arXiv subjects

Dajiang Zhu

Publications and source records attributed to Dajiang Zhu.

82 records · Page 5Linked to original sources

arXiv2022

Rectify ViT Shortcut Learning by Visual Saliency

Shortcut learning is common but harmful to deep learning models, leading to degenerated feature representations and consequently jeopardizing the model's generalizability and interpretability. However, shortcut learning in the widely used Vision Transformer framework is largely unknown. Meanwhile, introducing domain-specific knowledge is a major approach to rectifying the shortcuts, which are predominated by background related factors. For example, in the medical imaging field, eye-gaze data from radiologists is an effective human visual prior knowledge that has the great potential to guide the deep learning models to focus on meaningful foreground regions of interest. However, obtaining eye-gaze data is time-consuming, labor-intensive and sometimes even not practical. In this work, we propose a novel and effective saliency-guided vision transformer (SGT) model to rectify shortcut learning in ViT with the absence of eye-gaze data. Specifically, a computational visual saliency model is adopted to predict saliency maps for input image samples. Then, the saliency maps are used to distil the most informative image patches. In the proposed SGT, the self-attention among image patches focus only on the distilled informative ones. Considering this distill operation may lead to global information lost, we further introduce, in the last encoder layer, a residual connection that captures the self-attention across all the image patches. The experiment results on four independent public datasets show that our SGT framework can effectively learn and leverage human prior knowledge without eye gaze data and achieves much better performance than baselines. Meanwhile, it successfully rectifies the harmful shortcut learning and significantly improves the interpretability of the ViT model, demonstrating the promise of transferring human prior knowledge derived visual saliency in rectifying shortcut learning

cs.CV↗

arXiv2022

Discovering Dynamic Functional Brain Networks via Spatial and Channel-wise Attention

Using deep learning models to recognize functional brain networks (FBNs) in functional magnetic resonance imaging (fMRI) has been attracting increasing interest recently. However, most existing work focuses on detecting static FBNs from entire fMRI signals, such as correlation-based functional connectivity. Sliding-window is a widely used strategy to capture the dynamics of FBNs, but it is still limited in representing intrinsic functional interactive dynamics at each time step. And the number of FBNs usually need to be set manually. More over, due to the complexity of dynamic interactions in brain, traditional linear and shallow models are insufficient in identifying complex and spatially overlapped FBNs across each time step. In this paper, we propose a novel Spatial and Channel-wise Attention Autoencoder (SCAAE) for discovering FBNs dynamically. The core idea of SCAAE is to apply attention mechanism to FBNs construction. Specifically, we designed two attention modules: 1) spatial-wise attention (SA) module to discover FBNs in the spatial domain and 2) a channel-wise attention (CA) module to weigh the channels for selecting the FBNs automatically. We evaluated our approach on ADHD200 dataset and our results indicate that the proposed SCAAE method can effectively recover the dynamic changes of the FBNs at each fMRI time step, without using sliding windows. More importantly, our proposed hybrid attention modules (SA and CA) do not enforce assumptions of linearity and independence as previous methods, and thus provide a novel approach to better understanding dynamic functional brain networks.

cs.CV↗

arXiv2022

Representing Brain Anatomical Regularity and Variability by Few-Shot Embedding

Effective representation of brain anatomical architecture is fundamental in understanding brain regularity and variability. Despite numerous efforts, it is still difficult to infer reliable anatomical correspondence at finer scale, given the tremendous individual variability in cortical folding patterns. It is even more challenging to disentangle common and individual patterns when comparing brains at different neuro-developmental stages. In this work, we developed a novel learning-based few-shot embedding framework to encode the cortical folding patterns into a latent space represented by a group of anatomically meaningful embedding vectors. Specifically, we adopted 3-hinge (3HG) network as the substrate and designed an autoencoder-based embedding framework to learn a common embedding vector for each 3HG's multi-hop feature: each 3HG can be represented as a combination of these feature embeddings via a set of individual specific coefficients to characterize individualized anatomical information. That is, the regularity of folding patterns is encoded into the embeddings, while the individual variations are preserved by the multi=hop combination coefficients. To effectively learn the embeddings for the population with very limited samples, few-shot learning was adopted. We applied our method on adult HCP and pediatric datasets with 1,000+ brains (from 34 gestational weeks to young adult). Our experimental results show that: 1) the learned embedding vectors can quantitatively encode the commonality and individuality of cortical folding patterns; 2) with the embeddings we can robustly infer the complicated many-to-many anatomical correspondences among different brains and 3) our model can be successfully transferred to new populations with very limited training samples.

