Writer adaptation with style transfer mapping

Self-Reinforcing Unsupervised Matching

Remarkable gains in deep learning usually benefit from large-scale supervised data. Ensuring the intra-class modality diversity in training set is critical for generalization capability of cutting-edge deep models, but it burdens human with heavy manual labor on data collection and annotation. In addition, some rare or unexpected modalities are new for the current model, causing reduced performance under such emerging modalities. Inspired by the achievements in speech recognition, psychology and behavioristics, we present a practical solution, self-reinforcing unsupervised matching (SUM), to annotate the images with 2D structure-preserving property in an emerging modality by cross-modality matching. Specifically, we propose a dynamic programming algorithm, dynamic position warping (DPW), to reveal the underlying element correspondence relationship between two matrix-form data in an order-preserving fashion, and devise a local feature adapter (LoFA) to allow for cross-modality similarity measurement. On these bases, we develop a two-tier self-reinforcing learning mechanism on both feature level and image level to optimize the LoFA. The proposed SUM framework requires no any supervision in emerging modality and only one template in seen modality, providing a promising route towards incremental learning and continual learning. Extensive experimental evaluation on two proposed challenging one-template visual matching tasks demonstrate its efficiency and superiority.

Fast writer adaptation with style extractor network for handwritten text recognition

Writing style is an abstract attribute in handwritten text. It plays an important role in recognition systems and is not easy to define explicitly. Considering the effect of writing style, a writer adaptation method is proposed to transform a writer-independent recognizer toward a particular writer. This transformation has the potential to significantly increase accuracy. In this paper, under the deep learning framework, we propose a general fast writer adaptation solution. Specifically, without depending on other complex skills, a well designed style extractor network (SEN) trained by identification loss (IDL) is introduced to explicitly extract personalized writer information. The architecture of SEN consists of a stack of convolutional layers followed by a recurrent neural network with gated recurrent units to remove semantic context and retain writer information. Then, the outputs of the GRU are further integrated into a one-dimensional vector that is adopted to represent writing style. Finally, the extracted style information is fed into the writer-independent recognizer to achieve adaptation. Validated on offline handwritten text recognition tasks, the proposed fast sentence-level adaptation achieves remarkable improvements in Chinese and English text recognition tasks. Specifically, in the HETR task, a multi-information fusion network that is equipped with a hybrid attention mechanism and that integrates visual features, context features and writing style is proposed. In addition, under the same condition (only one writer-specific text line used as adaptation data), the proposed solution, without consuming extra time, can significantly outperform the previous multiple-pass decoding method. The code is available at https://github.com/Wukong90/Handwritten-Text-Recognition.

Stroke Sequence-Dependent Deep Convolutional Neural Network for Online Handwritten Chinese Character Recognition

We propose a novel model, called stroke sequence-dependent deep convolutional neural network (SSDCNN), which uses the stroke sequence information and eight-directional features of Chinese characters for online handwritten Chinese character recognition (OLHCCR). SSDCNN learns the representation of OLHCCs by incorporating the natural sequence information of the strokes. Furthermore, it naturally incorporates the eight-directional features. First, SSDCNN inputs the stroke sequence and transforms it into stacks of feature maps following the writing order of the strokes. Second, the fixed-length, stroke sequence-dependent representations of OLHCC are derived through convolutional, residual, and max-pooling operations. Third, the stroke sequence-dependent representation is combined with the eight-directional features via a number of fully connected neural network layers. Finally, the Chinese characters are recognized using a softmax classifier. The SSDCNN is trained in two stages: 1) the whole architecture is pretrained using the training data until the performance converges to an acceptable degree. 2) The stroke sequence-dependent representation is combined with the eight-directional features by a fully connected neural network and a softmax layer for further training. The model was experimentally evaluated on the OLHCCR competition tasks of International Conference on Document Analysis and Recognition (ICDAR) 2013. The recognition error was a maximum 58.28% lower in SSDCNN than in a model using the eight-directional features alone (5.13% versus 2.14%). Owing to its high accuracy (97.86%), the proposed SSDCNN reduced the recognition error by approximately 18.0% as compared with that of the winning system in the ICDAR 2013 competition. SSDCNN integrated with an adaptation mechanism, called the SSDCNN+Adapt model, and reached a new state-of-the-art (SOTA) standard with an accuracy of 97.94%. The SSDCNN exploits the stroke sequence information to learn high-quality OLHCC representations. Moreover, the learned representation and the classical eight-directional features complement each other within the SSDCNN architecture.

