Vision Transformers (ViTs) have shown impressive per-formance but still require a high computation cost as compared to convolutional neural networks (CNNs), one rea-son is that ViTs' attention measures global similarities and thus has a quadratic complexity with the number of in-put tokens. Existing efficient ViTs adopt local attention or linear attention, which sacrifice ViTs' capabilities of capturing either global or local context. In this work, we ask an important research question: Can ViTs learn both global and local context while being more efficient during inference? To this end, we propose a framework called Castling- ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attention during inference. Our Castling- ViT leverages angular ker-nels to measure the similarities between queries and keys via spectral angles. And we further simplify it with two techniques: (1) a novel linear-angular attention mechanism: we decompose the angular kernels into linear terms and high-order residuals, and only keep the linear terms; and (2) we adopt two parameterized modules to approximate high-order residuals: a depthwise convolution and an aux-iliary masked softmax attention to help learn global and lo-cal information, where the masks for softmax attention are regularized to gradually become zeros and thus incur no overhead during inference. Extensive experiments validate the effectiveness of our Castling- ViT, e.g., achieving up to a 1.8% higher accuracy or 40% MACs reduction on classification and 1.2 higher mAP on detection under comparable FLOPs, as compared to ViTs with vanilla softmax-based at-tentions. Project page is available at here.
more »
« less
This content will become publicly available on June 1, 2024
PaCa-ViT: Learning Patch-to-Cluster Attention in Vision Transformers
Vision Transformers (ViTs) are built on the assumption of treating image patches as “visual tokens” and learn patch-to-patch attention. The patch embedding based tokenizer has a semantic gap with respect to its counterpart, the textual tokenizer. The patch-to-patch attention suffers from the quadratic complexity issue, and also makes it non-trivial to explain learned ViTs. To address these issues in ViT, this paper proposes to learn Patch-to-Cluster attention (PaCa) in ViT. Queries in our PaCa-ViT starts with patches, while keys and values are directly based on clustering (with a predefined small number of clusters). The clusters are learned end-to-end, leading to better tokenizers and inducing joint clustering-for-attention and attention-for-clustering for better and interpretable models. The quadratic complexity is relaxed to linear complexity. The proposed PaCa module is used in designing efficient and interpretable ViT backbones and semantic segmentation head networks. In experiments, the proposed methods are tested on ImageNet-1k image classification, MS-COCO object detection and instance segmentation and MIT-ADE20k semantic segmentation. Compared with the prior art, it obtains better performance in all the three benchmarks than the SWin [32] and the PVTs [47], [48] by significant margins in ImageNet-1k and MIT-ADE20k. It is also significantly more efficient than PVT models in MS-COCO and MIT-ADE20k due to the linear complexity. The learned clusters are semantically meaningful. Code and model checkpoints are available at https:/github.com/iVMCL/PaCaViT.
more »
« less
- NSF-PAR ID:
- 10468302
- Publisher / Repository:
- IEEE
- Date Published:
- Page Range / eLocation ID:
- 18568 to 18578
- Format(s):
- Medium: X
- Location:
- Vancouver, BC, Canada
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Vedaldi, A. (Ed.)In state-of-the-art deep neural networks, both feature normalization and feature attention have become ubiquitous. They are usually studied as separate modules, however. In this paper, we propose a light-weight integration between the two schema and present Attentive Normalization (AN). Instead of learning a single affine transformation, AN learns a mixture of affine transformations and utilizes their weighted-sum as the final affine transformation applied to re-calibrate features in an instance-specific way. The weights are learned by leveraging channel-wise feature attention. In experiments, we test the proposed AN using four representative neural architectures. In the ImageNet-1000 classification benchmark and the MS-COCO 2017 object detection and instance segmentation benchmark. AN obtains consistent performance improvement for different neural architectures in both benchmarks with absolute increase of top-1 accuracy in ImageNet-1000 between 0.5\% and 2.7\%, and absolute increase up to 1.8\% and 2.2\% for bounding box and mask AP in MS-COCO respectively. We observe that the proposed AN provides a strong alternative to the widely used Squeeze-and-Excitation (SE) module. The source codes are publicly available at \href{https://github.com/iVMCL/AOGNet-v2}{the ImageNet Classification Repo} and \href{https://github.com/iVMCL/AttentiveNorm\_Detection}{the MS-COCO Detection and Segmentation Repo}.more » « less
-
null (Ed.)Training a semantic segmentation model requires large densely-annotated image datasets that are costly to obtain. Once the training is done, it is also difficult to add new object categories to such segmentation models. In this paper, we tackle the few-shot semantic segmentation problem, which aims to perform image segmentation task on unseen object categories merely based on one or a few support example(s). The key to solving this few-shot segmentation problem lies in effectively utilizing object information from support examples to separate target objects from the background in a query image. While existing methods typically generate object-level representations by averaging local features in support images, we demonstrate that such object representations are typically noisy and less distinguishing. To solve this problem, we design an object representation generator (ORG) module which can effectively aggregate local object features from support image( s) and produce better object-level representation. The ORG module can be embedded into the network and trained end-to-end in a weakly-supervised fashion without extra human annotation. We incorporate this design into a modified encoder-decoder network to present a powerful and efficient framework for few-shot semantic segmentation. Experimental results on the Pascal-VOC and MS-COCO datasets show that our approach achieves better performance compared to existing methods under both one-shot and five-shot settings.more » « less
-
null (Ed.)Training a semantic segmentation model requires large densely-annotated image datasets that are costly to obtain. Once the training is done, it is also difficult to add new ob- ject categories to such segmentation models. In this pa- per, we tackle the few-shot semantic segmentation prob- lem, which aims to perform image segmentation task on un- seen object categories merely based on one or a few sup- port example(s). The key to solving this few-shot segmen- tation problem lies in effectively utilizing object informa- tion from support examples to separate target objects from the background in a query image. While existing meth- ods typically generate object-level representations by av- eraging local features in support images, we demonstrate that such object representations are typically noisy and less distinguishing. To solve this problem, we design an ob- ject representation generator (ORG) module which can ef- fectively aggregate local object features from support im- age(s) and produce better object-level representation. The ORG module can be embedded into the network and trained end-to-end in a weakly-supervised fashion without extra hu- man annotation. We incorporate this design into a modified encoder-decoder network to present a powerful and efficient framework for few-shot semantic segmentation. Experimen- tal results on the Pascal-VOC and MS-COCO datasets show that our approach achieves better performance compared to existing methods under both one-shot and five-shot settings.more » « less
-
In natural language processing, most models try to learn semantic representations merely from texts. The learned representations encode the “distributional semantics” but fail to connect to any knowledge about the physical world. In contrast, humans learn language by grounding concepts in perception and action and the brain encodes “grounded semantics” for cognition. Inspired by this notion and recent work in vision-language learning, we design a two-stream model for grounding language learning in vision. The model includes a VGG-based visual stream and a Bert-based language stream. The two streams merge into a joint representational space. Through cross-modal contrastive learning, the model first learns to align visual and language representations with the MS COCO dataset. The model further learns to retrieve visual objects with language queries through a cross-modal attention module and to infer the visual relations between the retrieved objects through a bilinear operator with the Visual Genome dataset. After training, the model’s language stream is a stand-alone language model capable of embedding concepts in a visually grounded semantic space. This semantic space manifests principal dimensions explainable with human intuition and neurobiological knowledge. Word embeddings in this semantic space are predictive of human-defined norms of semantic features and are segregated into perceptually distinctive clusters. Furthermore, the visually grounded language model also enables compositional language understanding based on visual knowledge and multimodal image search with queries based on images, texts, or their combinations.more » « less