q-bio.NC↗

arXiv2022

Eye-gaze-guided Vision Transformer for Rectifying Shortcut Learning

Learning harmful shortcuts such as spurious correlations and biases prevents deep neural networks from learning the meaningful and useful representations, thus jeopardizing the generalizability and interpretability of the learned representation. The situation becomes even more serious in medical imaging, where the clinical data (e.g., MR images with pathology) are limited and scarce while the reliability, generalizability and transparency of the learned model are highly required. To address this problem, we propose to infuse human experts' intelligence and domain knowledge into the training of deep neural networks. The core idea is that we infuse the visual attention information from expert radiologists to proactively guide the deep model to focus on regions with potential pathology and avoid being trapped in learning harmful shortcuts. To do so, we propose a novel eye-gaze-guided vision transformer (EG-ViT) for diagnosis with limited medical image data. We mask the input image patches that are out of the radiologists' interest and add an additional residual connection in the last encoder layer of EG-ViT to maintain the correlations of all patches. The experiments on two public datasets of INbreast and SIIM-ACR demonstrate our EG-ViT model can effectively learn/transfer experts' domain knowledge and achieve much better performance than baselines. Meanwhile, it successfully rectifies the harmful shortcut learning and significantly improves the EG-ViT model's interpretability. In general, EG-ViT takes the advantages of both human expert's prior knowledge and the power of deep neural networks. This work opens new avenues for advancing current artificial intelligence paradigms by infusing human intelligence.

cs.CV↗

arXiv2022

Brain Cortical Functional Gradients Predict Cortical Folding Patterns via Attention Mesh Convolution

Since gyri and sulci, two basic anatomical building blocks of cortical folding patterns, were suggested to bear different functional roles, a precise mapping from brain function to gyro-sulcal patterns can provide profound insights into both biological and artificial neural networks. However, there lacks a generic theory and effective computational model so far, due to the highly nonlinear relation between them, huge inter-individual variabilities and a sophisticated description of brain function regions/networks distribution as mosaics, such that spatial patterning of them has not been considered. we adopted brain functional gradients derived from resting-state fMRI to embed the "gradual" change of functional connectivity patterns, and developed a novel attention mesh convolution model to predict cortical gyro-sulcal segmentation maps on individual brains. The convolution on mesh considers the spatial organization of functional gradients and folding patterns on a cortical sheet and the newly designed channel attention block enhances the interpretability of the contribution of different functional gradients to cortical folding prediction. Experiments show that the prediction performance via our model outperforms other state-of-the-art models. In addition, we found that the dominant functional gradients contribute less to folding prediction. On the activation maps of the last layer, some well-studied cortical landmarks are found on the borders of, rather than within, the highly activated regions. These results and findings suggest that a specifically designed artificial neural network can improve the precision of the mapping between brain functions and cortical folding patterns, and can provide valuable insight of brain anatomy-function relation for neuroscience.

q-bio.NC↗

arXiv2022

Mask-guided Vision Transformer (MG-ViT) for Few-Shot Learning

Learning with little data is challenging but often inevitable in various application scenarios where the labeled data is limited and costly. Recently, few-shot learning (FSL) gained increasing attention because of its generalizability of prior knowledge to new tasks that contain only a few samples. However, for data-intensive models such as vision transformer (ViT), current fine-tuning based FSL approaches are inefficient in knowledge generalization and thus degenerate the downstream task performances. In this paper, we propose a novel mask-guided vision transformer (MG-ViT) to achieve an effective and efficient FSL on ViT model. The key idea is to apply a mask on image patches to screen out the task-irrelevant ones and to guide the ViT to focus on task-relevant and discriminative patches during FSL. Particularly, MG-ViT only introduces an additional mask operation and a residual connection, enabling the inheritance of parameters from pre-trained ViT without any other cost. To optimally select representative few-shot samples, we also include an active learning based sample selection method to further improve the generalizability of MG-ViT based FSL. We evaluate the proposed MG-ViT on both Agri-ImageNet classification task and ACFR apple detection task with gradient-weighted class activation mapping (Grad-CAM) as the mask. The experimental results show that the MG-ViT model significantly improves the performance when compared with general fine-tuning based ViT models, providing novel insights and a concrete approach towards generalizing data-intensive and large-scale deep learning models for FSL.

cs.CV↗

arXiv2022

A Unified and Biologically-Plausible Relational Graph Representation of Vision Transformers

