CIFAR-10 Dendritic Comparison

Inspiration

Explore whether biologically inspired dendritic computation can make standard CNNs smarter and more parameter-efficient on CIFAR-10.

What it does

Applies Perforated AI’s Dendritic Optimization to a baseline PyTorch CIFAR-10 model and compares accuracy, model size, and efficiency against the non-dendritic version.

How we built it

Integrated Perforated AI’s dendritic layers into an existing CIFAR-10 CNN, then ran Weights & Biases hyperparameter sweeps to optimize dendritic configurations with minimal code changes.

Challenges we ran into

Tuning dendritic hyperparameters, ensuring training stability, and fairly comparing performance under identical training budgets.

Compute requirements, because we do not have steady access of GPU.

Accomplishments that we're proud of

Achieved competitive or improved accuracy with fewer effective parameters and demonstrated dendritic benefits on a real-world vision dataset.

What we learned

Dendritic computation can increase expressiveness without scaling model size, and works well with existing PyTorch workflows.

What's next for CIFAR-10 Dendritic Comparison

Extend experiments to ResNet-style architectures, test latency/energy metrics, and validate generalization on larger vision datasets.

Built With

• colab
• jupyter
• kaggle
• perforatedai
• perforatedbp
• python
• pytorch