Neural networks for visual content understanding have recently evolved from convolutional ones (CNNs) to transformers. The prior (CNN) relies on small-windowed kernels to capture the regional clues, demonstrating solid local expressiveness. On the contrary, the latter (transformer) establishes long-range global connections between localities for holistic learning. Inspired by this complementary nature, there is a growing interest in designing hybrid models to best utilize each technique. Current hybrids merely replace convolutions as simple approximations of linear projection or juxtapose a convolution branch with attention, without concerning the importance of local/global modeling. To tackle this, we propose a new hybrid named Adaptive Split-Fusion Transformer (ASF-former) to treat convolutional and attention branches differently with adaptive weights. Specifically, an ASF-former encoder equally splits feature channels into half to fit dual-path inputs. Then, the outputs of dual-path are fused with weighting scalars calculated from visual cues. We also design the convolutional path compactly for efficiency concerns. Extensive experiments on standard benchmarks, such as ImageNet-1K, CIFAR-10, and CIFAR-100, show that our ASF-former outperforms its CNN, transformer counterparts, and hybrid pilots in terms of accuracy (83.9% on ImageNet-1K), under similar conditions (12.9G MACs/56.7M Params, without large-scale pre-training). The code is available at: https://github.com/szx503045266/ASF-former.
We present DINO (\textbf{D}ETR with \textbf{I}mproved de\textbf{N}oising anch\textbf{O}r boxes), a state-of-the-art end-to-end object detector. % in this paper. DINO improves over previous DETR-like models in performance and efficiency by using a contrastive way for denoising training, a mixed query selection method for anchor initialization, and a look forward twice scheme for box prediction. DINO achieves $48.3$AP in $12$ epochs and $51.0$AP in $36$ epochs on COCO with a ResNet-50 backbone and multi-scale features, yielding a significant improvement of $\textbf{+4.9}$\textbf{AP} and $\textbf{+2.4}$\textbf{AP}, respectively, compared to DN-DETR, the previous best DETR-like model. DINO scales well in both model size and data size. Without bells and whistles, after pre-training on the Objects365 dataset with a SwinL backbone, DINO obtains the best results on both COCO \texttt{val2017} ($\textbf{63.2}$\textbf{AP}) and \texttt{test-dev} (\textbf{$\textbf{63.3}$AP}). Compared to other models on the leaderboard, DINO significantly reduces its model size and pre-training data size while achieving better results. Our code will be available at \url{https://github.com/IDEACVR/DINO}.
As an emerging technology, federated learning (FL) involves training machine learning models over distributed edge devices, which attracts sustained attention and has been extensively studied. However, the heterogeneity of client data severely degrades the performance of FL compared with that in centralized training. It causes the locally trained models of clients to move in different directions. On the one hand, it slows down or even stalls the global updates, leading to inefficient communication. On the other hand, it enlarges the distances between local models, resulting in an aggregated global model with poor performance. Fortunately, these shortcomings can be mitigated by reducing the angle between the directions that local models move in. Based on this fact, we propose FedCos, which reduces the directional inconsistency of local models by introducing a cosine-similarity penalty. It promotes the local model iterations towards an auxiliary global direction. Moreover, our approach is auto-adapt to various non-IID settings without an elaborate selection of hyperparameters. The experimental results show that FedCos outperforms the well-known baselines and can enhance them under a variety of FL scenes, including varying degrees of data heterogeneity, different number of participants, and cross-silo and cross-device settings. Besides, FedCos improves communication efficiency by 2 to 5 times. With the help of FedCos, multiple FL methods require significantly fewer communication rounds than before to obtain a model with comparable performance.
Inspired by the complexity and diversity of biological neurons, a quadratic neuron is proposed to replace the inner product in the current neuron with a simplified quadratic function. Employing such a novel type of neurons offers a new perspective on developing deep learning. When analyzing quadratic neurons, we find that there exists a function such that a heterogeneous network can approximate it well with a polynomial number of neurons but a purely conventional or quadratic network needs an exponential number of neurons to achieve the same level of error. Encouraged by this inspiring theoretical result on heterogeneous networks, we directly integrate conventional and quadratic neurons in an autoencoder to make a new type of heterogeneous autoencoders. Anomaly detection experiments confirm that heterogeneous autoencoders perform competitively compared to other state-of-the-art models.