Vision transformer (ViT) and its variants have achieved remarkable successes in various visual tasks. The key characteristic of these ViT models is to adopt different aggregation strategies of spatial patch information within the artificial neural networks (ANNs). However, there is still a key lack of unified representation of different ViT architectures for systematic understanding and assessment of model representation performance. Moreover, how those well-performing ViT ANNs are similar to real biological neural networks (BNNs) is largely unexplored. To answer these fundamental questions, we, for the first time, propose a unified and biologically-plausible relational graph representation of ViT models. Specifically, the proposed relational graph representation consists of two key sub-graphs: aggregation graph and affine graph. The former one considers ViT tokens as nodes and describes their spatial interaction, while the latter one regards network channels as nodes and reflects the information communication between channels. Using this unified relational graph representation, we found that: a) a sweet spot of the aggregation graph leads to ViTs with significantly improved predictive performance; b) the graph measures of clustering coefficient and average path length are two effective indicators of model prediction performance, especially when applying on the datasets with small samples; c) our findings are consistent across various ViT architectures and multiple datasets; d) the proposed relational graph representation of ViT has high similarity with real BNNs derived from brain science data. Overall, our work provides a novel unified and biologically-plausible paradigm for more interpretable and effective representation of ViT ANNs.

cs.NE↗

arXiv2021

Representative Functional Connectivity Learning for Multiple Clinical groups in Alzheimer's Disease

Mild cognitive impairment (MCI) is a high-risk dementia condition which progresses to probable Alzheimer's disease (AD) at approximately 10% to 15% per year. Characterization of group-level differences between two subtypes of MCI - stable MCI (sMCI) and progressive MCI (pMCI) is the key step to understand the mechanisms of MCI progression and enable possible delay of transition from MCI to AD. Functional connectivity (FC) is considered as a promising way to study MCI progression since which may show alterations even in preclinical stages and provide substrates for AD progression. However, the representative FC patterns during AD development for different clinical groups, especially for sMCI and pMCI, have been understudied. In this work, we integrated autoencoder and multi-class classification into a single deep model and successfully learned a set of clinical group related feature vectors. Specifically, we trained two non-linear mappings which realized the mutual transformations between original FC space and the feature space. By mapping the learned clinical group related feature vectors to the original FC space, representative FCs were constructed for each group. Moreover, based on these feature vectors, our model achieves a high classification accuracy - 68% for multi-class classification (NC vs SMC vs sMCI vs pMCI vs AD).

q-bio.NC↗

arXiv2017

Classification of Major Depressive Disorder via Multi-Site Weighted LASSO Model

Large-scale collaborative analysis of brain imaging data, in psychiatry and neu-rology, offers a new source of statistical power to discover features that boost ac-curacy in disease classification, differential diagnosis, and outcome prediction. However, due to data privacy regulations or limited accessibility to large datasets across the world, it is challenging to efficiently integrate distributed information. Here we propose a novel classification framework through multi-site weighted LASSO: each site performs an iterative weighted LASSO for feature selection separately. Within each iteration, the classification result and the selected features are collected to update the weighting parameters for each feature. This new weight is used to guide the LASSO process at the next iteration. Only the fea-tures that help to improve the classification accuracy are preserved. In tests on da-ta from five sites (299 patients with major depressive disorder (MDD) and 258 normal controls), our method boosted classification accuracy for MDD by 4.9% on average. This result shows the potential of the proposed new strategy as an ef-fective and practical collaborative platform for machine learning on large scale distributed imaging and biobank data.

cs.LG↗

arXiv2017

Large-scale Feature Selection of Risk Genetic Factors for Alzheimer's Disease via Distributed Group Lasso Regression

Genome-wide association studies (GWAS) have achieved great success in the genetic study of Alzheimer's disease (AD). Collaborative imaging genetics studies across different research institutions show the effectiveness of detecting genetic risk factors. However, the high dimensionality of GWAS data poses significant challenges in detecting risk SNPs for AD. Selecting relevant features is crucial in predicting the response variable. In this study, we propose a novel Distributed Feature Selection Framework (DFSF) to conduct the large-scale imaging genetics studies across multiple institutions. To speed up the learning process, we propose a family of distributed group Lasso screening rules to identify irrelevant features and remove them from the optimization. Then we select the relevant group features by performing the group Lasso feature selection process in a sequence of parameters. Finally, we employ the stability selection to rank the top risk SNPs that might help detect the early stage of AD. To the best of our knowledge, this is the first distributed feature selection model integrated with group Lasso feature selection as well as detecting the risk genetic factors across multiple research institutions system. Empirical studies are conducted on 809 subjects with 5.9 million SNPs which are distributed across several individual institutions, demonstrating the efficiency and effectiveness of the proposed method.

cs.LG↗