Learning Temporal Sequence Model from Partially Labeled Data

Density-aware generative algorithms learning from positive examples have verified high recall for visual object detection, but such generative methods suffer from excessive false positives which leads to low precision. Inspired by the recent success of detection-recognition pipeline with deep neural networks, this paper proposes a two-step framework by training a generative detector with positive samples first and then utilising a discriminative model to get rid of false positives in those detected bounding box candidates by the generative detector. Evidently, the discriminative model can be viewed as a post-processing step which improves the robustness by distinguishing true positives from false positives that confuse the generative detector. We exemplify the proposed approach on public ImageNet classes to demonstrate the significant improvement on precision while using only positive examples in training.

A common obstacle in effective learning of visual concept classifiers is the scarcity of positive training examples due to expensive labeling cost. This paper explores the sampling of weakly tagged web images for concept learning without human assistance. In particular, ontology knowledge is incorporated for semantic pooling of positive examples from ontologically neighboring concepts. This effectively widens the coverage of the positive samples with visually more diversified content, which is important for learning a good concept classifier. We experiment with two learning strategies: aggregate and incremental. The former strategy re-trains a new classifier by combining existing and newly collected examples, while the latter updates the existing model using the new samples incrementally. Extensive experiments on NUS-WIDE and VOC 2010 datasets show very encouraging results, even when comparing with classifiers learnt using expert labeled training examples.

Establishing structural or functional relationship between sequences, for instance to infer the structural class of an unannotated protein, is a key task in biological analysis. Protein sequences undergo complex transformations such as mutation, insertion and deletion during the evolutionary process and typically share low sequence similarity on the superfamily level, making the task for remote homology detection based on primary sequence only very challenging. Based on previous studies stating that knowledge based on only a subset of critical positions and the preferred symbols on such positions are sufficient for remote homology detection, we present a series of works, each enforcing different notion of sparsity, to recover such critical positions. We first start with a generative model and present the sparse profile hidden Markov models. Such generative approach recovers some critical patterns and motivates the need for discriminative learning. In our second study, we present a discriminative approach to recover such critical positions and the preferred symbols. In our third study, we address the issue of very few positive training examples, accompanied by a large number of negative training examples, which is typical in many remote homology detection task. Such issue motivates the need for semi-supervised learning. However, though containing abundant useful and critical information, large uncurated sequence databases also contain a lot of noise, which may compromise the quality of the classifiers. As a result, we present a systematic and biologically motivated framework for semi-supervised learning with large uncurated sequence databases. Combined with a very fast string kernel, our method not only realizes rapid and accurate remote homology detection and show state-of-the-art performance, but also recovers some critical patterns conserved in superfamilies.

BackgroundThe rapid growth of biomedical literature requires accurate text analysis and text processing tools. Detecting abbreviations and identifying their definitions is an important component of such tools. Most existing approaches for the abbreviation definition identification task employ rule-based methods. While achieving high precision, rule-based methods are limited to the rules defined and fail to capture many uncommon definition patterns. Supervised learning techniques, which offer more flexibility in detecting abbreviation definitions, have also been applied to the problem. However, they require manually labeled training data.MethodsIn this work, we develop a machine learning algorithm for abbreviation definition identification in text which makes use of what we term naturally labeled data. Positive training examples are naturally occurring potential abbreviation-definition pairs in text. Negative training examples are generated by randomly mixing potential abbreviations with unrelated potential definitions. The machine learner is trained to distinguish between these two sets of examples. Then, the learned feature weights are used to identify the abbreviation full form. This approach does not require manually labeled training data.ResultsWe evaluate the performance of our algorithm on the Ab3P, BIOADI and Medstract corpora. Our system demonstrated results that compare favourably to the existing Ab3P and BIOADI systems. We achieve an F-measure of 91.36% on Ab3P corpus, and an F-measure of 87.13% on BIOADI corpus which are superior to the results reported by Ab3P and BIOADI systems. Moreover, we outperform these systems in terms of recall, which is one of our goals.

