Confidence May Cheat: Self-Training on Graph Neural Networks under Distribution Shift

Graph Convolutional Networks (GCNs) have recently attracted vast interest and achieved state-of-the-art performance on graphs, but its success could typically hinge on careful training with amounts of expensive and time-consuming labeled data. To alleviate labeled data scarcity, self-training methods have been widely adopted on graphs by labeling high-confidence unlabeled nodes and then adding them to the training step. In this line, we empirically make a thorough study for current self-training methods on graphs. Surprisingly, we find that high-confidence unlabeled nodes are not always useful, and even introduce the distribution shift issue between the original labeled dataset and the augmented dataset by self-training, severely hindering the capability of self-training on graphs. To this end, in this paper, we propose a novel Distribution Recovered Graph Self-Training framework (DR-GST), which could recover the distribution of the original labeled dataset. Specifically, we first prove the equality of loss function in self-training framework under the distribution shift case and the population distribution if each pseudo-labeled node is weighted by a proper coefficient. Considering the intractability of the coefficient, we then propose to replace the coefficient with the information gain after observing the same changing trend between them, where information gain is respectively estimated via both dropout variational inference and dropedge variational inference in DR-GST. However, such a weighted loss function will enlarge the impact of incorrect pseudo labels. As a result, we apply the loss correction method to improve the quality of pseudo labels. Both our theoretical analysis and extensive experiments on five benchmark datasets demonstrate the effectiveness of the proposed DR-GST, as well as each well-designed component in DR-GST.

Debiased Graph Neural Networks with Agnostic Label Selection Bias

Most existing Graph Neural Networks (GNNs) are proposed without considering the selection bias in data, i.e., the inconsistent distribution between the training set with test set. In reality, the test data is not even available during the training process, making selection bias agnostic. Training GNNs with biased selected nodes leads to significant parameter estimation bias and greatly impacts the generalization ability on test nodes. In this paper, we first present an experimental investigation, which clearly shows that the selection bias drastically hinders the generalization ability of GNNs, and theoretically prove that the selection bias will cause the biased estimation on GNN parameters. Then to remove the bias in GNN estimation, we propose a novel Debiased Graph Neural Networks (DGNN) with a differentiated decorrelation regularizer. The differentiated decorrelation regularizer estimates a sample weight for each labeled node such that the spurious correlation of learned embeddings could be eliminated. We analyze the regularizer in causal view and it motivates us to differentiate the weights of the variables based on their contribution on the confounding bias. Then, these sample weights are used for reweighting GNNs to eliminate the estimation bias, thus help to improve the stability of prediction on unknown test nodes. Comprehensive experiments are conducted on several challenging graph datasets with two kinds of label selection biases. The results well verify that our proposed model outperforms the state-of-the-art methods and DGNN is a flexible framework to enhance existing GNNs.

Event-based Video Reconstruction via Potential-assisted Spiking Neural Network

Neuromorphic vision sensor is a new bio-inspired imaging paradigm that reports asynchronous, continuously per-pixel brightness changes called `events' with high temporal resolution and high dynamic range. So far, the event-based image reconstruction methods are based on artificial neural networks (ANN) or hand-crafted spatiotemporal smoothing techniques. In this paper, we first implement the image reconstruction work via fully spiking neural network (SNN) architecture. As the bio-inspired neural networks, SNNs operating with asynchronous binary spikes distributed over time, can potentially lead to greater computational efficiency on event-driven hardware. We propose a novel Event-based Video reconstruction framework based on a fully Spiking Neural Network (EVSNN), which utilizes Leaky-Integrate-and-Fire (LIF) neuron and Membrane Potential (MP) neuron. We find that the spiking neurons have the potential to store useful temporal information (memory) to complete such time-dependent tasks. Furthermore, to better utilize the temporal information, we propose a hybrid potential-assisted framework (PA-EVSNN) using the membrane potential of spiking neuron. The proposed neuron is referred as Adaptive Membrane Potential (AMP) neuron, which adaptively updates the membrane potential according to the input spikes. The experimental results demonstrate that our models achieve comparable performance to ANN-based models on IJRR, MVSEC, and HQF datasets. The energy consumptions of EVSNN and PA-EVSNN are 19.36$\times$ and 7.75$\times$ more computationally efficient than their ANN architectures, respectively.

Compact Graph Structure Learning via Mutual Information Compression

Graph Structure Learning (GSL) recently has attracted considerable attentions in its capacity of optimizing graph structure as well as learning suitable parameters of Graph Neural Networks (GNNs) simultaneously. Current GSL methods mainly learn an optimal graph structure (final view) from single or multiple information sources (basic views), however the theoretical guidance on what is the optimal graph structure is still unexplored. In essence, an optimal graph structure should only contain the information about tasks while compress redundant noise as much as possible, which is defined as "minimal sufficient structure", so as to maintain the accurancy and robustness. How to obtain such structure in a principled way? In this paper, we theoretically prove that if we optimize basic views and final view based on mutual information, and keep their performance on labels simultaneously, the final view will be a minimal sufficient structure. With this guidance, we propose a Compact GSL architecture by MI compression, named CoGSL. Specifically, two basic views are extracted from original graph as two inputs of the model, which are refinedly reestimated by a view estimator. Then, we propose an adaptive technique to fuse estimated views into the final view. Furthermore, we maintain the performance of estimated views and the final view and reduce the mutual information of every two views. To comprehensively evaluate the performance of CoGSL, we conduct extensive experiments on several datasets under clean and attacked conditions, which demonstrate the effectiveness and robustness of CoGSL.

