State-of-the-art Active Learning (AL) methods often encounter challenges associated with a hysteretic learning process and an expensive data sampling mechanism. The former implies that data selection in the ($i+1$)-th round is solely based on the learned model’s results in the$i$-th round. The latter involves using model inference to calculate data value (e.g., uncertainty estimation based on model inference), which can be cumbersome, particularly when working with large datasets or Deep Neural Networks (DNNs). To address these challenges, we propose FastAL, an efficient and dynamic deep AL framework. Our approach includes an efficient method for calculating data value from the frequency domain perspective, generating multiple candidates. Then, we introduce the Fast Evaluation Module, which directly calculates each candidate’s contribution to future model training and selects the best options. In addition, current AL methods, particularly those based on uncertainty, are susceptible to data bias, which implies that selected data may not represent the original unlabeled data adequately. To alleviate this issue, we propose the De-similar Module, which removes partially similar data. The above three modules are model-agnostic and thus can be seamlessly integrated into any Active Learning framework. We conducted rigorous experiments on various benchmark datasets to validate our approach’s effectiveness. Our results demonstrate that FastAL outperforms other state-of-the-art methods by a significant margin, including those based on uncertainty, diversity, and expected model change.

Despite the large-scale adoption of Artificial Intelligence (AI) models in healthcare, there is an urgent need for trustworthy tools to rigorously backtrack the model decisions so that they behave reliably. Counterfactual explanations take a counter-intuitive approach to allow users to explore "what if" scenarios gradually becoming popular in the trustworthy field. However, most previous work on model's counterfactual explanation cannot generate in-distribution attribution credibly, produces adversarial examples, or fails to give a confidence interval for the explanation. Hence, in this paper, we propose a novel approach that generates counterfactuals in locally smooth directed semantic embedding space, and at the same time gives an uncertainty estimate in the counterfactual generation process. Specifically, we identify low-dimensional directed semantic embedding space based on Principal Component Analysis (PCA) applied in differential generative model. Then, we propose latent space smoothing regularization to rectify counterfactual search within in-distribution, such that visually-imperceptible changes are more robust to adversarial perturbations. Moreover, we put forth an uncertainty estimation framework for evaluating counterfactual uncertainty. Extensive experiments on several challenging realistic Chest X-ray and CelebA datasets show that our approach performs consistently well and better than the existing several state-of-the-art baseline approaches.

Source-free domain adaptation (SFDA), where only a pre-trained source model is used to adapt to the target distribution, is a more general approach to achieving domain adaptation in the real world. However, it can be challenging to capture the inherent structure of the target features accurately due to the lack of supervised information on the target domain. By analyzing the clustering performance of the target features, we show that they still contain core features related to discriminative attributes but lack the collation of semantic information. Inspired by this insight, we present Chaos to Order (CtO), a novel approach for SFDA that strives to constrain semantic credibility and propagate label information among target subpopulations. CtO divides the target data into inner and outlier samples based on the adaptive threshold of the learning state, customizing the learning strategy to fit the data properties best. Specifically, inner samples are utilized for learning intra-class structure thanks to their relatively well-clustered properties. The low-density outlier samples are regularized by input consistency to achieve high accuracy with respect to the ground truth labels. In CtO, by employing different learning strategies to propagate the labels from the inner local to outlier instances, it clusters the global samples from chaos to order. We further adaptively regulate the neighborhood affinity of the inner samples to constrain the local semantic credibility. In theoretical and empirical analyses, we demonstrate that our algorithm not only propagates from inner to outlier but also prevents local clustering from forming spurious clusters. Empirical evidence demonstrates that CtO outperforms the state of the arts on three public benchmarks: Office-31, Office-Home, and VisDA.

Deep reinforcement learning (RL) has achieved re-markable performance in sequential decision-making problems. However, it is a challenge for deep RL methods to extract task-relevant semantic information when interacting with limited data from the environment. In this paper, we propose Mix-up Consistent Cross Representations (MCCR), a novel self-supervised auxiliary task, which aims to improve data efficiency and encourage representation prediction. Specifically, we calculate the contrastive loss between low-dimensional and high-dimensional representations of different state observations to boost the mutual information between states, thus improving data efficiency. Furthermore, we employ a mixed strategy to generate intermediate samples, increasing data diversity and the smoothness of representations prediction in nearby timesteps. Experimental results show that MCCR achieves competitive results over the state-of-the-art approaches for complex control tasks in DeepMind Control Suite, notably improving the ability of pretrained encoders to generalize to unseen tasks.

