Redesigning Embedding Layers for Queries, Keys, and Values in Cross-Covariance Image Transformers

초록

There are several attempts in vision transformers to reduce quadratic time complexity to linear time complexity according to increases in the number of tokens. Cross-covariance image transformers (XCiT) are also one of the techniques utilized to address the issue. However, despite these efforts, the increase in token dimensions still results in quadratic growth in time complexity, and the dimension is a key parameter for achieving superior generalization performance. In this paper, a novel method is proposed to improve the generalization performances of XCiT models without increasing token dimensions. We redesigned the embedding layers of queries, keys, and values, such as separate non-linear embedding (SNE), partially-shared non-linear embedding (P-SNE), and fully-shared non-linear embedding (F-SNE). Finally, a proposed structure with different model size settings achieved 71.4%,77.8%, and 82.1% on ImageNet-1k compared with 69.9%,77.1%, and 82.0% acquired by the original XCiT models, namely XCiT-N12, XCiT-T12, and XCiT-S12, respectively. Additionally, the proposed model achieved 94.8% in transfer learning experiments, on average, for CIFAR-10, CIFAR-100, Stanford Cars, and STL-10, which is superior to the baseline model of XCiT-S12 (94.5%). In particular, the proposed models demonstrated considerable improvements on the out-of-distribution detection task compared to the original XCiT models.

키워드