Perforated Vision Pro
75.4% Accuracy Improvement on ResNet-50 via Dendritic Neural Plasticity
---
## Results
| Metric | Original ResNet-50 | With Dendrites | Improvement |
|--------|-------------------|----------------|-------------|
| **Validation Accuracy** | 35.00% | **61.39%** | **+75.4%** |
| **Training Accuracy** | ~40% | ~75% | +87.5% |
| **Error Reduction** | — | — | **40.6%** |
> Our model underwent **10 autonomous restructuring cycles**, growing dendrites to break through performance plateaus.
---
## Introduction
By integrating **Perforated AI's Dendritic Optimization** into **ResNet-50** on **CIFAR-100**, we achieve a **75.4% relative accuracy improvement** through biologically-inspired structural plasticity.
The network **grows** new computational pathways during training, **adapts** across 10 learning cycles, and **compounds** improvements through cyclical restructuring.
---
## Team
**Shobhit Kapoor**
Case Study - https://docs.google.com/document/d/15h97tnoIzcxxOs5xVFLeYHde-76A5MoNnJOkImHeprA/edit?usp=sharing
W&B sweep report(BONUS) - [https://wandb.ai/shobhit1kapoor-university-of-texas-at-dall/Perforated-ResNet-CIFAR100/reports/Perforated-Vision-Pro-Maximizing-Visual-Intelligence-with-Minimal-Data--VmlldzoxNTY4MzE5Mw?accessToken=0qg28wqpzl00nsek3m6tf1sw7940cyurosm67xnhu0n6qyq4hmt3qsnq6dplel9y
](https://wandb.ai/shobhit1kapoor-university-of-texas-at-dall/Perforated-ResNet-CIFAR100/reports/Perforated-Vision-Pro-Maximizing-Visual-Intelligence-with-Minimal-Data--VmlldzoxNTY4MzE5Mw?accessToken=0qg28wqpzl00nsek3m6tf1sw7940cyurosm67xnhu0n6qyq4hmt3qsnq6dplel9y)
---
## Project Impact
### The Challenge
- CIFAR-100: **100 fine-grained classes**, only 500 training images per class
- Standard ResNet-50 achieved only **35%** with default settings
### The Solution
- **Dendritic Neural Optimization** — network grows additional computational pathways
- **Perforated Backpropagation** — intelligent gradient routing through dendrites
- **Cyclical Restructuring** — 10 phases of architecture evolution
### Business Impact
| Impact Area | Benefit |
|-------------|---------|
| Model Performance | +75.4% accuracy without architecture changes |
| Edge AI | Same ResNet-50, better accuracy for mobile/IoT |
| Compute Savings | Better accuracy at same parameter count |
**Applications:** Autonomous vehicles, medical imaging, manufacturing QC, retail visual search
---
## The 10-Cycle Evolution
| Cycle | Epoch | Val Accuracy |
|-------|-------|--------------|
| 1 | ~25 | 38% |
| 2 | ~60 | 51% |
| 5 | ~170 | 58% |
| 10 | ~330 | **61.4%** |
---
## Visualizations
### Training Graph (PAI Output)
### Accuracy Comparison
### W&B Training Dynamics
---
## Usage
```bash
# Clone and setup
git clone https://github.com/Shobhitkapoor1/PerforatedAI.git
cd PerforatedAI/Examples/hackathonProjects/Perforated_Vision_Pro
pip install -r requirements.txt
wandb login
# Run hyperparameter sweep
python resnet_cifar_dendrites.py --count 100
```
### Winning Configuration
```python
{
"improvement_threshold": 1,
"candidate_weight_init": 0.01,
"pai_forward_function": 1, # ReLU
"dendrite_mode": 1, # Gradient Dendrites
"weight_decay": 0.0005
}
```
---
## Technical Stack
- **Model:** ResNet-50 (23M+ parameters, industry standard)
- **Dataset:** CIFAR-100 (100 classes, 50k training images)
- **Framework:** PyTorch + torchvision + Weights & Biases
- **PAI Config:** 10 max dendrites, perforated backpropagation enabled
---
## Bonus Points
| Category | Achievement |
|----------|-------------|
| **Prevalence** | ResNet-50 + CIFAR-100 (industry standard) |
| **Quality** | +75.4% relative gain, 40.6% error reduction |
| **Framework** | PyTorch, torchvision, W&B, CUDA compatible |
| **Reproducibility** | Deterministic seeding, full W&B logging |
---
## Files
| File | Description |
|------|-------------|
| `resnet_cifar_dendrites.py` | Training script with PAI integration |
| `requirements.txt` | Dependencies |
| `ResNet_CIFAR.png` | PAI training visualization |
| `Accuracy_Comparison.png.png` | Before/after comparison |
---
## Resources
- [W&B Dashboard](https://wandb.ai/shobhit1kapoor-university-of-texas-at-dall/Perforated-ResNet-CIFAR100)
- [Perforated AI](https://perforatedai.com)
- [ResNet Paper](https://arxiv.org/abs/1512.03385)
---
## Final Statement
> **Neural networks don't need to be static.**
Our ResNet-50 achieved **75.4% improvement** not through more data or compute, but through **structural intelligence**—the ability to grow and evolve during training.
This is not an incremental improvement. This is a **paradigm shift**.
---
Built with Perforated AI | ResNet-50 | CIFAR-100 | PyTorch