlabeled samples in the target domain. We conduct MS-DTT experiments on two tasks: digit recognition and action recognition, and demonstrate the advantageous performance of the proposed method against multiple baselines.

cs.CV↗

arXiv2020

Constructing Multipartite Bell inequalities from stabilizers

Bell inequality with self-testing property has played an important role in quantum information field with both fundamental and practical applications. However, it is generally challenging to find Bell inequalities with self-testing property for multipartite states and actually there are not many known candidates. In this work, we propose a systematical framework to construct Bell inequalities from stabilizers which are maximally violated by general stabilizer states, with two observables for each local party. We show that the constructed Bell inequalities can self-test any stabilizer state which is essentially device-independent, if and only if these stabilizers can uniquely determine the state in a device-dependent manner. This bridges the gap between device-independent and device-dependent verification methods. Our framework can provide plenty of Bell inequalities for self-testing stabilizer states. Among them, we give two families of Bell inequalities with different advantages: (1) a family of Bell inequalities with a constant ratio of quantum and classical bounds using 2N correlations, (2) Single pair inequalities improving on all previous robustness self-testing bounds using N+1 correlations, which are both efficient and suitable for realizations in multipartite systems. Our framework can not only inspire more fruitful multipartite Bell inequalities from conventional verification methods, but also pave the way for their practical applications.

quant-ph↗

arXiv2020

Direction Concentration Learning: Enhancing Congruency in Machine Learning

One of the well-known challenges in computer vision tasks is the visual diversity of images, which could result in an agreement or disagreement between the learned knowledge and the visual content exhibited by the current observation. In this work, we first define such an agreement in a concepts learning process as congruency. Formally, given a particular task and sufficiently large dataset, the congruency issue occurs in the learning process whereby the task-specific semantics in the training data are highly varying. We propose a Direction Concentration Learning (DCL) method to improve congruency in the learning process, where enhancing congruency influences the convergence path to be less circuitous. The experimental results show that the proposed DCL method generalizes to state-of-the-art models and optimizers, as well as improves the performances of saliency prediction task, continual learning task, and classification task. Moreover, it helps mitigate the catastrophic forgetting problem in the continual learning task. The code is publicly available at https://github.com/luoyan407/congruency.

cs.LG↗

arXiv2019

Learning metrics for persistence-based summaries and applications for graph classification

Recently a new feature representation and data analysis methodology based on a topological tool called persistent homology (and its corresponding persistence diagram summary) has started to attract momentum. A series of methods have been developed to map a persistence diagram to a vector representation so as to facilitate the downstream use of machine learning tools, and in these approaches, the importance (weight) of different persistence features are often preset. However often in practice, the choice of the weight function should depend on the nature of the specific type of data one considers, and it is thus highly desirable to learn a best weight function (and thus metric for persistence diagrams) from labelled data. We study this problem and develop a new weighted kernel, called WKPI, for persistence summaries, as well as an optimization framework to learn a good metric for persistence summaries. Both our kernel and optimization problem have nice properties. We further apply the learned kernel to the challenging task of graph classification, and show that our WKPI-based classification framework obtains similar or (sometimes significantly) better results than the best results from a range of previous graph classification frameworks on a collection of benchmark datasets.

cs.CG↗

arXiv2019

Double Generalized Linear Model Reveals Those with High Intelligence are More Similar in Cortical Thickness

Most studies indicate that intelligence (g) is positively correlated with cortical thickness. However, the interindividual variability of cortical thickness has not been taken into account. In this study, we aimed to identify the association between intelligence and cortical thickness in adolescents from both the group's mean and dispersion point of view, utilizing the structural brain imaging from the Adolescent Brain and Cognitive Development (ABCD) Consortium, the largest cohort in early adolescents around 10 years old. The mean and dispersion parameters of cortical thickness and their association with intelligence were estimated using double generalized linear models(DGLM). We found that for the mean model part, the thickness of the frontal lobe like superior frontal gyrus was negatively related to intelligence, while the surface area was most positively associated with intelligence in the frontal lobe. In the dispersion part, intelligence was negatively correlated with the dispersion of cortical thickness in widespread areas, but not with the surface area. These results suggested that people with higher IQ are more similar in cortical thickness, which may be related to less differentiation or heterogeneity in cortical columns.

