The "Roaring 20s" of visual recognition began with the introduction of Vision Transformers (ViTs), which quickly superseded ConvNets as the state-of-the-art image classification model. A vanilla ViT, on the other hand, faces difficulties when applied to general computer vision tasks such as object detection and semantic segmentation. It is the hierarchical Transformers (e.g., Swin Transformers) that reintroduced several ConvNet priors, making Transformers practically viable as a generic vision backbone and demonstrating remarkable performance on a wide variety of vision tasks. However, the effectiveness of such hybrid approaches is still largely credited to the intrinsic superiority of Transformers, rather than the inherent inductive biases of convolutions. In this work, we reexamine the design spaces and test the limits of what a pure ConvNet can achieve. We gradually "modernize" a standard ResNet toward the design of a vision Transformer, and discover several key components that contribute to the performance difference along the way. The outcome of this exploration is a family of pure ConvNet models dubbed ConvNeXt. Constructed entirely from standard ConvNet modules, ConvNeXts compete favorably with Transformers in terms of accuracy and scalability, achieving 87.8% ImageNet top-1 accuracy and outperforming Swin Transformers on COCO detection and ADE20K segmentation, while maintaining the simplicity and efficiency of standard ConvNets.
This paper explores the feasibility of finding an optimal sub-model from a vision transformer and introduces a pure vision transformer slimming (ViT-Slim) framework that can search such a sub-structure from the original model end-to-end across multiple dimensions, including the input tokens, MHSA and MLP modules with state-of-the-art performance. Our method is based on a learnable and unified l1 sparsity constraint with pre-defined factors to reflect the global importance in the continuous searching space of different dimensions. The searching process is highly efficient through a single-shot training scheme. For instance, on DeiT-S, ViT-Slim only takes ~43 GPU hours for searching process, and the searched structure is flexible with diverse dimensionalities in different modules. Then, a budget threshold is employed according to the requirements of accuracy-FLOPs trade-off on running devices, and a re-training process is performed to obtain the final models. The extensive experiments show that our ViT-Slim can compress up to 40% of parameters and 40% FLOPs on various vision transformers while increasing the accuracy by ~0.6% on ImageNet. We also demonstrate the advantage of our searched models on several downstream datasets. Our source code will be publicly available.
We present MSeg, a composite dataset that unifies semantic segmentation datasets from different domains. A naive merge of the constituent datasets yields poor performance due to inconsistent taxonomies and annotation practices. We reconcile the taxonomies and bring the pixel-level annotations into alignment by relabeling more than 220,000 object masks in more than 80,000 images, requiring more than 1.34 years of collective annotator effort. The resulting composite dataset enables training a single semantic segmentation model that functions effectively across domains and generalizes to datasets that were not seen during training. We adopt zero-shot cross-dataset transfer as a benchmark to systematically evaluate a model's robustness and show that MSeg training yields substantially more robust models in comparison to training on individual datasets or naive mixing of datasets without the presented contributions. A model trained on MSeg ranks first on the WildDash-v1 leaderboard for robust semantic segmentation, with no exposure to WildDash data during training. We evaluate our models in the 2020 Robust Vision Challenge (RVC) as an extreme generalization experiment. MSeg training sets include only three of the seven datasets in the RVC; more importantly, the evaluation taxonomy of RVC is different and more detailed. Surprisingly, our model shows competitive performance and ranks second. To evaluate how close we are to the grand aim of robust, efficient, and complete scene understanding, we go beyond semantic segmentation by training instance segmentation and panoptic segmentation models using our dataset. Moreover, we also evaluate various engineering design decisions and metrics, including resolution and computational efficiency. Although our models are far from this grand aim, our comprehensive evaluation is crucial for progress. We share all the models and code with the community.
