HLM-Vision-Mix - CIFAR-10
HLM-Vision-Mix is a compact image-classification research model that combines a convolutional stem, a 2D spatial mixer, polynomial-Hopfield blocks, and mean pooling.
This repository stages two CIFAR-10 variants:
| Variant | Parameters | Reported validation accuracy | File |
|---|
| Small | 4.35M | 78.6% | small_model.pt |
| Large | 22.57M | 83.9% | large_model.pt |
These are not state-of-the-art CIFAR-10 models. The value of the release is the architecture/probe record: the spatial mixer fixed an earlier patch-only failure mode, and patch-shuffle probes were substantially more damaging than color-shuffle probes in the small model.
Files
| File | Purpose |
|---|
small_model.pt | Sanitized model-only checkpoint for the 4.35M variant |
large_model.pt | Sanitized model-only checkpoint for the 22.57M variant |
config.json | Public architecture and metrics metadata |
small_training_meta.json | Sanitized training metadata for the small variant |
large_training_meta.json | Sanitized training metadata for the large variant |
small_probe_v4.json | Small-model perturbation probe summary |
Intended Use
- Research on HLM-style vision blocks.
- Architecture and perturbation-probe comparisons.
- Educational CIFAR-10 experiments.
Limitations
- Accuracy is below standard CIFAR-10 CNN/ResNet recipes with modern augmentation.
- No robustness, safety, or production-vision claim.
- The probe results support a research hypothesis; they are not a full vision benchmark suite.
