AAAI 2023 Papers

Knowledge-aware question answering (KAQA) requires the model to answer questions over a knowledge base, which is essential for both open-domain QA and domain-specific QA, especially when language models alone cannot provide all the knowledge needed. Despite the promising result of recent KAQA systems which tend to integrate linguistic knowledge from pre-trained language models (PLM) and factual knowledge from knowledge graphs (KG) to answer complex questions, a bottleneck exists in effectively fusing the representations from PLMs and KGs because of (i) the semantic and distributional gaps between them, and (ii) the difficulties in joint reasoning over the provided knowledge from both modalities. To address the above two problems, we propose a Fine-grained Two-stage training framework (FiTs) to boost the KAQA system performance: The first stage aims at aligning representations from the PLM and the KG, thus bridging the modality gaps between them, named knowledge adaptive post-training. The second stage, called knowledge-aware fine-tuning, aims to improve the model's joint reasoning ability based on the aligned representations. In detail, we fine-tune the post-trained model via two auxiliary self-supervised tasks in addition to the QA supervision. Extensive experiments demonstrate that our approach achieves state-of-the-art performance on three benchmarks in the commonsense reasoning (i.e., CommonsenseQA, OpenbookQA) and medical question answering (i.e., MedQA-USMILE) domains.

As few-shot object detectors are often trained with abundant base samples and fine-tuned on few-shot novel examples, the learned models are usually biased to base classes and sensitive to the variance of novel examples. To address this issue, we propose a meta-learning framework with two novel feature aggregation schemes. More precisely, we first present a Class-Agnostic Aggregation (CAA) method, where the query and support features can be aggregated regardless of their categories. The interactions between different classes encourage class-agnostic representations and reduce confusion between base and novel classes. Based on the CAA, we then propose a Variational Feature Aggregation (VFA) method, which encodes support examples into class-level support features for robust feature aggregation. We use a variational autoencoder to estimate class distributions and sample variational features from distributions that are more robust to the variance of support examples. Besides, we decouple classification and regression tasks so that VFA is performed on the classification branch without affecting object localization. Extensive experiments on PASCAL VOC and COCO demonstrate that our method significantly outperforms a strong baseline (up to 16%) and previous state-of-the-art methods (4% in average).

The performances of defect inspection have been severely hindered by insufficient defect images in industries, which can be alleviated by generating more samples as data augmentation. We propose the first defect image generation method in the challenging few-shot cases. Given just a handful of defect images and relatively more defect-free ones, our goal is to augment the dataset with new defect images. Our method consists of two training stages. First, we train a data-efficient StyleGAN2 on defect-free images as the backbone. Second, we attach defect-aware residual blocks to the backbone, which learn to produce reasonable defect masks and accordingly manipulate the features within the masked regions by training the added modules on limited defect images. Extensive experiments on MVTec AD dataset not only validate the effectiveness of our method in generating realistic and diverse defect images, but also manifest the benefits it brings to downstream defect inspection tasks. Codes are available at https://github.com/Ldhlwh/DFMGAN.

We study the problem of composition learning for image retrieval, for which we learn to retrieve target images with search queries in the form of a composition of a reference image and a modification text that describes desired modifications of the image. Existing models of composition learning for image retrieval are generally built with large-scale datasets, demanding extensive training samples, i.e., query-target pairs, as supervision, which restricts their application for the scenario of few-shot learning with only few query-target pairs available. Recently, prompt tuning with frozen pretrained language models has shown remarkable performance when the amount of training data is limited. Inspired by this, we propose a prompt tuning mechanism with the pretrained CLIP model for the task of few-shot composition learning for image retrieval. Specifically, we regard the representation of the reference image as a trainable visual prompt, prefixed to the embedding of the text sequence. One challenge is to efficiently train visual prompt with few-shot samples. To deal with this issue, we further propose a self-upervised auxiliary task via ensuring that the reference image can retrieve itself when no modification information is given from the text, which facilitates training for the visual prompt, while not requiring additional annotations for query-target pairs. Experiments on multiple benchmarks show that our proposed model can yield superior performance when trained with only few query-target pairs.

