End-to-end object detection methods have attracted extensive interest recently since they alleviate the need for complicated human-designed components and simplify the detection pipeline. However, these methods suffer from slower training convergence and inferior detection performance compared to conventional detectors, as their feature fusion and selection processes are constrained by insufficient positive supervision. To address this issue, we introduce a novel query-selection encoder (QSE) designed for end-to-end object detectors to improve the training convergence speed and detection accuracy. QSE is composed of multiple encoder layers stacked on top of the backbone. A lightweight head network is added after each encoder layer to continuously optimize features in a cascading manner, providing more positive supervision for efficient training. Additionally, a hierarchical feature-aware attention (HFA) mechanism is incorporated in each encoder layer, including in- and cross-level feature attention, to enhance the interaction between features from different levels. HFA can effectively suppress similar feature representations and highlight discriminative ones, thereby accelerating the feature selection process. Our method is highly versatile in accommodating both CNN- and Transformer-based detectors. Extensive experiments were conducted on the popular benchmark datasets MS COCO, CrowdHuman, and PASCAL VOC to demonstrate the effectiveness of our method. The results showed that CNN- and Transformer-based detectors using QSE can achieve better end-to-end performance within fewer training epochs.

Abstract The one‐stage object detector has been widely applied in many computer vision applications due to its high detection efficiency and simple framework. However, one‐stage detectors heavily rely on Non‐maximum Suppression to remove the duplicated predictions for the same objects, and the detectors produce detection confidence to measure the quality of those predictions. The localisation quality is an important factor to evaluate the predicted bounding boxes, but its role has not been fully utilised in previous works. To alleviate the problem, the Quality Prediction Block (QPB), a lightweight sub‐network, is designed by the authors, which strengthens the effect of localisation quality evaluation on detection confidence by leveraging the features of predicted bounding boxes. The QPB is simple in structure and applies to different forms of detection confidence. Extensive experiments are conducted on the public benchmarks, MS COCO, PASCAL VOC and Berkeley DeepDrive. The results demonstrate the effectiveness of our method in the detectors with various forms of detection confidence. The proposed approach also achieves better performance in the stronger one‐stage detectors.

Mandarin text-to-speech (TTS) systems heavily depend on front-end processing, such as grapheme-to-phoneme conversion and prosodic boundary prediction, to produce expressive, human-like speech. Utilizing a pre-trained language model, such as the bidirectional encoder representations from Transformers (BERT), could significantly improve the TTS front-end’s performance. However, the original BERT is too big for edge TTS applications with tight limits on memory costs and inference latency. Although a distilled BERT can alleviate this problem, a considerable efficiency barrier may still exist due to the self-attention module's quadratic complexity and the feed-forward module's enormous computation. To this end, we propose a lightweight distilled convolution network as an alternative to the distilled BERT. Unlike previous knowledge distillation methods, which commonly used the same self-attention network for the teacher and student models, we transfer knowledge from the self-attention network to a convolution network. Experiments on two major Mandarin TTS front-end tasks have shown that our distilled convolution model can achieve comparable results to various distilled BERT variants while drastically reducing the model size and inference latency.

The one-stage object detector has recently attracted extensive interest due to its high detection efficiency and simple framework. However, one-stage detectors suffer much from the extreme positive-negative imbalance since the majority of samples are labeled as the negative. To address this problem, a novel activation function, the Gradient Enhanced Function (GEF), is devised for the last output layer of the one-stage detector to strengthen the role of positive samples during the optimization process. The Quality-Guided Loss (QGL) is proposed for the classification task in object detection, which further makes the training focus on the high-quality positive samples. In addition, QGL is devised to handle different classification labels (i.e. the one-hot label and the soft label). Owing to its simplicity and effectiveness, the QGL together with the GEF might be applicable to various one-stage methods, including anchor-based and anchor-free detectors. Comprehensive experiments are conducted on multiple public benchmarks, MS COCO, PASCAL VOC and Berkeley DeepDrive. The results demonstrate that the proposed approach achieves notable improvements in both anchor-based and anchor-free detectors with various classification labels. The effectiveness of the GEF and the QGL is further verified in the stronger one-stage detectors.

