MRAM: Multi-Level Recurrent Attention Model

MRAM is a multi-level recurrent attention model for active visual perception. It extends recurrent hard-attention models with a hierarchical recurrent structure and studies the emergence of fixation-like and saccade-like gaze behavior in image classification.

The work analyzes both recognition performance and gaze dynamics, including fixation patterns, saccade behavior, and scanpath structure. It was presented as an oral paper at ICONIP 2025.

Code: pengcheng-pan97/MRAM

Keywords: recurrent attention, hard attention, active perception, human-like gaze, PyTorch.