AAI_2025_Capstone_Chronicles_Combined

effectively parallelize and is actually losing performance by trying to vectorize the calculation across the swarm.

Figure 5.4 CPU Performance Comparison for Large Networks

5.2) MNIST Ensemble Model

The experiments on the timing are useful but it is also desirable to see actual performance of the library in real world situations. A classic problem to test algorithmic performance is the MNIST digit classification dataset. Both libraries use the same model architecture, two hidden layers with 64 and 32 nodes respectively and ReLu activations between them. A PyTorch model is trained for a baseline of performance and compared to a swarm of TurbaNet models. Training a swarm of TurbaNet models on the same data would result in incredibly similar models that would provide no real benefit over a single model, so instead we use the swarm of trained models to create an ensemble model which averages the predictions from each constituent model. Each of the models can be trained on a portion of the dataset and while it may be a weaker predictor on its own, combined they can be stronger.

140