Vision Transformers (ViT) is known for its scalability. In this work, we target to scale down a ViT to fit in an environment with dynamic-changing resource constraints. We observe that smaller ViTs are intrinsically the sub-networks of a larger ViT with different widths. Thus, we propose a general framework, named Scala, to enable a single network to represent multiple smaller ViTs with flexible inference capability, which aligns with the inherent design of ViT to vary from widths. Concretely, Scala activates several subnets during training, introduces Isolated Activation to disentangle the smallest sub-network from other subnets, and leverages Scale Coordination to ensure each sub-network receives simplified, steady, and accurate learning objectives. Comprehensive empirical validations on different tasks demonstrate that with only one-shot training, Scala learns slimmable representation without modifying the original ViT structure and matches the performance of Separate Training. Compared with the prior art, Scala achieves an average improvement of 1.6% on ImageNet-1K with fewer parameters.

Image retrieval is a fundamental task in computer vision. Despite recent advances in this field, many techniques have been evaluated on a limited number of domains, with a small number of instance categories. Notably, most existing works only consider domains like 3D landmarks, making it difficult to generalize the conclusions made by these works to other domains, e.g., logo and other 2D flat objects. To bridge this gap, we introduce a new dataset for benchmarking visual search methods on flat images with diverse patterns. Our flat object retrieval benchmark (FORB) supplements the commonly adopted 3D object domain, and more importantly, it serves as a testbed for assessing the image embedding quality on out-of-distribution domains. In this benchmark we investigate the retrieval accuracy of representative methods in terms of candidate ranks, as well as matching score margin, a viewpoint which is largely ignored by many works. Our experiments not only highlight the challenges and rich heterogeneity of FORB, but also reveal the hidden properties of different retrieval strategies. The proposed benchmark is a growing project and we expect to expand in both quantity and variety of objects. The dataset and supporting codes are available at https://github.com/pxiangwu/FORB/.

Despite the remarkable accuracy of deep neural networks in object detection, they are costly to train and scale due to supervision requirements. Particularly, learning more object categories typically requires proportionally more bounding box annotations. Weakly supervised and zero-shot learning techniques have been explored to scale object detectors to more categories with less supervision, but they have not been as successful and widely adopted as supervised models. In this paper, we put forth a novel formulation of the object detection problem, namely open-vocabulary object detection, which is more general, more practical, and more effective than weakly supervised and zero-shot approaches. We propose a new method to train object detectors using bounding box annotations for a limited set of object categories, as well as image-caption pairs that cover a larger variety of objects at a significantly lower cost. We show that the proposed method can detect and localize objects for which no bounding box annotation is provided during training, at a significantly higher accuracy than zero-shot approaches. Meanwhile, objects with bounding box annotation can be detected almost as accurately as supervised methods, which is significantly better than weakly supervised baselines. Accordingly, we establish a new state of the art for scalable object detection.

Music emotion recognition (MER) aims to recognize the affective content of a piece of music, which is important for applications such as automatic soundtrack generation and music recommendation. MER is commonly formulated as a supervised learning problem. In practice, except for Pop music, there is little labeled data in most genres. In addition, emotion is genre specific in music and thus the labeled data of Pop music cannot be used for other genres. In this paper, we aim to solve the genre-specific MER problem by exploiting two kinds of auxiliary data: unlabeled songs and social tags. However, using these two kinds of data effectively is a non-trivial task, e.g. tags are noisy and therefore cannot be treated as fully trustworthy. To build an accurate model with the help from the unlabeled songs and noisy tags, we present SMART, which stands for Semi-Supervised Music Affective Emotion Recognition with Social Tagging, combining of a graph-based semi-supervised learning algorithm with a novel tag refinement method. Experiments on the Million Song Dataset show that our proposed approach, trained with only 10 labeled instances, is as accurate as Support Vector Regression trained with 750 labeled songs.

Bluetooth information can efficiently capture characteristics of user-centric surrounding contexts, such as formal meeting or chatting with friends, shopping with friends or alone, etc. In this paper, we extract novel features from Bluetooth traces and use these features for recognizing contextual behavior as well as inferring continuous episode transition. Evaluation results show that extracted novel features are very effective, which enable the model to achieve an average of 87% accuracy for specific context classification and the ability of episode inference from real-life Bluetooth traces.

Activity recognition aims to identify and predict human activities based on a series of sensor readings. In recent years, machine learning methods have become popular in solving activity recognition problems. A special difficulty for adopting machine learning methods is the workload to annotate a large number of sensor readings as training data. Labeling sensor readings for their corresponding activities is a time-consuming task. In practice, we often have a set of labeled training instances ready for an activity recognition task. If we can transfer such knowledge to a new activity recognition scenario that is different from, but related to, the source domain, it will ease our effort to perform manual labeling of training data for the new scenario. In this paper, we propose a transfer learning framework based on automatically learning a correspondence between different sets of sensors to solve this transfer-learning in activity recognition problem. We validate our framework on two different datasets and compare it against previous approaches of activity recognition, and demonstrate its effectiveness.

Designing Quantitative Structure-Activity Relationship (QSAR) models has been a recurrent research interest for biologists and computer scientists. An example is to predict the toxicity of chemical compounds using their structural properties as features represented by graphs. A popular method to classify these graphs is to exploit classifiers such as support vector machines (SVMs) and graph kernels to incorporate the sequential, structural and chemical information. Previous works have focused on designing specific graph kernels for this task, amongst which graph alignment kernels are one of the most popular approach. Graph alignment kernels align the nodes of one graph to the nodes of the second graph so that the total overall similarity is maximized with respect to all possible alignments. However, taking both vertex and edge similarities into account makes the problem NP-Hard. In this paper, we present a novel general graph-matching based method for QSAR. We view the problem of calculating optimal assignments of two attributed graphs from a different perspective. Instead of first designing an atom kernel function and a bond kernel function, we first provide a training set of pairs of graphs with their corresponding matchings. We then try to learn the compatibility function over atoms and use only the atom kernel function to compute graph matchings. Our algorithm has the advantage of being more general and yet efficient than previous approaches for the QSAR problem. We evaluate our method on a set of chemical structure-activity prediction benchmark datasets, and show that our algorithm can achieve better or comparable accuracies over the optimal assignment kernel method.

