We just released a pretrained dendritic ResNet-18 model that achieves 4x better parameter efficiency compared to scaling up to ResNet-34.
ImageNet training (from scratch):
- ResNet-18 (11.7M params): 69.76%
- Dendritic ResNet-18 (13.3M params): 71.95%
- ResNet-34 (21.8M params): 73.30%
Adding 1.6M parameters via dendritic connections yields +2.19% accuracy (1.37% per million params), while jumping to ResNet-34 adds 10.1M parameters for +3.54% accuracy (0.35% per million params). **Inference performance:** 4.37ms vs ResNet-34’s 7.48ms (41% faster), 8% slower than ResNet-18’s 4.04ms.
Transfer learning benchmarks:
| Dataset | ResNet-18 | Dendritic-18 | ResNet-34 |
|---|---|---|---|
| Flowers-101 | 87.1% | 87.9% | 87.9% |
| Oxford Pets | 90.8% | 91.4% | 92.6% |
| Food-101 | 81.7% | 82.1% | 83.9% |
Model and code:
HuggingFace link | Open source repo
This is a drop-in replacement for standard ResNet-18. The model is pretrained on ImageNet and ready for transfer learning. We welcome feedback on how it performs on your datasets, this is the first publicly available pretrained dendritic model.