MedMNIST is a medical dataset proposed to block the need for medical knowledge, but there is currently no model that can generalize well on all its sub-datasets. Owing to the inadequacy of long-range relation modeling, models based on convolutional neural networks (CNNs) cannot fully learn the information of images. Besides, relying only on high-level features limits the generalization effect as well. All of these remain challenges for MedMNIST Classification Decathlon. In this paper, we proposed Feature Pyramid Vision Transformer (FPViT), a strong alternative for MedMNIST Classification Decathlon. Our FPViT exhibits enhanced feature learning and modeling capabilities, which merits both residual network (ResNet) and Vision Transformer (ViT). Transformers in our model take the features extracted by ResNet as sequences to capture global contexts which compensate for the lack of locality of convolution operations. Moreover, the feature pyramid designed in our model effectively utilizes the multi-scale feature maps from basic layers of ResNet. These multi-scale features from low-level to high level enable our model to have better adaptability. And, the final prediction is based on the multi-scale ViT and the original ResNet heads. Through experiments, our FPViT can achieve superior classification and generalization on MedMNIST than state-of-the-art methods.

Imporving the generalization ability of an agent is an important and challenging task in deep reinforcement learning (RL). Procedually generated environment is an important benchmark for testing generalization in deep RL. In this benchmark, each game consists of multiple levels, each level is an algorithmically created environment instance with a unique configuration of its factors of variation. Existing methods (e.g., regularization, data augmentation) for improving the generalization of RL agent do not learn well the invariant representation among multiple levels. Besides, existing methods for learning invariant representations in RL using adversarial training can only learn invariant information across two levels. To solve this problem, we propose Adversarial Discriminative Feature Separate (ADFS). First, ADFS design a new discriminator for distinguishing whether two observations belong to the same level. Thus, the policy encoder is encouraged to learn invariant information between multiple levels. Second, it separates the representation of observation into level-invariant features and level-discriminative features, so that correction of the optimization direction of the discriminator. The discriminative features are learned by reducing the similarity of specific features intra-levels and increasing that of inter-levels, respectively. Experimental results demonstrate that our method is quite competitive with existing state-of-the-art methods on Procgen Benchmark.

Data augmentation has been widely introduced into graph-based tasks to improve the generalizability of models. Based on empirical hypothesis, we show that the node in a graph has different importance, the important one is critical for classification task while the unimportant one hurts the performance. However, there are few works in data augmentation addressing the information propagation of nodes with different importance. In this work, we propose a novel graph data augmentation algorithm for graph classification task, called PassAugment, aiming to pass these importance in graph data augmentation. After distinguishing the importance of all nodes in each graph using the saliency map, we design a data augmentation approach including two strategies: (i) randomly adding edges between the important nodes and the other nodes to globally improve the effective information passing, and (ii) randomly removing edges between the unimportant nodes and their neighbors to locally reduce the ineffective information passing. More importantly, our proposed approach as a standalone module can be combined with many GNNs architectures. Experimental results on graph classification task show that our approach consistently improves the accuracy and achieves or closely matches the state-of-the-art performance.

While Deep Neural Networks (DNNs) are achieving state-of-the-art performance on medical domains across a variety of tasks, the need for explainability of model predictions in these high-stakes tasks is still lacking. Current for the explainability in model predictions potentially relies on the supervised counterfactual generation that is time-consuming and direction uncontrollable. Yet, the counterfactual generation needs to be easy to implement and have a controllable direction. In light of this trend, we propose an approach for the unsupervised latent direction search of black-box models that are steerable to the user by enabling the user to effectively explore counterfactual generation in a directional way, without relying on domain- or data-specific assumptions. To identify these explainable directions, we use Principal Component Analysis (PCA), a general manifold learning framework to extract low-dimensional subspaces based on a local noise injection of the pre-trained generative model, so that a small perturbation in the subspaces would provide enough change in the resulting data. With experiments on three real-world CXR datasets involving 6 tasks, we find that our approach is capable of learning explainable predictions that discard unrelated confounding factors. Moreover, our method enables practitioners to edit directions to better understand which features are used for predictions.