3D point cloud semantic segmentation aims to group all points into different semantic categories, which benefits important applications such as point cloud scene reconstruction and understanding. Existing supervised point cloud semantic segmentation methods usually require large-scale annotated point clouds for training and cannot handle new categories. While a few-shot learning method was proposed recently to address these two problems, it suffers from high computational complexity caused by graph construction and inability to learn fine-grained relationships among points due to the use of pooling operations. In this paper, we further address these problems by developing a new multi-layer transformer network for few-shot point cloud semantic segmentation. In the proposed network, the query point cloud features are aggregated based on the class-specific support features in different scales. Without using pooling operations, our method makes full use of all pixel-level features from the support samples. By better leveraging the support features for few-shot learning, the proposed method achieves the new state-of-the-art performance, with 15% less inference time, over existing few-shot 3D point cloud segmentation models on the S3DIS dataset and the ScanNet dataset. Our code is available at https://github.com/czzhang179/SCAT.

With the success of Vision Transformer (ViT) in image classification, its variants have yielded great success in many downstream vision tasks. Among those, the semantic segmentation task has also benefited greatly from the advance of ViT variants. However, most studies of the transformer for semantic segmentation only focus on designing efficient transformer encoders, rarely giving attention to designing the decoder. Several studies make attempts in using the transformer decoder as the segmentation decoder with class-wise learnable query. Instead, we aim to directly use the encoder features as the queries. This paper proposes the Feature Enhancing Decoder transFormer (FeedFormer) that enhances structural information using the transformer decoder. Our goal is to decode the high-level encoder features using the lowest-level encoder feature. We do this by formulating high-level features as queries, and the lowest-level feature as the key and value. This enhances the high-level features by collecting the structural information from the lowest-level feature. Additionally, we use a simple reformation trick of pushing the encoder blocks to take the place of the existing self-attention module of the decoder to improve efficiency. We show the superiority of our decoder with various light-weight transformer-based decoders on popular semantic segmentation datasets. Despite the minute computation, our model has achieved state-of-the-art performance in the performance computation trade-off. Our model FeedFormer-B0 surpasses SegFormer-B0 with 1.8% higher mIoU and 7.1% less computation on ADE20K, and 1.7% higher mIoU and 14.4% less computation on Cityscapes, respectively. Code will be released at: https://github.com/jhshim1995/FeedFormer.

Federated learning (FL) emerges as a popular distributed learning schema that learns a model from a set of participating users without sharing raw data. One major challenge of FL comes with heterogeneous users, who may have distributionally different (or non-iid) data and varying computation resources. As federated users would use the model for prediction, they often demand the trained model to be robust against malicious attackers at test time. Whereas adversarial training (AT) provides a sound solution for centralized learning, extending its usage for federated users has imposed significant challenges, as many users may have very limited training data and tight computational budgets, to afford the data-hungry and costly AT. In this paper, we study a novel FL strategy: propagating adversarial robustness from rich-resource users that can afford AT, to those with poor resources that cannot afford it, during federated learning. We show that existing FL techniques cannot be effectively integrated with the strategy to propagate robustness among non-iid users and propose an efficient propagation approach by the proper use of batch-normalization. We demonstrate the rationality and effectiveness of our method through extensive experiments. Especially, the proposed method is shown to grant federated models remarkable robustness even when only a small portion of users afford AT during learning. Source code can be accessed at https://github.com/illidanlab/FedRBN.