The aim of this work is to localize a query photograph by finding other images depicting the same place in a large geotagged image database. This is a challenging task due to changes in viewpoint, imaging conditions and the large size of the image database. The contribution of this work is two-fold. First, we cast the place recognition problem as a classification task and use the available geotags to train a classifier for each location in the database in a similar manner to per-exemplar SVMs in object recognition. Second, as only few positive training examples are available for each location, we propose a new approach to calibrate all the per-location SVM classifiers using only the negative examples. The calibration we propose relies on a significance measure essentially equivalent to the p-values classically used in statistical hypothesis testing. Experiments are performed on a database of 25,000 geotagged street view images of Pittsburgh and demonstrate improved place recognition accuracy of the proposed approach over the previous work.

For hyper-rectangles in $$\mathbb{R}^{d}$$ Auer (1997) proved a PAC bound of $$O(\frac{1}{\varepsilon}(d+\log \frac{1}{\delta}))$$ , where $$\varepsilon$$ and $$\delta$$ are the accuracy and confidence parameters. It is still an open question whether one can obtain the same bound for intersection-closed concept classes of VC-dimension $$d$$ in general. We present a step towards a solution of this problem showing on one hand a new PAC bound of $$O(\frac{1}{\varepsilon}(d\log d + \log \frac{1}{\delta}))$$ for arbitrary intersection-closed concept classes, complementing the well-known bounds $$O(\frac{1}{\varepsilon}(\log \frac{1}{\delta}+d\log \frac{1}{\varepsilon}))$$ and $$O(\frac{d}{\varepsilon}\log \frac{1}{\delta})$$ of Blumer et al. and (1989) and Haussler, Littlestone and Warmuth (1994). Our bound is established using the closure algorithm, that generates as its hypothesis the intersection of all concepts that are consistent with the positive training examples. On the other hand, we show that many intersection-closed concept classes including e.g. maximum intersection-closed classes satisfy an additional combinatorial property that allows a proof of the optimal bound of $$O(\frac{1}{\varepsilon}(d+\log \frac{1}{\delta}))$$ . For such improved bounds the choice of the learning algorithm is crucial, as there are consistent learning algorithms that need $$\Omega(\frac{1}{\varepsilon}(d\log\frac{1}{\varepsilon} +\log\frac{1}{\delta}))$$ examples to learn some particular maximum intersection-closed concept classes.

An object detection method is developed based on combining top-down recognition with bottom-up image segmentation. There are three main steps in this method: a shape learning step, a hypothesis generation step and a verification step. In the shape learning step, training images are used to building codebook dictionary, and matching template is also constructed based on codebook dictionary. In the top-down hypothesis generation step, the Probability Shape Context (PSC) is used to generate a set of hypotheses of object locations. In the verification step, the hypotheses of object locations are segmented based on Partial Differential Equation (PDE). Based on segmented outlines and matching template, the inference is used to prune out false positives. Experimental results demonstrate that our proposed approach is able to accurately detect objects with only a few positive training examples.

Hierarchical learning strategy in semantic relation extraction

Amongst all the methods of machine learning, learning from examples is considered as a key to acquire knowledge automatically. Learning from examples is to obtain a general cover through induction from a given set of positive and negative examples of a concept, which may describe all the positive examples, but reject all the negative examples of that concept. Extension matrix theory, a branch of learning from examples, presents the training examples as two matrices which consist of positive and negative examples respectively, and then finds a path satisfied with a positive example against the background of negative examples. Obviously, such a path is a conjunction of conditions that is satisfied by the positive example. However, most of all of pattern recognition problems, such as handwritten Chinese character recognition, have an overlay area. It is necessary to solve the problem of nonlinear boundaries among different classes when machine learning is applied to acquire the recognition rules. An advanced extension matrix algorithm is proposed in this paper, in which a heuristic search based on the average entropy is used to get the approximate solutions of the minimal complex. A potential function is used to estimate the probability density function of the overlay area between positive and negative examples, so that the nonlinear interfaces of the interclass areas may be obtained. This algorithm is applied to supervised learning for Chinese character recognition to acquire recognition rules, and the implementations upon 751,000 loosely-constrained handwritten Chinese characters indicate that these methodologies can be applied to a practical recognition system with promising results.

