metadata
license: other
license_name: other
license_link: LICENSE
tags:
- tinyml
- edge-ai
- cifar100
- cnn
ฯNet Tiny โ Technical Proof One-Pager
Inventor: Luigi Calcagno
๐ง [email protected]
Patent-pending ฯ-loss + Tiny CNN for True Edge Inference
โ What This Shows
ฯNet Tiny is a custom convolutional neural network with:
- Proprietary ฯ-loss function for tail-class lift.
- Real ONNX โ TensorFlow โ TFLite flow.
- Proven microcontroller-friendly footprint.
Tested vs MobileNetV2 using real CIFAR-100 sample.
๐ Combined Results
Sample: torch.Size([1, 3, 32, 32])
Label: 49 (CIFAR-100)
| Metric | ฯNet Tiny | MobileNetV2 |
|---|---|---|
| Param Size | 0.1148 MB | ~13.37 MB |
| TFLite Size | 117.52 KB | Not feasible |
| ฯ-loss Output | 4.5335 | N/A |
| Inference Time (10 runs) | ~0.0094 s | ~0.0785 s |
โ
ฯNet Tiny deploys .tflite ~117 KB โ runs on STM32, ESP32, Pi Zero.
โ MobileNetV2 cannot run at MCU scale.
๐ Key Points
- Reproducible ฯ-loss benchmark.
- Fully local inference, no cloud.
- Open pilot opportunities.
Built & tested by Luigi Calcagno
Contact: [email protected]