q-bio.NC↗

arXiv2019

Human Annotations Improve GAN Performances

Generative Adversarial Networks (GANs) have shown great success in many applications. In this work, we present a novel method that leverages human annotations to improve the quality of generated images. Unlike previous paradigms that directly ask annotators to distinguish between real and fake data in a straightforward way, we propose and annotate a set of carefully designed attributes that encode important image information at various levels, to understand the differences between fake and real images. Specifically, we have collected an annotated dataset that contains 600 fake images and 400 real images. These images are evaluated by 10 workers from the Amazon Mechanical Turk (AMT) based on eight carefully defined attributes. Statistical analyses have revealed different distributions of the proposed attributes between real and fake images. These attributes are shown to be useful in discriminating fake images from real ones, and deep neural networks are developed to automatically predict the attributes. We further utilize the information by integrating the attributes into GANs to generate better images. Experimental results evaluated by multiple metrics show performance improvement of the proposed model.

cs.CV↗

arXiv2019

Simultaneous Certification of Entangled States and Measurements in Bounded Dimensional Semi-Quantum Games

Certification of quantum systems and operations is a central task in quantum information processing. Most current schemes rely on a tomography with fully characterised devices, while this may not be met in real experiments. Device characterisations can be removed with device-independent tests, it is technically challenging at the moment, though. In this letter, we investigate the problem of certifying entangled states and measurements via semi-quantum games, a type of non-local quantum games with well characterised quantum inputs, balancing practicality and device-independence. We first design a specific bounded-dimensional measurement-device-independent game, with which we simultaneously certify any pure entangled state and Bell state measurement operators. Afterwards via a duality treatment of state and measurement, we interpret the dual form of this game as a source-independent bounded-dimensional entanglement swapping protocol and show the whole process, including any entangled projector and Bell states, can be certified with this protocol. In particular, our results do not require a complete Bell state measurement, which is beneficial for experiments and practical use.

quant-ph↗

arXiv2019

Quantum Coherence and Intrinsic Randomness

The peculiar uncertainty or randomness of quantum measurements stems from coherence, whose information-theoretic characterization is currently under investigation. Under the resource theory of coherence, it is interesting to investigate interpretations of coherence measures and the interplay with other quantum properties, such as quantum correlations and intrinsic randomness. Coherence can be viewed as the resource for the intrinsic randomness in the measurement outcomes of a state in the computational basis. We observed in our previous work that the coherence of formation, which measures the asymptotic coherence dilution rate, indeed quantifies the uncertainty of a (classical) correlated party about the system measurement outcome. In this work, we re-derive the result from a quantum point of view and then connect the intrinsic randomness to the relative entropy of coherence, another important coherence measure that quantifies the asymptotic distillable coherence. Even though there does not exist bound coherent states, these two intrinsic randomness quantified by coherence of formation and the relative entropy of coherence are different. Interestingly, we show that this gap is equal to the quantum discord, a general form of quantum correlations, in the state of the system of interest and the correlated party, after a local measurement on the former system.

quant-ph↗

arXiv2019

Detecting multipartite entanglement structure with minimal resources

Recently, there are tremendous developments on the number of controllable qubits in several quantum computing systems. For these implementations, it is crucial to determine the entanglement structure of the prepared multipartite quantum state as a basis for further information processing tasks. In reality, evaluation of a multipartite state is in general a very challenging task owing to the exponential increase of the Hilbert space with respect to the number of system components. In this work, we propose a systematic method using very few local measurements to detect multipartite entanglement structures based on the graph state --- one of the most important classes of quantum states for quantum information processing. Thanks to the close connection between the Schmidt coefficient and quantum entropy in graph states, we develop a family of efficient witness operators to detect the entanglement between subsystems under any partitions and hence the entanglement intactness. We show that the number of local measurements equals to the chromatic number of the underlying graph, which is a constant number, independent of the number of qubits. In reality, the optimization problem involved in the witnesses can be challenging with large system size. For several widely-used graph states, such as 1-D and 2-D cluster states and the Greenberger-Horne-Zeilinger state, by taking advantage of the area law of entanglement entropy, we derive analytical solutions for the witnesses, which only employ two local measurements. Our method offers a standard tool for entanglement structure detection to benchmark multipartite quantum systems.

quant-ph↗

arXiv2019

Theory of variational quantum simulation

The variational method is a versatile tool for classical simulation of a variety of quantum systems. Great efforts have recently been devoted to its extension to quantum computing for efficiently solving static many-body problems and simulating real and imaginary time dynamics. In this work, we first review the conventional variational principles, including the Rayleigh-Ritz method for solving static problems, and the Dirac and Frenkel variational principle, the McLachlan's variational principle, and the time-dependent variational principle, for simulating real time dynamics. We focus on the simulation of dynamics and discuss the connections of the three variational principles. Previous works mainly focus on the unitary evolution of pure states. In this work, we introduce variational quantum simulation of mixed states under general stochastic evolution. We show how the results can be reduced to the pure state case with a correction term that takes accounts of global phase alignment. For variational simulation of imaginary time evolution, we also extend it to the mixed state scenario and discuss variational Gibbs state preparation. We further elaborate on the design of ansatz that is compatible with post-selection measurement and the implementation of the generalised variational algorithms with quantum circuits. Our work completes the theory of variational quantum simulation of general real and imaginary time evolution and it is applicable to near-term quantum hardware.

