Xingchen Zhao

Xu Ma, Huan Wang, Can Qin et al.

Vision Transformers have shown great promise recently for many vision tasks due to the insightful architecture design and attention mechanism. By revisiting the self-attention responses in Transformers, we empirically observe two interesting issues. First, Vision Transformers present a queryirrelevant behavior at deep layers, where the attention maps exhibit nearly consistent contexts in global scope, regardless of the query patch position (also head-irrelevant). Second, the attention maps are intrinsically sparse, few tokens dominate the attention weights; introducing the knowledge from ConvNets would largely smooth the attention and enhance the performance. Motivated by above observations, we generalize self-attention formulation to abstract a queryirrelevant global context directly and further integrate the global context into convolutions. The resulting model, a Fully Convolutional Vision Transformer (i.e., FCViT), purely consists of convolutional layers and firmly inherits the merits of both attention mechanism and convolutions, including dynamic property, weight sharing, and short- and long-range feature modeling, etc. Experimental results demonstrate the effectiveness of FCViT. With less than 14M parameters, our FCViT-S12 outperforms related work ResT-Lite by 3.7% top1 accuracy on ImageNet-1K. When scaling FCViT to larger models, we still perform better than previous state-of-the-art ConvNeXt with even fewer parameters. FCViT-based models also demonstrate promising transferability to downstream tasks, like object detection, instance segmentation, and semantic segmentation. Codes and models are made available at: https://github.com/ma-xu/FCViT.

Xingchen Zhao, Niluthpol Chowdhury Mithun, Abhinav Rajvanshi et al.

Deep learning-based solutions for semantic segmentation suffer from significant performance degradation when tested on data with different characteristics than what was used during the training. Adapting the models using annotated data from the new domain is not always practical. Unsupervised Domain Adaptation (UDA) approaches are crucial in deploying these models in the actual operating conditions. Recent state-of-the-art (SOTA) UDA methods employ a teacher-student self-training approach, where a teacher model is used to generate pseudo-labels for the new data which in turn guide the training process of the student model. Though this approach has seen a lot of success, it suffers from the issue of noisy pseudo-labels being propagated in the training process. To address this issue, we propose an auxiliary pseudo-label refinement network (PRN) for online refining of the pseudo labels and also localizing the pixels whose predicted labels are likely to be noisy. Being able to improve the quality of pseudo labels and select highly reliable ones, PRN helps self-training of segmentation models to be robust against pseudo label noise propagation during different stages of adaptation. We evaluate our approach on benchmark datasets with three different domain shifts, and our approach consistently performs significantly better than the previous state-of-the-art methods.

Xingchen Zhao, Chang Liu, Anthony Sicilia et al.

The topic of generalizing machine learning models learned on a collection of source domains to unknown target domains is challenging. While many domain generalization (DG) methods have achieved promising results, they primarily rely on the source domains at train-time without manipulating the target domains at test-time. Thus, it is still possible that those methods can overfit to source domains and perform poorly on target domains. Driven by the observation that domains are strongly related to styles, we argue that reducing the gap between source and target styles can boost models' generalizability. To solve the dilemma of having no access to the target domain during training, we introduce Test-time Fourier Style Calibration (TF-Cal) for calibrating the target domain style on the fly during testing. To access styles, we utilize Fourier transformation to decompose features into amplitude (style) features and phase (semantic) features. Furthermore, we present an effective technique to Augment Amplitude Features (AAF) to complement TF-Cal. Extensive experiments on several popular DG benchmarks and a segmentation dataset for medical images demonstrate that our method outperforms state-of-the-art methods.

Dabing Cheng, Haosen Zhan, Xingchen Zhao et al.

The exponential growth of short-video content has ignited a surge in the necessity for efficient, automated solutions to video editing, with challenges arising from the need to understand videos and tailor the editing according to user requirements. Addressing this need, we propose an innovative end-to-end foundational framework, ultimately actualizing precise control over the final video content editing. Leveraging the flexibility and generalizability of Multimodal Large Language Models (MLLMs), we defined clear input-output mappings for efficient video creation. To bolster the model's capability in processing and comprehending video content, we introduce a strategic combination of a denser frame rate and a slow-fast processing technique, significantly enhancing the extraction and understanding of both temporal and spatial video information. Furthermore, we introduce a text-to-edit mechanism that allows users to achieve desired video outcomes through textual input, thereby enhancing the quality and controllability of the edited videos. Through comprehensive experimentation, our method has not only showcased significant effectiveness within advertising datasets, but also yields universally applicable conclusions on public datasets.

Anthony Sicilia, Xingchen Zhao, Anastasia Sosnovskikh et al.

