Transformers have made much progress in dealing with visual tasks. However, existing vision transformers still do not possess an ability that is important to visual input: building the attention among features of different scales. The reasons for thi
s problem are two-fold: (1) Input embeddings of each layer are equal-scale without cross-scale features; (2) Some vision transformers sacrifice the small-scale features of embeddings to lower the cost of the self-attention module. To make up this defect, we propose Cross-scale Embedding Layer (CEL) and Long Short Distance Attention (LSDA). In particular, CEL blends each embedding with multiple patches of different scales, providing the model with cross-scale embeddings. LSDA splits the self-attention module into a short-distance and long-distance one, also lowering the cost but keeping both small-scale and large-scale features in embeddings. Through these two designs, we achieve cross-scale attention. Besides, we propose dynamic position bias for vision transformers to make the popular relative position bias apply to variable-sized images. Based on these proposed modules, we construct our vision architecture called CrossFormer. Experiments show that CrossFormer outperforms other transformers on several representative visual tasks, especially object detection and segmentation. The code has been released: https://github.com/cheerss/CrossFormer.
In self-supervised representation learning, a common idea behind most of the state-of-the-art approaches is to enforce the robustness of the representations to predefined augmentations. A potential issue of this idea is the existence of completely co
llapsed solutions (i.e., constant features), which are typically avoided implicitly by carefully chosen implementation details. In this work, we study a relatively concise framework containing the most common components from recent approaches. We verify the existence of complete collapse and discover another reachable collapse pattern that is usually overlooked, namely dimensional collapse. We connect dimensional collapse with strong correlations between axes and consider such connection as a strong motivation for feature decorrelation (i.e., standardizing the covariance matrix). The gains from feature decorrelation are verified empirically to highlight the importance and the potential of this insight.
Deep learning techniques have achieved remarkable performance in wide-ranging tasks. However, when trained on privacy-sensitive datasets, the model parameters may expose private information in training data. Prior attempts for differentially private
training, although offering rigorous privacy guarantees, lead to much lower model performance than the non-private ones. Besides, different runs of the same training algorithm produce models with large performance variance. To address these issues, we propose DPlis--Differentially Private Learning wIth Smoothing. The core idea of DPlis is to construct a smooth loss function that favors noise-resilient models lying in large flat regions of the loss landscape. We provide theoretical justification for the utility improvements of DPlis. Extensive experiments also demonstrate that DPlis can effectively boost model quality and training stability under a given privacy budget.
The width of a neural network matters since increasing the width will necessarily increase the model capacity. However, the performance of a network does not improve linearly with the width and soon gets saturated. In this case, we argue that increas
ing the number of networks (ensemble) can achieve better accuracy-efficiency trade-offs than purely increasing the width. To prove it, one large network is divided into several small ones regarding its parameters and regularization components. Each of these small networks has a fraction of the original ones parameters. We then train these small networks together and make them see various views of the same data to increase their diversity. During this co-training process, networks can also learn from each other. As a result, small networks can achieve better ensemble performance than the large one with few or no extra parameters or FLOPs. Small networks can also achieve faster inference speed than the large one by concurrent running on different devices. We validate our argument with 8 different neural architectures on common benchmarks through extensive experiments. The code is available at url{https://github.com/mzhaoshuai/Divide-and-Co-training}.
Most neural network pruning methods, such as filter-level and layer-level prunings, prune the network model along one dimension (depth, width, or resolution) solely to meet a computational budget. However, such a pruning policy often leads to excessi
ve reduction of that dimension, thus inducing a huge accuracy loss. To alleviate this issue, we argue that pruning should be conducted along three dimensions comprehensively. For this purpose, our pruning framework formulates pruning as an optimization problem. Specifically, it first casts the relationships between a certain models accuracy and depth/width/resolution into a polynomial regression and then maximizes the polynomial to acquire the optimal values for the three dimensions. Finally, the model is pruned along the three optimal dimensions accordingly. In this framework, since collecting too much data for training the regression is very time-costly, we propose two approaches to lower the cost: 1) specializing the polynomial to ensure an accurate regression even with less training data; 2) employing iterative pruning and fine-tuning to collect the data faster. Extensive experiments show that our proposed algorithm surpasses state-of-the-art pruning algorithms and even neural architecture search-based algorithms.
3D object detection based on point clouds has become more and more popular. Some methods propose localizing 3D objects directly from raw point clouds to avoid information loss. However, these methods come with complex structures and significant compu
tational overhead, limiting its broader application in real-time scenarios. Some methods choose to transform the point cloud data into compact tensors first and leverage off-the-shelf 2D detectors to propose 3D objects, which is much faster and achieves state-of-the-art results. However, because of the inconsistency between 2D and 3D data, we argue that the performance of compact tensor-based 3D detectors is restricted if we use 2D detectors without corresponding modification. Specifically, the distribution of point clouds is uneven, with most points gather on the boundary of objects, while detectors for 2D data always extract features evenly. Motivated by this observation, we propose DENse Feature Indicator (DENFI), a universal module that helps 3D detectors focus on the densest region of the point clouds in a boundary-aware manner. Moreover, DENFI is lightweight and guarantees real-time speed when applied to 3D object detectors. Experiments on KITTI dataset show that DENFI improves the performance of the baseline single-stage detector remarkably, which achieves new state-of-the-art performance among previous 3D detectors, including both two-stage and multi-sensor fusion methods, in terms of mAP with a 34FPS detection speed.
Training deep neural networks from scratch could be computationally expensive and requires a lot of training data. Recent work has explored different watermarking techniques to protect the pre-trained deep neural networks from potential copyright inf
ringements. However, these techniques could be vulnerable to watermark removal attacks. In this work, we propose REFIT, a unified watermark removal framework based on fine-tuning, which does not rely on the knowledge of the watermarks, and is effective against a wide range of watermarking schemes. In particular, we conduct a comprehensive study of a realistic attack scenario where the adversary has limited training data, which has not been emphasized in prior work on attacks against watermarking schemes. To effectively remove the watermarks without compromising the model functionality under this weak threat model, we propose two techniques that are incorporated into our fine-tuning framework: (1) an adaption of the elastic weight consolidation (EWC) algorithm, which is originally proposed for mitigating the catastrophic forgetting phenomenon; and (2) unlabeled data augmentation (AU), where we leverage auxiliary unlabeled data from other sources. Our extensive evaluation shows the effectiveness of REFIT against diverse watermark embedding schemes. In particular, both EWC and AU significantly decrease the amount of labeled training data needed for effective watermark removal, and the unlabeled data samples used for AU do not necessarily need to be drawn from the same distribution as the benign data for model evaluation. The experimental results demonstrate that our fine-tuning based watermark removal attacks could pose real threats to the copyright of pre-trained models, and thus highlight the importance of further investigating the watermarking problem and proposing more robust watermark embedding schemes against the attacks.
