Vision Permutator: A Permutable MLP-Like Architecture for Visual Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current MLP-like models often flatten spatial dimensions, losing crucial positional information.
• Vision Transformers (ViTs) use attention, which can be computationally intensive.
• Convolutional Neural Networks (CNNs) rely on spatial convolutions for feature extraction.

Purpose of the Study:

• To introduce Vision Permutator, a novel MLP-like architecture for visual recognition.
• To develop a data-efficient model that effectively captures spatial information.
• To offer a competitive alternative to CNNs and ViTs without relying on attention or convolutions.

Main Methods:

• Vision Permutator encodes height and width dimensions separately using linear projections.
• It aggregates features in a mutually complementing manner.
• The architecture avoids flattened spatial dimensions and attention mechanisms.

Main Results:

• Vision Permutator achieves 81.5% top-1 accuracy on ImageNet with 25M parameters, outperforming comparable CNNs and ViTs.
• Scaling to 88M parameters yields 83.2% top-1 accuracy.
• The model demonstrates strong performance without large-scale pre-training data.

Conclusions:

• Vision Permutator offers a conceptually simple and data-efficient approach to visual recognition.
• The model effectively captures long-range dependencies and positional information.
• This work encourages new methods for encoding spatial information in MLP-like models.