evMLP: An Efficient Event-Driven MLP Architecture for Vision

The introduction of evMLP marks a significant step in the evolution of neural network architectures for computer vision. Traditionally dominated by Convolutional Neural Networks (CNNs) and more recently Vision Transformers (ViTs), the exploration of multi-layer perceptrons (MLPs) offers new insights. The evMLP architecture employs an event-driven local update mechanism, allowing it to process only relevant patches in images or feature maps, thus enhancing computational efficiency. By defining 'events' as changes between consecutive frames, evMLP minimizes redundant computations, which is particularly beneficial for sequential image data like video. This innovative approach not only reduces computational costs but also maintains competitive accuracy, as demonstrated through rigorous ImageNet classification experiments and evaluations on various video datasets. The results indicate that evMLP stands as a viable alternative to existing models, potentially reshaping the landscape of vision…