A major problem of classification learning is the lack of ground-truth labeled data. It is usually expensive to label new data instances for training a model. To solve this problem, domain adaptation in transfer learning has been proposed to classify target domain data by using some other source domain data, even when the data may have different distributions. However, domain adaptation may not work well when the differences between the source and target domains are large. In this paper, we design a novel transfer learning approach, called BIG (Bridging Information Gap), to effectively extract useful knowledge in a worldwide knowledge base, which is then used to link the source and target domains for improving the classification performance. BIG works when the source and target domains share the same feature space but different underlying data distributions. Using the auxiliary source data, we can extract a ¿bridge¿ that allows cross-domain text classification problems to be solved using standard semisupervised learning algorithms. A major contribution of our work is that with BIG, a large amount of worldwide knowledge can be easily adapted and used for learning in the target domain. We conduct experiments on several real-world cross-domain text classification tasks and demonstrate that our proposed approach can outperform several existing domain adaptation approaches significantly.

Query classification (QC) is a task that aims to classify Web queries into topical categories. Since queries are usually short in length and ambiguous, the same query may need to be classified to different categories according to different people's perspectives. In this paper, we propose the Personalized Query Classification (PQC) task and develop an algorithm based on user preference learning as a solution. Users' preferences that are hidden in clickthrough logs are quite helpful for search engines to improve their understandings of users' queries. We propose to connect query classification with users' preference learning from clickthrough logs for PQC. To tackle the sparseness problem in clickthrough logs, we propose a collaborative ranking model to leverage similar users' information. Experiments on a real world clickthrough log data show that our proposed PQC algorithm can gain significant improvement compared with general QC as well as natural baselines. Our method can be applied to a wide range of applications including personalized search and online advertising.

In activity recognition, one major challenge is huge manual effort in labeling when a new domain of activities is to be tested. In this paper, we ask an interesting question: can we transfer the available labeled data from a set of existing activities in one domain to help recognize the activities in another different but related domain? Our answer is "yes", provided that the sensor data from the two domains are related in some way. We develop a bridge between the activities in two domains by learning a similarity function via Web search, under the condition that the sensor data are from the same feature space. Based on the learned similarity measures, our algorithm interprets the data from the source domain as the data in the domain with different confidence levels, thus accomplishing the cross-domain knowledge transfer task. Our algorithm is evaluated on several real-world datasets to demonstrate its effectiveness.

Understanding users'search intent expressed through their search queries is crucial to Web search and online advertisement. Web query classification (QC) has been widely studied for this purpose. Most previous QC algorithms classify individual queries without considering their context information. However, as exemplified by the well-known example on query "jaguar", many Web queries are short and ambiguous, whose real meanings are uncertain without the context information. In this paper, we incorporate context information into the problem of query classification by using conditional random field (CRF) models. In our approach, we use neighboring queries and their corresponding clicked URLs (Web pages) in search sessions as the context information. We perform extensive experiments on real world search logs and validate the effectiveness and effciency of our approach. We show that we can improve the F1 score by 52% as compared to other state-of-the-art baselines.

BACKGROUND: Protein subcellular localization is concerned with predicting the location of a protein within a cell using computational method. The location information can indicate key functionalities of proteins. Accurate predictions of subcellular localizations of protein can aid the prediction of protein function and genome annotation, as well as the identification of drug targets. Computational methods based on machine learning, such as support vector machine approaches, have already been widely used in the prediction of protein subcellular localization. However, a major drawback of these machine learning-based approaches is that a large amount of data should be labeled in order to let the prediction system learn a classifier of good generalization ability. However, in real world cases, it is laborious, expensive and time-consuming to experimentally determine the subcellular localization of a protein and prepare instances of labeled data. RESULTS: In this paper, we present an approach based on a new learning framework, semi-supervised learning, which can use much fewer labeled instances to construct a high quality prediction model. We construct an initial classifier using a small set of labeled examples first, and then use unlabeled instances to refine the classifier for future predictions. CONCLUSION: Experimental results show that our methods can effectively reduce the workload for labeling data using the unlabeled data. Our method is shown to enhance the state-of-the-art prediction results of SVM classifiers by more than 10%.

Recognizing and understanding the activities of people from sensor readings is an important task in ubiquitous computing. Activity recognition is also a particularly difficult task because of the inherent uncertainty and complexity of the data collected by the sensors. Many researchers have tackled this problem in an overly simplistic setting by assuming that users often carry out single activities one at a time or multiple activities consecutively, one after another. However, so far there has been no formal exploration on the degree in which humans perform concurrent or interleaving activities, and no thorough study on how to detect multiple goals in a real world scenario. In this article, we ask the fundamental questions of whether users often carry out multiple concurrent and interleaving activities or single activities in their daily life, and if so, whether such complex behavior can be detected accurately using sensors. We define several classes of complexity levels under a goal taxonomy that describe different granularities of activities, and relate the recognition accuracy with different complexity levels or granularities. We present a theoretical framework for recognizing multiple concurrent and interleaving activities, and evaluate the framework in several real-world ubiquitous computing environments.