Non-exemplar class-incremental learning is to recognize both the old and new classes when old class samples cannot be saved. It is a challenging task since representation optimization and feature retention can only be achieved under supervision from new classes. To address this problem, we propose a novel self-sustaining representation expansion scheme. Our scheme consists of a structure reorganization strategy that fuses main-branch expansion and side-branch updating to maintain the old features, and a main-branch distillation scheme to transfer the invariant knowledge. Furthermore, a prototype selection mechanism is proposed to enhance the discrimination between the old and new classes by selectively incorporating new samples into the distillation process. Extensive experiments on three benchmarks demonstrate significant incremental performance, outperforming the state-of-the-art methods by a margin of 3%, 3% and 6%, respectively.
Graph Convolutional Network (GCN) plays pivotal roles in many real-world applications. Despite the successes of GCN deployment, GCN often exhibits performance disparity with respect to node degrees, resulting in worse predictive accuracy for low-degree nodes. We formulate the problem of mitigating the degree-related performance disparity in GCN from the perspective of the Rawlsian difference principle, which is originated from the theory of distributive justice. Mathematically, we aim to balance the utility between low-degree nodes and high-degree nodes while minimizing the task-specific loss. Specifically, we reveal the root cause of this degree-related unfairness by analyzing the gradients of weight matrices in GCN. Guided by the gradients of weight matrices, we further propose a pre-processing method RawlsGCN-Graph and an in-processing method RawlsGCN-Grad that achieves fair predictive accuracy in low-degree nodes without modification on the GCN architecture or introduction of additional parameters. Extensive experiments on real-world graphs demonstrate the effectiveness of our proposed RawlsGCN methods in significantly reducing degree-related bias while retaining comparable overall performance.
To achieve point cloud denoising, traditional methods heavily rely on geometric priors, and most learning-based approaches suffer from outliers and loss of details. Recently, the gradient-based method was proposed to estimate the gradient fields from the noisy point clouds using neural networks, and refine the position of each point according to the estimated gradient. However, the predicted gradient could fluctuate, leading to perturbed and unstable solutions, as well as a large inference time. To address these issues, we develop the momentum gradient ascent method that leverages the information of previous iterations in determining the trajectories of the points, thus improving the stability of the solution and reducing the inference time. Experiments demonstrate that the proposed method outperforms state-of-the-art methods with a variety of point clouds and noise levels.
In video denoising, the adjacent frames often provide very useful information, but accurate alignment is needed before such information can be harnassed. In this work, we present a multi-alignment network, which generates multiple flow proposals followed by attention-based averaging. It serves to mimics the non-local mechanism, suppressing noise by averaging multiple observations. Our approach can be applied to various state-of-the-art models that are based on flow estimation. Experiments on a large-scale video dataset demonstrate that our method improves the denoising baseline model by 0.2dB, and further reduces the parameters by 47% with model distillation.
In this study, we aim to predict the plausible future action steps given an observation of the past and study the task of instructional activity anticipation. Unlike previous anticipation tasks that aim at action label prediction, our work targets at generating natural language outputs that provide interpretable and accurate descriptions of future action steps. It is a challenging task due to the lack of semantic information extracted from the instructional videos. To overcome this challenge, we propose a novel knowledge distillation framework to exploit the related external textual knowledge to assist the visual anticipation task. However, previous knowledge distillation techniques generally transfer information within the same modality. To bridge the gap between the visual and text modalities during the distillation process, we devise a novel cross-modal contrastive distillation (CCD) scheme, which facilitates knowledge distillation between teacher and student in heterogeneous modalities with the proposed cross-modal distillation loss. We evaluate our method on the Tasty Videos dataset. CCD improves the anticipation performance of the visual-alone student model by a large margin of 40.2% relatively in BLEU4. Our approach also outperforms the state-of-the-art approaches by a large margin.