BACKGROUND: The main limitations of most existing clustering methods used in genomic data analysis include heuristic or random algorithm initialization, the potential of finding poor local optima, the lack of cluster number detection, an inability to incorporate prior/expert knowledge, black-box and non-adaptive designs, in addition to the curse of dimensionality and the discernment of uninformative, uninteresting cluster structure associated with confounding variables. RESULTS: In an effort to partially address these limitations, we develop the VIsual Statistical Data Analyzer (VISDA) for cluster modeling, visualization, and discovery in genomic data. VISDA performs progressive, coarse-to-fine (divisive) hierarchical clustering and visualization, supported by hierarchical mixture modeling, supervised/unsupervised informative gene selection, supervised/unsupervised data visualization, and user/prior knowledge guidance, to discover hidden clusters within complex, high-dimensional genomic data. The hierarchical visualization and clustering scheme of VISDA uses multiple local visualization subspaces (one at each node of the hierarchy) and consequent subspace data modeling to reveal both global and local cluster structures in a "divide and conquer" scenario. Multiple projection methods, each sensitive to a distinct type of clustering tendency, are used for data visualization, which increases the likelihood that cluster structures of interest are revealed. Initialization of the full dimensional model is based on first learning models with user/prior knowledge guidance on data projected into the low-dimensional visualization spaces. Model order selection for the high dimensional data is accomplished by Bayesian theoretic criteria and user justification applied via the hierarchy of low-dimensional visualization subspaces. Based on its complementary building blocks and flexible functionality, VISDA is generally applicable for gene clustering, sample clustering, and phenotype clustering (wherein phenotype labels for samples are known), albeit with minor algorithm modifications customized to each of these tasks. CONCLUSION: VISDA achieved robust and superior clustering accuracy, compared with several benchmark clustering schemes. The model order selection scheme in VISDA was shown to be effective for high dimensional genomic data clustering. On muscular dystrophy data and muscle regeneration data, VISDA identified biologically relevant co-expressed gene clusters. VISDA also captured the pathological relationships among different phenotypes revealed at the molecular level, through phenotype clustering on muscular dystrophy data and multi-category cancer data.

Though supervised and unsupervised analyses of genomic data have been intensively studied in recent years, little effort has been made to discover the structural information contained in the data. In this work, we propose a stability analysis guided supervised clustering and visualization method aiming to discover the hierarchical structure in gene expression data, which we call the "tree of phenotypes". We applied the method on two multiclass gene expression microarray data sets and presented the biological plausibility of the learned trees. We also tested the multiclass classifiers built on the learned trees and demonstrated their good classification performance

MOTIVATION: Multilayer perceptrons (MLP) represent one of the widely used and effective machine learning methods currently applied to diagnostic classification based on high-dimensional genomic data. Since the dimensionalities of the existing genomic data often exceed the available sample sizes by orders of magnitude, the MLP performance may degrade owing to the curse of dimensionality and over-fitting, and may not provide acceptable prediction accuracy. RESULTS: Based on Fisher linear discriminant analysis, we designed and implemented an MLP optimization scheme for a two-layer MLP that effectively optimizes the initialization of MLP parameters and MLP architecture. The optimized MLP consistently demonstrated its ability in easing the curse of dimensionality in large microarray datasets. In comparison with a conventional MLP using random initialization, we obtained significant improvements in major performance measures including Bayes classification accuracy, convergence properties and area under the receiver operating characteristic curve (A(z)). SUPPLEMENTARY INFORMATION: The Supplementary information is available on http://www.cbil.ece.vt.edu/publications.htm

The exploration of high dimensional gene expression microarray data demands powerful analytical tools. Our data mining software, visual data analyzer (VISDA) for cluster discovery, reveals many distinguishing patterns among gene expression profiles. The model-supported hierarchical data exploration tool has two complementary schemes: discriminatory dimensionality reduction for structure-focused data visualization, and cluster decomposition by probabilistic clustering. Reducing dimensionality generates the visualization of the complete data set at the top level. This data set is then partitioned into subclusters that can consequently be visualized at lower levels and if necessary partitioned again. These approaches produce different visualizations that are compared against known phenotypes from the microarray experiments. For class prediction on cancers using miroarray data, multilayer perceptrons (MLPs) are trained and optimized, whose architecture and parameters are regularized and initialized by weighted Fisher criterion (wFC)-based discriminatory component analysis (DCA). The prediction performance is compared and evaluated via multifold cross-validation.