We present in this paper a novel denoising training method to speedup DETR (DEtection TRansformer) training and offer a deepened understanding of the slow convergence issue of DETR-like methods. We show that the slow convergence results from the instability of bipartite graph matching which causes inconsistent optimization goals in early training stages. To address this issue, except for the Hungarian loss, our method additionally feeds ground-truth bounding boxes with noises into Transformer decoder and trains the model to reconstruct the original boxes, which effectively reduces the bipartite graph matching difficulty and leads to a faster convergence. Our method is universal and can be easily plugged into any DETR-like methods by adding dozens of lines of code to achieve a remarkable improvement. As a result, our DN-DETR results in a remarkable improvement ($+1.9$AP) under the same setting and achieves the best result (AP $43.4$ and $48.6$ with $12$ and $50$ epochs of training respectively) among DETR-like methods with ResNet-$50$ backbone. Compared with the baseline under the same setting, DN-DETR achieves comparable performance with $50\%$ training epochs. Code is available at \url{https://github.com/FengLi-ust/DN-DETR}.
By driving optimizers to converge to flat minima, sharpness-aware minimization (SAM) has shown the power to improve the model generalization. However, SAM requires to perform two forward-backward propagations for one parameter update, which largely burdens the practical computation. In this paper, we propose a novel and efficient training scheme, called Stochastic Scheduled SAM (SS-SAM). Specifically, in SS-SAM, the optimizer is arranged by a predefined scheduling function to perform a random trial at each update step, which would randomly select to perform the SGD optimization or the SAM optimization. In this way, the overall count of propagation pair could be largely reduced. Then, we empirically investigate four typical types of scheduling functions, and demonstrates the computational efficiency and their impact on model performance respectively. We show that with proper scheduling functions, models could be trained to achieve comparable or even better performance with much lower computation cost compared to models trained with only SAM training scheme.
Learning discriminative representation from the complex spatio-temporal dynamic space is essential for video recognition. On top of those stylized spatio-temporal computational units, further refining the learnt feature with axial contexts is demonstrated to be promising in achieving this goal. However, previous works generally focus on utilizing a single kind of contexts to calibrate entire feature channels and could hardly apply to deal with diverse video activities. The problem can be tackled by using pair-wise spatio-temporal attentions to recompute feature response with cross-axis contexts at the expense of heavy computations. In this paper, we propose an efficient feature refinement method that decomposes the feature channels into several groups and separately refines them with different axial contexts in parallel. We refer this lightweight feature calibration as group contextualization (GC). Specifically, we design a family of efficient element-wise calibrators, i.e., ECal-G/S/T/L, where their axial contexts are information dynamics aggregated from other axes either globally or locally, to contextualize feature channel groups. The GC module can be densely plugged into each residual layer of the off-the-shelf video networks. With little computational overhead, consistent improvement is observed when plugging in GC on different networks. By utilizing calibrators to embed feature with four different kinds of contexts in parallel, the learnt representation is expected to be more resilient to diverse types of activities. On videos with rich temporal variations, empirically GC can boost the performance of 2D-CNN (e.g., TSN and TSM) to a level comparable to the state-of-the-art video networks. Code is available at https://github.com/haoyanbin918/Group-Contextualization.
We present a hierarchical neural network model called SemText to detect HTML boilerplate based on a novel semantic representation of HTML tags, class names, and text blocks. We train SemText on three published datasets of news webpages and fine-tune it using a small number of development data in CleanEval and GoogleTrends-2017. We show that SemText achieves the state-of-the-art accuracy on these datasets. We then demonstrate the robustness of SemText by showing that it also detects boilerplate effectively on out-of-domain community-based question-answer webpages.
We construct a contextual network to represent a document with syntactic and semantic relations between word-sentence pairs, based on which we devise an unsupervised algorithm called CNATAR (Contextual Network And Text Analysis Rank) to score sentences, and rank them through a bi-objective 0-1 knapsack maximization problem over topic analysis and sentence scores. We show that CNATAR outperforms the combined ranking of the three human judges provided on the SummBank dataset under both ROUGE and BLEU metrics, which in term significantly outperforms each individual judge's ranking. Moreover, CNATAR produces so far the highest ROUGE scores over DUC-02, and outperforms previous supervised algorithms on the CNN/DailyMail and NYT datasets. We also compare the performance of CNATAR and the latest supervised neural-network summarization models and compute oracle results.