CIFAR-10 ResNet18 PerforatedAI Dendritic Optimization

Inspiration

I wanted to test whether Perforated AI’s Dendritic Optimization could improve a real computer vision model without changing the dataset or switching to a larger architecture. CIFAR-10 with ResNet18 is a strong baseline, so it was a good benchmark to validate accuracy improvement using dendrites.

What I built

I built a CIFAR-10 image classifier using ResNet18 in PyTorch and created two versions:

A baseline ResNet18 training pipeline (no dendrites)
The same ResNet18 architecture initialized and trained with PerforatedAI dendrites

How I built it

Dataset: CIFAR-10 with standard augmentation (random crop + horizontal flip)
Model: ResNet18 modified for CIFAR-10 (3x3 conv, removed maxpool, 10-class final layer)
Baseline training: standard Adam optimizer + CrossEntropyLoss
Dendritic training: initialized the same model using PerforatedAI's initialize_pai() and trained it using the same setup
I also generated the required PerforatedAI output graph (PAI.png) to verify dendrites were applied correctly

Results

Baseline ResNet18 test accuracy: 85.71%
Dendritic ResNet18 test accuracy: 86.97%

Remaining Error Reduction: Baseline error = 100 - 85.71 = 14.29
Dendritic error = 100 - 86.97 = 13.03

$$ \text{Remaining Error Reduction} = \frac{14.29 - 13.03}{14.29} \times 100 = 8.81\% $$

This shows dendritic optimization reduced the remaining error by 8.81%.

Challenges I faced

PerforatedAI initialization produced warnings related to BatchNorm layers not being wrapped, which I learned is expected behavior.
Training took longer than the baseline due to dendritic optimization overhead.
Locating and organizing the correct PerforatedAI output graph for submission required careful checking of generated artifacts.

What I learned

This project helped me understand how dendritic optimization can improve accuracy while keeping the model architecture mostly the same. I also learned how to structure a clean hackathon submission with reproducible code, results, and required output visualizations.