Anytime inference requires a model to make a progression of predictions which might be halted at any time. Prior research on anytime visual recognition has mostly focused on image classification. We propose the first unified and end-to-end model approach for anytime pixel-level recognition. A cascade of "exits" is attached to the model to make multiple predictions and direct further computation. We redesign the exits to account for the depth and spatial resolution of the features for each exit. To reduce total computation, and make full use of prior predictions, we develop a novel spatially adaptive approach to avoid further computation on regions where early predictions are already sufficiently confident. Our full model with redesigned exit architecture and spatial adaptivity enables anytime inference, achieves the same level of final accuracy, and even significantly reduces total computation. We evaluate our approach on semantic segmentation and human pose estimation. On Cityscapes semantic segmentation and MPII human pose estimation, our approach enables anytime inference while also reducing the total FLOPs of its base models by 44.4% and 59.1% without sacrificing accuracy. As a new anytime baseline, we measure the anytime capability of deep equilibrium networks, a recent class of model that is intrinsically iterative, and we show that the accuracy-computation curve of our architecture strictly dominates it.
Recommender systems, which analyze users' preference patterns to suggest potential targets, are indispensable in today's society. Collaborative Filtering (CF) is the most popular recommendation model. Specifically, Graph Neural Network (GNN) has become a new state-of-the-art for CF. In the GNN-based recommender system, message dropout is usually used to alleviate the selection bias in the user-item bipartite graph. However, message dropout might deteriorate the recommender system's performance due to the randomness of dropping out the outgoing messages based on the user-item bipartite graph. To solve this problem, we propose a graph contrastive learning module for a general recommender system that learns the embeddings in a self-supervised manner and reduces the randomness of message dropout. Besides, many recommender systems optimize models with pairwise ranking objectives, such as the Bayesian Pairwise Ranking (BPR) based on a negative sampling strategy. However, BPR has the following problems: suboptimal sampling and sample bias. We introduce a new debiased contrastive loss to solve these problems, which provides sufficient negative samples and applies a bias correction probability to alleviate the sample bias. We integrate the proposed framework, including graph contrastive module and debiased contrastive module with several Matrix Factorization(MF) and GNN-based recommendation models. Experimental results on three public benchmarks demonstrate the effectiveness of our framework.
Meta-learning has become a popular framework for few-shot learning in recent years, with the goal of learning a model from collections of few-shot classification tasks. While more and more novel meta-learning models are being proposed, our research has uncovered simple baselines that have been overlooked. We present a Meta-Baseline method, by pre-training a classifier on all base classes and meta-learning on a nearest-centroid based few-shot classification algorithm, it outperforms recent state-of-the-art methods by a large margin. Why does this simple method work so well? In the meta-learning stage, we observe that a model generalizing better on unseen tasks from base classes can have a decreasing performance on tasks from novel classes, indicating a potential objective discrepancy. We find both pre-training and inheriting a good few-shot classification metric from the pre-trained classifier are important for Meta-Baseline, which potentially helps the model better utilize the pre-trained representations with stronger transferability. Furthermore, we investigate when we need meta-learning in this Meta-Baseline. Our work sets up a new solid benchmark for this field and sheds light on further understanding the phenomenons in the meta-learning framework for few-shot learning.
In supervised learning, smoothing label/prediction distribution in neural network training has been proven useful in preventing the model from being over-confident, and is crucial for learning more robust visual representations. This observation motivates us to explore the way to make predictions flattened in unsupervised learning. Considering that human annotated labels are not adopted in unsupervised learning, we introduce a straightforward approach to perturb input image space in order to soften the output prediction space indirectly. Despite its conceptual simplicity, we show empirically that with the simple solution -- image mixture, we can learn more robust visual representations from the transformed input, and the benefits of representations learned from this space can be inherited by the linear classification and downstream tasks.
Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion.Whereas traditional convolutional networks with L layers have L connections - one between each layer and its subsequent layer - our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, encourage feature reuse and substantially improve parameter efficiency. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less parameters and computation to achieve high performance.
Automatically extracting Protein-Protein Interactions (PPI) from biomedical literature provides additional support for precision medicine efforts. This paper proposes a novel memory network-based model (MNM) for PPI extraction, which leverages prior knowledge about protein-protein pairs with memory networks. The proposed MNM captures important context clues related to knowledge representations learned from knowledge bases. Both entity embeddings and relation embeddings of prior knowledge are effective in improving the PPI extraction model, leading to a new state-of-the-art performance on the BioCreative VI PPI dataset. The paper also shows that multiple computational layers over an external memory are superior to long short-term memory networks with the local memories.