Semi-supervised learning has achieved many successes in medical image segmentation since it reduces the costs of manually annotating by leveraging abundant unlabeled data. However, these semi-supervised methods lack attention to ambiguous regions (e.g., some edges or corners around the targets), which may lead to meaningless and unreliable guidance. In this paper, we propose a novel semi-supervised segmentation method called Shape-aware Multi-task Learning (SMTL) to address the above issue. Our multi-task framework includes three tasks namely i) the main task for segmentation ii) one auxiliary task for signed distance regression iii) another auxiliary task for contour detection. The multi-task framework jointly predicts probabilistic segmentation maps, signed distance maps (SDMs) and edge maps to collect complementary information in the existing target label. Specifically, these two auxiliary tasks explicitly enforce shape-priors on the segmentation output to generate more accurate masks. Moreover, we design a region-attention-based adversarial learning strategy that enforces the consistency of two auxiliary tasks prediction distributions on the unlabeled and labeled data to make a meaningful and reliable guidance. We evaluate our SMTL on the datasets of the 2018 Atrial Segmentation Challenge and the 2017 Liver Tumor Segmentation Challenge. The results demonstrate that our SMTL achieves improvements and outperforms the state-of-the-art semi-supervised methods.

Patients infected with COVID-19 can lead to their Chest X-rays (CXRs) with opacifications rendered regions, which may produce incomplete lung segmentation in automated image analysis models. To tackle this issue, we propose a Group structured Conditional Variational data Imputation model to capture the missing data accurately with conditional distribution, where the high-dimensional probability distribution is narrow down to a small latent space to account for unobserved features. This work particularly arises in the fight against COVID-19 that effectively modeling a segmentation of plausible can be presented to a subsequent automated risk scoring and treatment. We train this model with limited CXRs data to demonstrate the abilities on the task of data imputation and proved to be effective though with relatively small datasets.

Lung cancer is the leading cause of cancer deaths worldwide with its mortality rate higher than that of other leading cancers. Nodules in the lungs can grow quite large without any obvious symptoms until the condition has reached a certain stage. Early detection and diagnosis of growing nodules can lay a good foundation for further treatment and potentially improve lung cancer survival rate. In this paper a unified framework is proposed for visual prediction and diagnosis of follow-up lung nodules. Future nodule growth is predicted by modeling the nodule growth between consecutive Computed Tomography (CT) scans via spatial transformation using convolutional network. Nodule classification is made based on the predicted nodule growth and previous diagnosis. Experiments are conducted on a longitudinal follow-up dataset of 615 LDCT scans of 153 lung nodules in early stages from 125 patients. Quantitative and qualitative results demonstrate the effectiveness of the proposed method.

Group equivariant Convolutional Neural Networks (G-CNNs) has led to big empirical success in the medical domain, one fundamental assumption is that equivariance provides a powerful inductive bias for medical images. By leveraging concepts from group representation theory, we can generalize vanilla Convolutional Neural Networks (CNNs) to G-CNN. Currently, although embedding an arbitrary equivariance to CNNs can learn powerful disentangled representations in a higher dimensional domain, they lack explicit means to learn meaningful relationships among the equivariant convolutional kernels. In this paper, we propose a generalization of the dynamic convolutional method, named as dynamic group equivariant convolution, to strengthen the relationships and increase model capability by aggregating multiple group convolutional kernels via attention. Meanwhile, we generalize attention to an equivariant one to preserve equivariant of dynamic group convolution. In our approach, this leads to a flexible framework that enables a dynamic convolutional in G-CNNs by means of a dynamic routing layer expansions. We demonstrate that breast tumor classification is substantial improvements when compared to a recent baseline architecture.

We present a scalable stacking-based deep neural network(S-DNN) for facial expression recognition. The network is a congregate of basic learning models in series to synthesize a deep neural network with feedforward network architecture. Thur, choosing trainable learning modules is the core to effectively build S-DNN in an end-to-end manner. Inspired by the manifold learning archetype, we implement a Patch Discriminative Analysis(PDA) as a basic learning model, followed by hashing and block histogram on the top, which sample image in a low discriminative space, and finding an efficient representation of the training data. As those self-learnable models trained, a low dimensional discriminative feature is implicitly learned, which proves to be useful in facial expression recognition. Experimental results on the facial expression dataset(CK+) show that the proposed model is superior to its counterparts, capable of achieving state-of-the-art performance.