dynamic resolution networks

**Dynamic Resolution Networks** are **networks that adaptively choose the input or feature map resolution for each sample** — processing easy images at low resolution (fast) and hard images at high resolution (accurate), optimizing the computation per sample based on difficulty. **Dynamic Resolution Methods** - **Input Resolution**: Downscale easy inputs before processing — less computation for smaller inputs. - **Feature Resolution**: Use early features at low resolution, upscale only for hard cases. - **Multi-Scale**: Process at multiple resolutions and fuse — attend more to resolution levels that help. - **Resolution Policy**: Train a lightweight policy network to select the optimal resolution per input. **Why It Matters** - **Quadratic Savings**: Computation in conv layers scales quadratically with spatial resolution — halving resolution gives 4× speedup. - **Natural Hierarchy**: Many images have easy-to-classify global structure — low resolution suffices. - **Defect Inspection**: Large wafer images with localized defects don't need full-resolution processing everywhere. **Dynamic Resolution** is **zooming in only where needed** — adapting spatial resolution to each input's complexity for efficient image processing.