Multiscale Self Attentive Convolutions for Vision and Language Modeling
This work addresses the need for more flexible attention mechanisms in computer vision and NLP, though it appears incremental as it builds on existing self-attention paradigms.
The paper tackled the problem of extending self-attention mechanisms to 2D signals like images and generalizing them beyond 1x1 convolutions, resulting in novel operators such as 2D self-attention, Self Attentive Convolutions (SAC), and Multiscale SAC (MSAC) for vision and language modeling.
Self attention mechanisms have become a key building block in many state-of-the-art language understanding models. In this paper, we show that the self attention operator can be formulated in terms of 1x1 convolution operations. Following this observation, we propose several novel operators: First, we introduce a 2D version of self attention that is applicable for 2D signals such as images. Second, we present the 1D and 2D Self Attentive Convolutions (SAC) operator that generalizes self attention beyond 1x1 convolutions to 1xm and nxm convolutions, respectively. While 1D and 2D self attention operate on individual words and pixels, SAC operates on m-grams and image patches, respectively. Third, we present a multiscale version of SAC (MSAC) which analyzes the input by employing multiple SAC operators that vary by filter size, in parallel. Finally, we explain how MSAC can be utilized for vision and language modeling, and further harness MSAC to form a cross attentive image similarity machinery.