Graph neural networks (GNNs) have shown their superiority in modeling graph data. Owing to the advantages of federated learning, federated graph learning (FGL) enables clients to train strong GNN models in a distributed manner without sharing their private data. A core challenge in federated systems is the non-IID problem, which also widely exists in real-world graph data. For example, local data of clients may come from diverse datasets or even domains, e.g., social networks and molecules, increasing the difficulty for FGL methods to capture commonly shared knowledge and learn a generalized encoder. From real-world graph datasets, we observe that some structural properties are shared by various domains, presenting great potential for sharing structural knowledge in FGL. Inspired by this, we propose FedStar, an FGL framework that extracts and shares the common underlying structure information for inter-graph federated learning tasks. To explicitly extract the structure information rather than encoding them along with the node features, we define structure embeddings and encode them with an independent structure encoder. Then, the structure encoder is shared across clients while the feature-based knowledge is learned in a personalized way, making FedStar capable of capturing more structure-based domain-invariant information and avoiding feature misalignment issues. We perform extensive experiments over both cross-dataset and cross-domain non-IID FGL settings, demonstrating the superiority of FedStar.

The study of generative models is a promising branch of deep learning techniques, which has been successfully applied to different scenarios, such as Artificial Intelligence and the Internet of Things. While in most of the existing works, the generative models are realized as a centralized structure, raising the threats of security and privacy and the overburden of communication costs. Rare efforts have been committed to investigating distributed generative models, especially when the training data comes from multiple heterogeneous sources under realistic IoT settings. In this paper, to handle this challenging problem, we design a federated generative model framework that can learn a powerful generator for the hierarchical IoT systems. Particularly, our generative model framework can solve the problem of distributed data generation on multi-source heterogeneous data in two scenarios, i.e., feature related scenario and label related scenario. In addition, in our federated generative models, we develop a synchronous and an asynchronous updating methods to satisfy different application requirements. Extensive experiments on a simulated dataset and multiple real datasets are conducted to evaluate the data generation performance of our proposed generative models through comparison with the state-of-the-arts.

Traditional federated learning (FL) algorithms, such as FedAvg, fail to handle non-i.i.d data because they learn a global model by simply averaging biased local models that are trained on non-i.i.d local data, therefore failing to model the global data distribution. In this paper, we present a novel Bayesian FL algorithm that successfully handles such a non-i.i.d FL setting by enhancing the local training task with an auxiliary task that explicitly estimates the global data distribution. One key challenge in estimating the global data distribution is that the data are partitioned in FL, and therefore the ground-truth global data distribution is inaccessible. To address this challenge, we propose an expectation-propagation-inspired probabilistic neural network, dubbed federated neural propagation (FedNP), which efficiently estimates the global data distribution given non-i.i.d data partitions. Our algorithm is sampling-free and end-to-end differentiable, can be applied with any conventional FL frameworks and learns richer global data representation. Experiments on both image classification tasks with synthetic non-i.i.d image data partitions and real-world non-i.i.d speech recognition tasks demonstrate that our framework effectively alleviates the performance deterioration caused by non-i.i.d data.

Fairness has been considered as a critical problem in federated learning (FL). In this work, we analyze two direct causes of unfairness in FL - an unfair direction and an improper step size when updating the model. To solve these issues, we introduce an effective way to measure fairness of the model through the cosine similarity, and then propose a federated multiple gradient descent algorithm with fair guidance (FedMDFG) to drive the model fairer. We first convert FL into a multi-objective optimization problem (MOP) and design an advanced multiple gradient descent algorithm to calculate a fair descent direction by adding a fair-driven objective to MOP. A low-communication-cost line search strategy is then designed to find a better step size for the model update. We further show the theoretical analysis on how it can enhance fairness and guarantee the convergence. Finally, extensive experiments in several FL scenarios verify that FedMDFG is robust and outperforms the SOTA FL algorithms in convergence and fairness. The source code is available at https://github.com/zibinpan/FedMDFG.