quant-ph↗

arXiv2019

Interact as You Intend: Intention-Driven Human-Object Interaction Detection

The recent advances in instance-level detection tasks lay strong foundation for genuine comprehension of the visual scenes. However, the ability to fully comprehend a social scene is still in its preliminary stage. In this work, we focus on detecting human-object interactions (HOIs) in social scene images, which is demanding in terms of research and increasingly useful for practical applications. To undertake social tasks interacting with objects, humans direct their attention and move their body based on their intention. Based on this observation, we provide a unique computational perspective to explore human intention in HOI detection. Specifically, the proposed human intention-driven HOI detection (iHOI) framework models human pose with the relative distances from body joints to the object instances. It also utilizes human gaze to guide the attended contextual regions in a weakly-supervised setting. In addition, we propose a hard negative sampling strategy to address the problem of mis-grouping. We perform extensive experiments on two benchmark datasets, namely V-COCO and HICO-DET. The efficacy of each proposed component has also been validated.

cs.CV↗

arXiv2019

$\mathcal{G}$-softmax: Improving Intra-class Compactness and Inter-class Separability of Features

Intra-class compactness and inter-class separability are crucial indicators to measure the effectiveness of a model to produce discriminative features, where intra-class compactness indicates how close the features with the same label are to each other and inter-class separability indicates how far away the features with different labels are. In this work, we investigate intra-class compactness and inter-class separability of features learned by convolutional networks and propose a Gaussian-based softmax ($\mathcal{G}$-softmax) function that can effectively improve intra-class compactness and inter-class separability. The proposed function is simple to implement and can easily replace the softmax function. We evaluate the proposed $\mathcal{G}$-softmax function on classification datasets (i.e., CIFAR-10, CIFAR-100, and Tiny ImageNet) and on multi-label classification datasets (i.e., MS COCO and NUS-WIDE). The experimental results show that the proposed $\mathcal{G}$-softmax function improves the state-of-the-art models across all evaluated datasets. In addition, analysis of the intra-class compactness and inter-class separability demonstrates the advantages of the proposed function over the softmax function, which is consistent with the performance improvement. More importantly, we observe that high intra-class compactness and inter-class separability are linearly correlated to average precision on MS COCO and NUS-WIDE. This implies that improvement of intra-class compactness and inter-class separability would lead to improvement of average precision.

cs.CV↗

arXiv2019

One-shot Coherence Distillation with Catalysts

The resource theory of quantum coherence is an important topic in quantum information science. Standard coherence distillation and dilution problems have been thoroughly studied. In this paper, we introduce and study the problem of one-shot coherence distillation with catalysts. In order to distill more coherence from a state of interest, a catalytic system can be involved and a jointly free operation is applied to both systems. The joint output state should be a maximally coherent state in tensor product with the unchanged catalysts, with some allowable fidelity error. We consider several different definitions of this problem. Firstly, with a small fidelity error in both systems, we show that, even via the smallest free operation class (PIO), the distillable coherence of any state with no restriction on the catalysts is infinite, which is a "coherence embezzling phenomenon". We then define and calculate a lower bound for the distillable coherence when the dimension of catalysts is restricted. Finally, in consideration of physical relevance, we define the "perfect catalysts" scenario where the catalysts are required to be pure and precisely unchanged. Interestingly, we show that in this setting catalysts basically provide no advantages in pure state distillation via IO and SIO under certain smoothing restriction. Our work enhances the understanding of catalytic effect in quantum resource theory.

quant-ph↗

arXiv2019

Learning to Learn from Noisy Labeled Data

Despite the success of deep neural networks (DNNs) in image classification tasks, the human-level performance relies on massive training data with high-quality manual annotations, which are expensive and time-consuming to collect. There exist many inexpensive data sources on the web, but they tend to contain inaccurate labels. Training on noisy labeled datasets causes performance degradation because DNNs can easily overfit to the label noise. To overcome this problem, we propose a noise-tolerant training algorithm, where a meta-learning update is performed prior to conventional gradient update. The proposed meta-learning method simulates actual training by generating synthetic noisy labels, and train the model such that after one gradient update using each set of synthetic noisy labels, the model does not overfit to the specific noise. We conduct extensive experiments on the noisy CIFAR-10 dataset and the Clothing1M dataset. The results demonstrate the advantageous performance of the proposed method compared to several state-of-the-art baselines.

cs.LG↗