SCALING VISION MAMBA MODELS

Summary from the TLDR AI newsletter

This paper introduces a stochastic layer-wise shuffle regularization technique to overcome overfitting in Vision Mamba models, enabling them to scale up to 300M parameters while maintaining competitive performance with Vision Transformers (ViTs).