While federated learning has shown strong results in opti- mizing a machine learning model without direct access to the original data, its performance may be hindered by in- termittent client availability which slows down the conver- gence and biases the final learned model. There are significant challenges to achieve both stable and bias-free training un- der arbitrary client availability. To address these challenges, we propose a framework named Federated Graph-based Sam- pling (FEDGS), to stabilize the global model update and mitigate the long-term bias given arbitrary client availabil- ity simultaneously. First, we model the data correlations of clients with a Data-Distribution-Dependency Graph (3DG) that helps keep the sampled clients data apart from each other, which is theoretically shown to improve the approximation to the optimal model update. Second, constrained by the far- distance in data distribution of the sampled clients, we fur- ther minimize the variance of the numbers of times that the clients are sampled, to mitigate long-term bias. To validate the effectiveness of FEDGS, we conduct experiments on three datasets under a comprehensive set of seven client availability modes. Our experimental results confirm FEDGS’s advantage in both enabling a fair client-sampling scheme and improving the model performance under arbitrary client availability. Our code is available at https://github.com/WwZzz/FedGS.

A key challenge in federated learning (FL) is the statistical heterogeneity that impairs the generalization of the global model on each client. To address this, we propose a method Federated learning with Adaptive Local Aggregation (FedALA) by capturing the desired information in the global model for client models in personalized FL. The key component of FedALA is an Adaptive Local Aggregation (ALA) module, which can adaptively aggregate the downloaded global model and local model towards the local objective on each client to initialize the local model before training in each iteration. To evaluate the effectiveness of FedALA, we conduct extensive experiments with five benchmark datasets in computer vision and natural language processing domains. FedALA outperforms eleven state-of-the-art baselines by up to 3.27% in test accuracy. Furthermore, we also apply ALA module to other federated learning methods and achieve up to 24.19% improvement in test accuracy. Code is available at https://github.com/TsingZ0/FedALA.

Federated learning aims to collaboratively train models without accessing their client's local private data. The data may be Non-IID for different clients and thus resulting in poor performance. Recently, personalized federated learning (PFL) has achieved great success in handling Non-IID data by enforcing regularization in local optimization or improving the model aggregation scheme on the server. However, most of the PFL approaches do not take into account the unfair competition issue caused by the imbalanced data distribution and lack of positive samples for some classes in each client. To address this issue, we propose a novel and generic PFL framework termed Federated Averaging via Binary Classification, dubbed FedABC. In particular, we adopt the ``one-vs-all'' training strategy in each client to alleviate the unfair competition between classes by constructing a personalized binary classification problem for each class. This may aggravate the class imbalance challenge and thus a novel personalized binary classification loss that incorporates both the under-sampling and hard sample mining strategies is designed. Extensive experiments are conducted on two popular datasets under different settings, and the results demonstrate that our FedABC can significantly outperform the existing counterparts.

We present a new algorithm to train a robust malware detector. Malware is a prolific problem and malware detectors are a front-line defense. Modern detectors rely on machine learning algorithms. Now, the adversarial objective is to devise alterations to the malware code to decrease the chance of being detected whilst preserving the functionality and realism of the malware. Adversarial learning is effective in improving robustness but generating functional and realistic adversarial malware samples is non-trivial. Because: i) in contrast to tasks capable of using gradient-based feedback, adversarial learning in a domain without a differentiable mapping function from the problem space (malware code inputs) to the feature space is hard; and ii) it is difficult to ensure the adversarial malware is realistic and functional. This presents a challenge for developing scalable adversarial machine learning algorithms for large datasets at a production or commercial scale to realize robust malware detectors. We propose an alternative; perform adversarial learning in the feature space in contrast to the problem space. We prove the projection of perturbed, yet valid malware, in the problem space into feature space will always be a subset of adversarials generated in the feature space. Hence, by generating a robust network against feature-space adversarial examples, we inherently achieve robustness against problem-space adversarial examples. We formulate a Bayesian adversarial learning objective that captures the distribution of models for improved robustness. To explain the robustness of the Bayesian adversarial learning algorithm, we prove that our learning method bounds the difference between the adversarial risk and empirical risk and improves robustness. We show that Bayesian neural networks (BNNs) achieve state-of-the-art results; especially in the False Positive Rate (FPR) regime. Adversarially trained BNNs achieve state-of-the-art robustness. Notably, adversarially trained BNNs are robust against stronger attacks with larger attack budgets by a margin of up to 15% on a recent production-scale malware dataset of more than 20 million samples. Importantly, our efforts create a benchmark for future defenses in the malware domain.

