The training of medical image analysis systems using machine learning approaches follows a common script: collect and annotate a large dataset, train the classifier on the training set, and test it on a hold-out test set. This process bears no direct resemblance with radiologist training, which is based on solving a series of tasks of increasing difficulty, where each task involves the use of significantly smaller datasets than those used in machine learning. In this paper, we propose a novel training approach inspired by how radiologists are trained. In particular, we explore the use of meta-training that models a classifier based on a series of tasks. Tasks are selected using teacher-student curriculum learning, where each task consists of simple classification problems containing small training sets. We hypothesize that our proposed meta-training approach can be used to pre-train medical image analysis models. This hypothesis is tested on the automatic breast screening classification from DCE-MRI trained with weakly labeled datasets. The classification performance achieved by our approach is shown to be the best in the field for that application, compared to state of art baseline approaches: DenseNet, multiple instance learning and multi-task learning.
Image style transfer has drawn broad attention in recent years. However, most existing methods aim to explicitly model the transformation between different styles, and the learned model is thus not generalizable to new styles. We here propose a unified style transfer framework for both character typeface transfer and neural style transfer tasks leveraging style and content separation. A key merit of such framework is its generalizability to new styles and contents. The overall framework consists of style encoder, content encoder, mixer and decoder. The style encoder and content encoder are used to extract the style and content representations from the corresponding reference images. The mixer integrates the above two representations and feeds it into the decoder to generate images with the target style and content. During training, the encoder networks learn to extract styles and contents from limited size of style/content reference images. This learning framework allows simultaneous style transfer among multiple styles and can be deemed as a special `multi-task' learning scenario. The encoders are expected to capture the underlying features for different styles and contents which is generalizable to new styles and contents. Under this framework, we design two individual networks for character typeface transfer and neural style transfer, respectively. For character typeface transfer, to separate the style features and content features, we leverage the conditional dependence of styles and contents given an image. For neural style transfer, we leverage the statistical information of feature maps in certain layers to represent style. Extensive experimental results have demonstrated the effectiveness and robustness of the proposed methods.
Time series prediction typically consists of a data reconstruction phase where the time series is broken into overlapping windows known as the timespan. The size of the timespan can be seen as a way of determining the extent of past information required for an effective prediction. In certain applications such as the prediction of wind-intensity of storms and cyclones, prediction models need to be dynamic in accommodating different values of the timespan. These applications require robust prediction as soon as the event takes place. We identify a new category of problem called dynamic time series prediction that requires a model to give prediction when presented with varying lengths of the timespan. In this paper, we propose a co-evolutionary multi-task learning method that provides a synergy between multi-task learning and co-evolutionary algorithms to address dynamic time series prediction. The method features effective use of building blocks of knowledge inspired by dynamic programming and multi-task learning. It enables neural networks to retain modularity during training for making a decision in situations even when certain inputs are missing. The effectiveness of the method is demonstrated using one-step-ahead chaotic time series and tropical cyclone wind-intensity prediction.
Recent Image-to-Image Translation algorithms have achieved significant progress in neural style transfer and image attribute manipulation tasks. However, existing approaches require exhaustively labelling training data, which is labor demanding, difficult to scale up, and hard to migrate into new domains. To overcome such a key limitation, we propose Sparsely Grouped Generative Adversarial Networks (SG-GAN) as a novel approach that can translate images on sparsely grouped datasets where only a few samples for training are labelled. Using a novel one-input multi-output architecture, SG-GAN is well-suited for tackling sparsely grouped learning and multi-task learning. The proposed model can translate images among multiple groups using only a single commonly trained model. To experimentally validate advantages of the new model, we apply the proposed method to tackle a series of attribute manipulation tasks for facial images. Experimental results demonstrate that SG-GAN can generate image translation results of comparable quality with baselines methods on adequately labelled datasets and results of superior quality on sparsely grouped datasets. The official implementation is publicly available:https://github.com/zhangqianhui/Sparsely-Grouped-GAN.
We consider the problem of learning a low-rank matrix, constrained to lie in a linear subspace, and introduce a novel factorization for modeling such matrices. A salient feature of the proposed factorization scheme is it decouples the low-rank and the structural constraints onto separate factors. We formulate the optimization problem on the Riemannian spectrahedron manifold, where the Riemannian framework allows to develop computationally efficient conjugate gradient and trust-region algorithms. Experiments on problems such as standard/robust/non-negative matrix completion, Hankel matrix learning and multi-task learning demonstrate the efficacy of our approach. A shorter version of this work has been published in ICML'18.
Neural text classification models typically treat output labels as categorical variables which lack description and semantics. This forces their parametrization to be dependent on the label set size, and, hence, they are unable to scale to large label sets and generalize to unseen ones. Existing joint input-label text models overcome these issues by exploiting label descriptions, but they are unable to capture complex label relationships, have rigid parametrization, and their gains on unseen labels happen often at the expense of weak performance on the labels seen during training. In this paper, we propose a new input-label model which generalizes over previous such models, addresses their limitations and does not compromise performance on seen labels. The model consists of a joint non-linear input-label embedding with controllable capacity and a joint-space-dependent classification unit which is trained with cross-entropy loss to optimize classification performance. We evaluate models on full-resource and low- or zero-resource text classification of multilingual news and biomedical text with a large label set. Our model outperforms monolingual and multilingual models which do not leverage label semantics and previous joint input-label space models in both scenarios.
We consider the problem of multi-task learning in the high dimensional setting. In particular, we introduce an estimator and investigate its statistical and computational properties for the problem of multiple connected linear regressions known as Data Enrichment/Sharing. The between-tasks connections are captured by a cross-tasks \emph{common parameter}, which gets refined by per-task \emph{individual parameters}. Any convex function, e.g., norm, can characterize the structure of both common and individual parameters. We delineate the sample complexity of our estimator and provide a high probability non-asymptotic bound for estimation error of all parameters under a geometric condition. We show that the recovery of the common parameter benefits from \emph{all} of the pooled samples. We propose an iterative estimation algorithm with a geometric convergence rate and supplement our theoretical analysis with experiments on synthetic data. Overall, we present a first thorough statistical and computational analysis of inference in the data-sharing model.