Neural Foundations · 2015
Multi-Scale Context Aggregation by Dilated Convolutions
Introduced a context module built from dilated convolutions that grows the receptive field exponentially with depth while keeping full resolution and coverage, improving dense prediction and establishing dilation as a standard tool for segmentation and audio models.
Editorial record
Plain-language summary
Image classification networks shrink their feature maps with pooling, which loses the detail needed for tasks like semantic segmentation. Dilated convolutions skip input positions inside the kernel so each layer sees a wider area without pooling or extra parameters. Stacking layers with increasing dilation gives a large receptive field at full resolution. The idea was adopted widely, from later segmentation networks to WaveNet's audio generation.
Source record
Provenance
- Record ID
- P-674
- Record created
- 2026-09-26
- Last reviewed
- 2026-09-26
- Record version
- 1
- https://arxiv.org/abs/1511.07122
- arXiv:1511.07122
Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.