Natural language understanding (NLU) models often rely on dataset biases rather than intended task-relevant features to achieve high performance on specific datasets. As a result, these models perform poorly on datasets outside the training distribution. Some recent studies address this issue by reducing the weights of biased samples during the training process. However, these methods still encode biased latent features in representations and neglect the dynamic nature of bias, which hinders model prediction. We propose an NLU debiasing method, named debiasing contrastive learning (DCT), to simultaneously alleviate the above problems based on contrastive learning. We devise a debiasing, positive sampling strategy to mitigate biased latent features by selecting the least similar biased positive samples. We also propose a dynamic negative sampling strategy to capture the dynamic influence of biases by employing a bias-only model to dynamically select the most similar biased negative samples. We conduct experiments on three NLU benchmark datasets. Experimental results show that DCT outperforms state-of-the-art baselines on out-of-distribution datasets while maintaining in-distribution performance. We also verify that DCT can reduce biased latent features from the model's representation.

To train a model in a traditional supervised learning classification system for natural language processing (NLP) tasks, it is essential to have labeled data, which is not present in large amounts for many tasks. Prompt-based learning methods attempt to combat the supervised learning need for labeled data by directly adapting pre-trained language models and modeling the probability of text itself. In this paper, we propose a novel data-agnostic strategy for prompt-based fine-tuning that leverages feature moments (a.k.a., mean and standard deviation) as a data augmentation technique and employs training dynamics (i.e., confidence and variability) to allow more informative samples to be concatenated for generating demonstrations as input context. Our approach is a strong method for few-shot learning that forces the language model to pay special attention to the feature moments and allows more informative samples to be concatenated for generating demonstrations as input context by selecting high confidence and low variance samples. To demonstrate its effectiveness given limited training data, we conduct extensive experiments in different few-shot settings on three empathy and emotion classification datasets (from various domains). We further evaluate our method's robustness by introducing noise to our few-shot input data and labels and show that exchanging moments between samples and incorporating cartography-based demonstrations are beneficial when the available data is limited and noisy.

Few-shot learning has received increasing attention and witnessed significant advances in recent years. However, most of the few-shot learning methods focus on the optimization of training process, and the learning of metric and sample generating networks. They ignore the importance of learning the ground-truth feature distributions of few-shot classes. This paper proposes a direction-driven weighting method to make the feature distributions of few-shot classes precisely fit the ground-truth distributions. The learned feature distributions can generate an unlimited number of training samples for the few-shot classes to avoid overfitting. Specifically, the proposed method consists of two optimization strategies. The direction-driven strategy is for capturing more complete direction information that can describe the feature distributions. The similarity-weighting strategy is proposed to estimate the impact of different classes in the fitting procedure and assign corresponding weights. Our method outperforms the current state-of-the-art performance by an average of 3% for 1-shot on standard few-shot learning benchmarks like miniImageNet, CIFAR-FS, and CUB. The excellent performance and compelling visualization show that our method can more accurately estimate the ground-truth distributions.

We propose a novel multi-task learning method termed Feature Decomposition Network (FDN). The key idea of the proposed FDN is to reduce the phenomenon of feature redundancy by explicitly decomposing features into task-specific features and task-shared features with carefully designed constraints. Experimental results show that our proposed FDN can outperform the state-of-the-art (SOTA) methods by a noticeable margin on Ali-CCP.