MutualFormer: Multi-Modality Representation Learning via Mutual Transformer

Aggregating multi-modality data to obtain accurate and reliable data representation attracts more and more attention. The pristine researchers generally adopt the CNN to extract features of independent modality and aggregate them with a fusion module. However, the overall performance is becoming saturated due to limited local convolutional features. Recent studies demonstrate that Transformer models usually work comparable or even better than CNN for multi-modality task, but they simply adopt concatenation or cross-attention for feature fusion which may just obtain sub-optimal results. In this work, we re-thinking the self-attention based Transformer and propose a novel MutualFormer for multi-modality data fusion and representation. The core of MutualFormer is the design of both token mixer and modality mixer to conduct the communication among both tokens and modalities. Specifically, it contains three main modules, i.e., i) Self-attention (SA) for intra-modality token mixer, ii) Cross-diffusion attention (CDA) for inter-modality mixer and iii) Aggregation module. The main advantage of the proposed CDA is that it is defined based on individual domain similarities in the metric space which thus can naturally avoid the issue of domain/modality gap in cross-modality similarities computation. We successfully apply the MutualFormer to the saliency detection problem and propose a novel approach to obtain the reinforced features of RGB and Depth images. Extensive experiments on six popular datasets demonstrate that our model achieves comparable results with 16 SOTA models.

MTFNet: Mutual-Transformer Fusion Network for RGB-D Salient Object Detection

Salient object detection (SOD) on RGB-D images is an active problem in computer vision. The main challenges for RGB-D SOD problem are how to 1) extract the accurate features for RGB and Depth image data with clutter background or poor image quality and 2) explore the complementary information between RGB and Depth image data. To address these challenges, we propose a novel Mutual-Transformer Fusion Network (MTFNet) for RGB-D SOD. MTFNet contains two main modules, $i.e.$, Focal Feature Extractor (FFE) and Mutual-Transformer Fusion (MTF). FFE aims to extract the more accurate CNN features for RGB and Depth images by introducing a novel pixel-level focal regularization to guide CNN feature extractor. MTF is designed to deeply exploit the multi-modal interaction between RGB and Depth images on both coarse and fine scales. The main benefit of MTF is that it conducts the learning of intra-modality and inter-modality simultaneously and thus can achieve communication across different modalities more directly and sufficiently. Comprehensive experimental results on six public benchmarks demonstrate the superiority of our proposed MTFNet.

Generalizing Graph Neural Networks on Out-Of-Distribution Graphs

Graph Neural Networks (GNNs) are proposed without considering the agnostic distribution shifts between training and testing graphs, inducing the degeneration of the generalization ability of GNNs on Out-Of-Distribution (OOD) settings. The fundamental reason for such degeneration is that most GNNs are developed based on the I.I.D hypothesis. In such a setting, GNNs tend to exploit subtle statistical correlations existing in the training set for predictions, even though it is a spurious correlation. However, such spurious correlations may change in testing environments, leading to the failure of GNNs. Therefore, eliminating the impact of spurious correlations is crucial for stable GNNs. To this end, we propose a general causal representation framework, called StableGNN. The main idea is to extract high-level representations from graph data first and resort to the distinguishing ability of causal inference to help the model get rid of spurious correlations. Particularly, we exploit a graph pooling layer to extract subgraph-based representations as high-level representations. Furthermore, we propose a causal variable distinguishing regularizer to correct the biased training distribution. Hence, GNNs would concentrate more on the stable correlations. Extensive experiments on both synthetic and real-world OOD graph datasets well verify the effectiveness, flexibility and interpretability of the proposed framework.

LiT: Zero-Shot Transfer with Locked-image Text Tuning

This paper presents contrastive-tuning, a simple method employing contrastive training to align image and text models while still taking advantage of their pre-training. In our empirical study we find that locked pre-trained image models with unlocked text models work best. We call this instance of contrastive-tuning "Locked-image Text tuning" (LiT-tuning), which just teaches a text model to read out good representations from a pre-trained image model for new tasks. A LiT-tuned model gains the capability of zero-shot transfer to new vision tasks, such as image classification or retrieval. The proposed LiT-tuning is widely applicable; it works reliably with multiple pre-training methods (supervised and unsupervised) and across diverse architectures (ResNet, Vision Transformers and MLP-Mixer) using three different image-text datasets. With the transformer-based pre-trained ViT-g/14 model, the LiT-tuned model achieves 84.5% zero-shot transfer accuracy on the ImageNet test set, and 81.1% on the challenging out-of-distribution ObjectNet test set.