Dynamic-Network-inference

This project implements a custom neural network model based on BranchyNet to classify CIFAR-10 images, using an early exit mechanism that enhances computational efficiency. The model is designed with five exit points across different layers, allowing samples to "exit" the network based on confidence thresholds calculated using entropy. This adaptive inference strategy reduces computation time, especially for straightforward samples, without compromising accuracy on more complex inputs.

The implementation includes dynamically set thresholds for each exit layer, enabling efficient decision-making for early exits. Additionally, I analyzed the performance of this approach by examining accuracy and inference time per sample across exit layers, providing insights into the trade-offs between speed and precision.

Tools and libraries like PyTorch, Torchvision, and SciPy were essential in developing and optimizing the model, while Sklearn provided valuable metrics for evaluating classification performance. This project demonstrates a practical, efficient design for neural network models, highlighting adaptive inference techniques ideal for real-time applications where speed and efficiency are critical.