AAI_2025_Capstone_Chronicles_Combined

3.) Data Summary

The most important exploration for this project is the runtime comparison to PyTorch for large numbers of concurrent networks. The initial comparison is based on a test case implemented by Will Whitney in his 2021 article on using JAX for parallel training (2021). This test case is a binary classification problem where two interesting spirals are classified (the negative vs. positive variant). The resulting data looks like this with the class labels:

Figure 3.1: Spirals Binary Classification

The goal is to have the model be able to classify any coordinates in the space to be assigned to the first or second spiral. Generally speaking, all of the orange points should be classified into the same group, and all of the blue points should be classified into the other with points in between that fall near those points also getting classified into their group. Of course the training data does not cover the entire space so we expect the model to develop boundaries where it would be uncertain between the two classes. Figure 3.2 provides an example of how we may expect a model to classify given this training data:

131