What it is
PyTorch visual backbones built around self-modifying memory updates during image processing, with pretrained models for classification, detection, and segmentation. Use a visual backbone whose internal memories and learning dynamics evolve while scanning an image, with ready-to-run PyTorch models and standalone SRNL/operator modules.
Key results
- ImageNet-1K: 84.1/85.2/85.6 top-1 for Tiny/Small/Base
- COCO Mask R-CNN: up to 50.5 box AP and 45.0 mask AP
- ADE20K UPerNet: up to 51.8 mIoU with released checkpoints
Shows a vision backbone can update coupled memories and learning behavior within a single image pass while remaining usable across classification and dense prediction.