The Jaccard index, also known as Intersection-over-Union (IoU), is one of the most critical evaluation metrics in image semantic segmentation. However, direct optimization of IoU score is very difficult because the learning objective is neither differentiable nor decomposable. Although some algorithms have been proposed to optimize its surrogates, there is no guarantee provided for the generalization ability. In this paper, we propose a margin calibration method, which can be directly used as a learning objective, for an improved generalization of IoU over the data-distribution, underpinned by a rigid lower bound. This scheme theoretically ensures a better segmentation performance in terms of IoU score. We evaluated the effectiveness of the proposed margin calibration method on seven image datasets, showing substantial improvements in IoU score over other learning objectives using deep segmentation models.
Recently, large pretrained models (e.g., BERT, StyleGAN, CLIP) have shown great knowledge transfer and generalization capability on various downstream tasks within their domains. Inspired by these efforts, in this paper we propose a unified model for open-domain image editing focusing on color and tone adjustment of open-domain images while keeping their original content and structure. Our model learns a unified editing space that is more semantic, intuitive, and easy to manipulate than the operation space (e.g., contrast, brightness, color curve) used in many existing photo editing softwares. Our model belongs to the image-to-image translation framework which consists of an image encoder and decoder, and is trained on pairs of before- and after-images to produce multimodal outputs. We show that by inverting image pairs into latent codes of the learned editing space, our model can be leveraged for various downstream editing tasks such as language-guided image editing, personalized editing, editing-style clustering, retrieval, etc. We extensively study the unique properties of the editing space in experiments and demonstrate superior performance on the aforementioned tasks.
Music sentiment transfer is a completely novel task. Sentiment transfer is a natural evolution of the heavily-studied style transfer task, as sentiment transfer is rooted in applying the sentiment of a source to be the new sentiment for a target piece of media; yet compared to style transfer, sentiment transfer has been only scantily studied on images. Music sentiment transfer attempts to apply the high level objective of sentiment transfer to the domain of music. We propose CycleGAN to bridge disparate domains. In order to use the network, we choose to use symbolic, MIDI, data as the music format. Through the use of a cycle consistency loss, we are able to create one-to-one mappings that preserve the content and realism of the source data. Results and literature suggest that the task of music sentiment transfer is more difficult than image sentiment transfer because of the temporal characteristics of music and lack of existing datasets.
Learning new skills by observing humans' behaviors is an essential capability of AI. In this work, we leverage instructional videos to study humans' decision-making processes, focusing on learning a model to plan goal-directed actions in real-life videos. In contrast to conventional action recognition, goal-directed actions are based on expectations of their outcomes requiring causal knowledge of potential consequences of actions. Thus, integrating the environment structure with goals is critical for solving this task. Previous works learn a single world model will fail to distinguish various tasks, resulting in an ambiguous latent space; planning through it will gradually neglect the desired outcomes since the global information of the future goal degrades quickly as the procedure evolves. We address these limitations with a new formulation of procedure planning and propose novel algorithms to model human behaviors through Bayesian Inference and model-based Imitation Learning. Experiments conducted on real-world instructional videos show that our method can achieve state-of-the-art performance in reaching the indicated goals. Furthermore, the learned contextual information presents interesting features for planning in a latent space.
Some cognitive research has discovered that humans accomplish event segmentation as a side effect of event anticipation. Inspired by this discovery, we propose a simple yet effective end-to-end self-supervised learning framework for event segmentation/boundary detection. Unlike the mainstream clustering-based methods, our framework exploits a transformer-based feature reconstruction scheme to detect event boundary by reconstruction errors. This is consistent with the fact that humans spot new events by leveraging the deviation between their prediction and what is actually perceived. Thanks to their heterogeneity in semantics, the frames at boundaries are difficult to be reconstructed (generally with large reconstruction errors), which is favorable for event boundary detection. Additionally, since the reconstruction occurs on the semantic feature level instead of pixel level, we develop a temporal contrastive feature embedding module to learn the semantic visual representation for frame feature reconstruction. This procedure is like humans building up experiences with "long-term memory". The goal of our work is to segment generic events rather than localize some specific ones. We focus on achieving accurate event boundaries. As a result, we adopt F1 score (Precision/Recall) as our primary evaluation metric for a fair comparison with previous approaches. Meanwhile, we also calculate the conventional frame-based MoF and IoU metric. We thoroughly benchmark our work on four publicly available datasets and demonstrate much better results.