Multisource Transfer Learning for Cross-Subject EEG Emotion Recognition

Electroencephalogram (EEG) has been widely used in emotion recognition due to its high temporal resolution and reliability. Since the individual differences of EEG are large, the emotion recognition models could not be shared across persons, and we need to collect new labeled data to train personal models for new users. In some applications, we hope to acquire models for new persons as fast as possible, and reduce the demand for the labeled data amount. To achieve this goal, we propose a multisource transfer learning method, where existing persons are sources, and the new person is the target. The target data are divided into calibration sessions for training and subsequent sessions for test. The first stage of the method is source selection aimed at locating appropriate sources. The second is style transfer mapping, which reduces the EEG differences between the target and each source. We use few labeled data in the calibration sessions to conduct source selection and style transfer. Finally, we integrate the source models to recognize emotions in the subsequent sessions. The experimental results show that the three-category classification accuracy on benchmark SEED improves by 12.72% comparing with the nontransfer method. Our method facilitates the fast deployment of emotion recognition models by reducing the reliance on the labeled data amount, which has practical significance especially in fast-deployment scenarios.

Analysis and Usage: Subject-to-subject Linear Domain Adaptation in sEMG Classification

Before the operation of a biosignal-based application, long-duration calibration is required to adjust the pre-trained classifier to a new user data (target data). For reducing such time-consuming step, linear domain adaptation (DA) transfer learning approaches, which transfer pooled data (source data) related to the target data, are highlighted. In the last decade, they have been applied to surface electromyogram (sEMG) data with the implicit assumption that sEMG data are linear. However, sEMGs typically have non-linear characteristics, and due to the discrepancy between the assumption and actual characteristics, linear DA approaches would cause a negative transfer. This study investigated how the correlation between the source and target data affects an 8-class forearm movement classification after applying linear DA approaches. As a result, we found significant positive correlations between the classification accuracy and the source-target correlation. Additionally, the source-target correlation depended on the motion class. Therefore, our results suggest that we should choose a non-linear DA approach when the source-target correlation among subjects or motion classes is low.

Discriminative structure learning of sum-product networks for data stream classification

Sum-product network (SPN) is a deep probabilistic representation that allows for exact and tractable inference. There has been a trend of online SPN structure learning from massive and continuous data streams. However, online structure learning of SPNs has been introduced only for the generative settings so far. In this paper, we present an online discriminative approach for SPNs for learning both the structure and parameters. The basic idea is to keep track of informative and representative examples to capture the trend of time-changing class distributions. Specifically, by estimating the goodness of model fitting of data points and dynamically maintaining a certain amount of informative examples over time, we generate new sub-SPNs in a recursive and top-down manner. Meanwhile, an outlier-robust margin-based log-likelihood loss is applied locally to each data point and the parameters of SPN are updated continuously using most probable explanation (MPE) inference. This leads to a fast yet powerful optimization procedure and improved discrimination capability between the genuine class and rival classes. Empirical results show that the proposed approach achieves better prediction performance than the state-of-the-art online structure learner for SPNs, while promising order-of-magnitude speedup. Comparison with state-of-the-art stream classifiers further proves the superiority of our approach.

Combining Structure and Parameter Adaptation of HMMs for Printed Text Recognition

We present two algorithms that extend existing HMM parameter adaptation algorithms (MAP and MLLR) by adapting the HMM structure. This improvement relies on a smart combination of MAP and MLLR with a structure optimization procedure. Our algorithms are semi-supervised: to adapt a given HMM model on new data, they require little labeled data for parameter adaptation and a moderate amount of unlabeled data to estimate the criteria used for HMM structure optimization. Structure optimization is based on state splitting and state merging operations and proceeds so as to optimize either the likelihood or a heuristic criterion. Our algorithms are successfully applied to the recognition of printed characters by adapting the HMM character models of a polyfont printed text recognizer to new fonts. Our experiments involve a total of 1,120,000 real and 3,100,000 synthetic character images and concern a set of 89 HMM models. A comparison of our results with those of state-of-the-art adaptation algorithms (MAP and MLLR) shows a significant increase in the accuracy of character recognition.