This paper presents a novel method to retrieve images by learning the embedded visual concept from a set of given examples. Through a user's relevance feedback, the visual concept can be effectively learned to classify images which contain common visual entities. The learning process is started by providing a set of either positive or negative training examples and is then interactively adjusted according to the user's relevance feedback. In contrast to traditional methods, the proposed method utilizes a novel way to overcome the under-training problem which is frequently suffered in the learning process. Since no time-consuming optimization process is involved, the proposed method learns the visual concepts extremely fast. Therefore, the target concept can be learned on-line and is user-adaptable for effective retrieval of image contents. Experimental results are provided to prove the superiority of the proposed method.

A data fusion method is extensively investigated to stabilize the motion of aircraft with the visual and inertial measurements in the unknown environment without the signal of the global positioning system and with the unknown inertial measurement biases. There is no information on the landmarks on the ground. What can be used are the sequence of collected images. The main idea of the method is to apply the correspondence between the two successive images based on a new formulation of the extended Kalman filtering to fuse the visual and inertial measurements to predict the necessary feedback information and the unknown inertial measurement biases. Neither computation nor decomposition of any essential matrix is needed. The observability of the new formulation of the extended Kalman filtering and the robustness against the measurement errors of the resulting closed-loop system are analyzed in simulation. The simulation results demonstrate that the attitude angles are observable, however they have no robustness against the visual measurement errors. The motion of aircraft can be stabilized with the 3D visual and inertial measurements as well as the measurable Euler angles in presence of the measurement errors.

VIO-DualProNet: Visual-inertial odometry with learning based process noise covariance

Due to the lack of direct distance measurements, robust and accurate state estimation at high altitude but GPS-denied environments is a challenging task. A possible solution is monocular visual–inertial estimators, in which visual and inertial measurements are properly fused to recover the metric estimates. However, these estimators suffer from initialization under poor numerical conditioning or even degeneration, due to difficulties in retrieving observations of visual features with sufficient parallax, and the excessive period of inertial measurement integration. In this letter, we introduce the joint formulation into the spline-based high altitude estimator initialization method for monocular visual–inertial navigation system, in which the fitting of the spline and the alignment of visual measurements and inertial measurements are jointly optimized to recover metric estimates. The method ensures that sufficient excitation is contained in the inertial measurements when initialized, which eliminates the numerical issues. Compared with the work of Liu and Shen, the joint formulation makes our initialization method insensitive to the choice of spline parameters. Thus, the adaptivity to various environments and motions is obtained, as well as higher accuracy. The method is applicable for both loosely coupled and tightly coupled visual–inertial estimators. Extensive experiments are conducted to validate our approach.

In the last few years, object detection techniques have progressed immensely. Impressive detection results have been achieved for many objects such as faces [11, 14, 9] and cars [11]. The robustness of these systems emerges from a training stage utilizing thousands of positive examples. One approach to enable learning from a small set of training examples is to find an efficient set of features that accurately represent the target object. Unfortunately, automatically selecting such a feature set is a difficult task in itself. In this paper we present a novel feature selection method that is based on the notion of object categories. We assume that when learning to recognize a new object (like an apple) we also know a category it belongs to (fruit). We further assume that features that are useful for learning other objects in the same category (e.g. pear or orange) will also be useful for learning the novel object. This leads to a simple criterion for selecting features and building classifiers. We show that our method gives significant improvement in detection performance in challenging domains.

Face detection systems have recently achieved high detection rates and real-time performance. However, these methods usually rely on a huge training database (around 5,000 positive examples for good performance). While such huge databases may be feasible for building a system that detects a single object, it is obviously problematic for scenarios where multiple objects (or multiple views of a single object) need to be detected. Indeed, even for multi-viewface detection the performance of existing systems is far from satisfactory. In this work we focus on the problem of learning to detect objects from a small training database. We show that performance depends crucially on the features that are used to represent the objects. Specifically, we show that using local edge orientation histograms (EOH) as features can significantly improve performance compared to the standard linear features used in existing systems. For frontal faces, local orientation histograms enable state of the art performance using only a few hundred training examples. For profile view faces, local orientation histograms enable learning a system that seems to outperform the state of the art in real-time systems even with a small number of training examples.