Application of deep neural networks to medical imaging tasks has in some sense become commonplace. Still, a "thorn in the side" of the deep learning movement is the argument that deep networks are prone to overfitting and are thus unable to generalize well when datasets are small (as is common in medical imaging tasks). One way to bolster confidence is to provide mathematical guarantees, or bounds, on network performance after training which explicitly quantify the possibility of overfitting. In this work, we explore recent advances using the PAC-Bayesian framework to provide bounds on generalization error for large (stochastic) networks. While previous efforts focus on classification in larger natural image datasets (e.g., MNIST and CIFAR-10), we apply these techniques to both classification and segmentation in a smaller medical imagining dataset: the ISIC 2018 challenge set. We observe the resultant bounds are competitive compared to a simpler baseline, while also being more explainable and alleviating the need for holdout sets.

Anthony Sicilia, Xingchen Zhao, Davneet Minhas et al.

We consider a model-agnostic solution to the problem of Multi-Domain Learning (MDL) for multi-modal applications. Many existing MDL techniques are model-dependent solutions which explicitly require nontrivial architectural changes to construct domain-specific modules. Thus, properly applying these MDL techniques for new problems with well-established models, e.g. U-Net for semantic segmentation, may demand various low-level implementation efforts. In this paper, given emerging multi-modal data (e.g., various structural neuroimaging modalities), we aim to enable MDL purely algorithmically so that widely used neural networks can trivially achieve MDL in a model-independent manner. To this end, we consider a weighted loss function and extend it to an effective procedure by employing techniques from the recently active area of learning-to-learn (meta-learning). Specifically, we take inner-loop gradient steps to dynamically estimate posterior distributions over the hyperparameters of our loss function. Thus, our method is model-agnostic, requiring no additional model parameters and no network architecture changes; instead, only a few efficient algorithmic modifications are needed to improve performance in MDL. We demonstrate our solution to a fitting problem in medical imaging, specifically, in the automatic segmentation of white matter hyperintensity (WMH). We look at two neuroimaging modalities (T1-MR and FLAIR) with complementary information fitting for our problem.

Xingchen Zhao, Anthony Sicilia, Davneet Minhas et al.

Typical machine learning frameworks heavily rely on an underlying assumption that training and test data follow the same distribution. In medical imaging which increasingly begun acquiring datasets from multiple sites or scanners, this identical distribution assumption often fails to hold due to systematic variability induced by site or scanner dependent factors. Therefore, we cannot simply expect a model trained on a given dataset to consistently work well, or generalize, on a dataset from another distribution. In this work, we address this problem, investigating the application of machine learning models to unseen medical imaging data. Specifically, we consider the challenging case of Domain Generalization (DG) where we train a model without any knowledge about the testing distribution. That is, we train on samples from a set of distributions (sources) and test on samples from a new, unseen distribution (target). We focus on the task of white matter hyperintensity (WMH) prediction using the multi-site WMH Segmentation Challenge dataset and our local in-house dataset. We identify how two mechanically distinct DG approaches, namely domain adversarial learning and mix-up, have theoretical synergy. Then, we show drastic improvements of WMH prediction on an unseen target domain.

Anthony Sicilia, Xingchen Zhao, Seong Jae Hwang

Theoretically, domain adaptation is a well-researched problem. Further, this theory has been well-used in practice. In particular, we note the bound on target error given by Ben-David et al. (2010) and the well-known domain-aligning algorithm based on this work using Domain Adversarial Neural Networks (DANN) presented by Ganin and Lempitsky (2015). Recently, multiple variants of DANN have been proposed for the related problem of domain generalization, but without much discussion of the original motivating bound. In this paper, we investigate the validity of DANN in domain generalization from this perspective. We investigate conditions under which application of DANN makes sense and further consider DANN as a dynamic process during training. Our investigation suggests that the application of DANN to domain generalization may not be as straightforward as it seems. To address this, we design an algorithmic extension to DANN in the domain generalization case. Our experimentation validates both theory and algorithm.

Xingchen Zhao, Xuehai He, Pengtao Xie

Learning by ignoring, which identifies less important things and excludes them from the learning process, is broadly practiced in human learning and has shown ubiquitous effectiveness. There has been psychological studies showing that learning to ignore certain things is a powerful tool for helping people focus. In this paper, we explore whether this useful human learning methodology can be borrowed to improve machine learning. We propose a novel machine learning framework referred to as learning by ignoring (LBI). Our framework automatically identifies pretraining data examples that have large domain shift from the target distribution by learning an ignoring variable for each example and excludes them from the pretraining process. We formulate LBI as a three-level optimization framework where three learning stages are involved: pretraining by minimizing the losses weighed by ignoring variables; finetuning; updating the ignoring variables by minimizing the validation loss. A gradient-based algorithm is developed to efficiently solve the three-level optimization problem in LBI. Experiments on various datasets demonstrate the effectiveness of our framework.