Search:
Match:
1 results

Analysis

This paper addresses the computational inefficiency of Vision Transformers (ViTs) due to redundant token representations. It proposes a novel approach using Hilbert curve reordering to preserve spatial continuity and neighbor relationships, which are often overlooked by existing token reduction methods. The introduction of Neighbor-Aware Pruning (NAP) and Merging by Adjacent Token similarity (MAT) are key contributions, leading to improved accuracy-efficiency trade-offs. The work emphasizes the importance of spatial context in ViT optimization.
Reference

The paper proposes novel neighbor-aware token reduction methods based on Hilbert curve reordering, which explicitly preserves the neighbor structure in a 2D space using 1D sequential representations.