Recent advances in deep learning have mainly relied on Transformers due to their data dependency and ability to learn at scale. The attention module in these architectures, however, exhibits quadratic time and space in input size, limiting their scalability for long-sequence modeling. State Space Models (SSMs), and more specifically Selective SSMs (S6), with efficient hardware-aware implementation, have shown promising potential for long causal sequence modeling. They, however, use separate blocks for each channel and fail to filter irrelevant channels and capture inter-channel dependencies. Natural attempt to mix information across channels using MLP, attention, or SSMs results in further instability in the training of SSMs for large networks and/or nearly double the number of parameters. We present the MambaMixer block, a new SSM-based architecture with data-dependent weights that uses a dual selection mechanism across tokens and channels-called Selective Token and Channel Mixer. To mitigate doubling the number of parameters, we present a new non-causal heuristic of the S6 block with a hardware-friendly implementation. We further present an efficient variant of MambaMixer, called QSMixer, that mixes information along both sequence and embedding dimensions. As a proof of concept, we design Vision MambaMixer (ViM2) and Vision QSMixer (ViQS) architectures. To enhance their ability to capture spatial information in images, we present Switch of Scans (SoS) that dynamically uses a set of useful image scans to traverse image patches. We evaluate the performance of our methods in image classification, segmentation, and object detection. Our results underline the importance of selectively mixing across both tokens and channels and show the competitive (resp. superior) performance of our methods with well-established vision models (resp. SSM-based models).
翻译:暂无翻译