MambaOut: Do We Really Need Mamba for Vision? This paper investigates whether Mamba, a state-space model (SSM)-based architecture with RNN-like properties, is necessary for vision tasks. Mamba was introduced to address the quadratic complexity of attention mechanisms and has been applied to vision tasks. However, its performance on vision tasks is often inferior to convolutional and attention-based models. The authors hypothesize that Mamba is not necessary for image classification on ImageNet, as this task does not align with the long-sequence or autoregressive characteristics that Mamba is suited for. However, Mamba may still be beneficial for tasks like object detection and segmentation, which align with the long-sequence characteristic. To test these hypotheses, the authors developed a series of models called MambaOut, which are based on Gated CNN blocks without the SSM. Experimental results show that MambaOut outperforms visual Mamba models on ImageNet image classification, indicating that Mamba is indeed unnecessary for this task. However, MambaOut does not match the performance of state-of-the-art visual Mamba models in detection and segmentation tasks, demonstrating the potential of Mamba for long-sequence visual tasks. The authors conclude that Mamba is ideally suited for tasks with long-sequence and autoregressive characteristics. However, since most vision tasks do not meet these criteria, Mamba may not be necessary for image classification. Nevertheless, Mamba may still be useful for tasks like object detection and segmentation, which align with the long-sequence characteristic. The results of this study suggest that MambaOut, which does not include the SSM, can serve as a natural baseline for future research on visual Mamba models.MambaOut: Do We Really Need Mamba for Vision? This paper investigates whether Mamba, a state-space model (SSM)-based architecture with RNN-like properties, is necessary for vision tasks. Mamba was introduced to address the quadratic complexity of attention mechanisms and has been applied to vision tasks. However, its performance on vision tasks is often inferior to convolutional and attention-based models. The authors hypothesize that Mamba is not necessary for image classification on ImageNet, as this task does not align with the long-sequence or autoregressive characteristics that Mamba is suited for. However, Mamba may still be beneficial for tasks like object detection and segmentation, which align with the long-sequence characteristic. To test these hypotheses, the authors developed a series of models called MambaOut, which are based on Gated CNN blocks without the SSM. Experimental results show that MambaOut outperforms visual Mamba models on ImageNet image classification, indicating that Mamba is indeed unnecessary for this task. However, MambaOut does not match the performance of state-of-the-art visual Mamba models in detection and segmentation tasks, demonstrating the potential of Mamba for long-sequence visual tasks. The authors conclude that Mamba is ideally suited for tasks with long-sequence and autoregressive characteristics. However, since most vision tasks do not meet these criteria, Mamba may not be necessary for image classification. Nevertheless, Mamba may still be useful for tasks like object detection and segmentation, which align with the long-sequence characteristic. The results of this study suggest that MambaOut, which does not include the SSM, can serve as a natural baseline for future research on visual Mamba models.