dilated attention in vision

**Dilated Attention** is the **sparse attention pattern that spreads tokens apart to cover a larger effective field while keeping the number of attended neighbors low** — by stepping through spatial positions with a stride greater than one, the model reaches distant tokens with fewer dot products, letting it process ultra-high-resolution inputs without computing full dense matrices. **What Is Dilated Attention?** - **Definition**: An attention variant that attends to keys and values sampled every d tokens along height and width, mirroring the dilation trick from dilated convolutions. - **Key Feature 1**: Stride d controls how sparse the attention grid becomes; d=1 recovers regular attention, while larger d zooms out to global cues. - **Key Feature 2**: Dilation can increase across layers so that early layers see local patches and later layers capture global structure. - **Key Feature 3**: Relative positional encodings pair with dilation to keep token alignment in check even with sparse sampling. - **Key Feature 4**: Works naturally with hybrid attention where some heads use dilation and others stay dense. **Why Dilated Attention Matters** - **Receptive Field Growth**: Each layer jumps farther without costing extra tokens, making it ideal for megapixel segmentation or remote sensing. - **Compute Savings**: Because keys/values are subsampled, complexity drops proportionally to 1/d^2. - **Multi-Scale Fusion**: Dilation matches multi-scale features such as edges, corners, and textures by sampling at appropriate granularities. - **Generalization**: Sparse yet structured sampling prevents overfitting to local noise while still preserving global alignment. - **Compatibility**: Dilation plays well with axial and windowed patterns, letting architects mix and match patterns per head. **Dilation Strategies** **Constant Dilation**: - Use a fixed d per layer for predictable receptive field. - Suitable when the entire dataset has similar spatial scale requirements (e.g., microscopy slides). **Progressive Dilation**: - Increase d with depth, like 1, 2, 4, to gradually enlarge coverage. - Matches the way CNNs increase receptive field by stacking dilated convolutions. **Hybrid Dilation**: - Assign dilation to half of the heads while keeping other heads dense, providing both detail and overview simultaneously. **How It Works / Technical Details** **Step 1**: Build attention neighborhoods by striding across the flattened spatial grid with step d, gathering keys and values only at those positions through strided slicing. **Step 2**: Compute scaled dot product attention on the collected subsets, apply softmax, and gather weighted sums; combine with residual projections and feed-forward blocks. **Comparison / Alternatives** | Aspect | Dilated Attention | Local Window (Swin) | Global Attention | |--------|-------------------|--------------------|------------------| | Coverage | Sparse global | Local with shifts | Dense global | | Compute | O(N/k^2) depending on d | O(Nw^2) | O(N^2) | | Detail | Varies with d | Fixed by w | Full detail | | Best Use | Very high resolution | Balanced performance | Moderate sizes | **Tools & Platforms** - **OpenMMLab**: Offers dilated attention modules for detection and segmentation backbones. - **ConvNeXt / Timm**: Provide dilation parameters inside attention_config dictionaries. - **TensorRT**: Can fuse dilated attention with other vision blocks for inference. - **Visualization Tools**: Use attention rollout maps to ensure dilation still covers critical objects. Dilated attention is **the sparse yet powerful lens that makes Vision Transformers see far without lifting a heavy quadratic burden** — it jumps over nearby redundancy and attends to tokens that truly matter at distant locations.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account