AAI_2025_Capstone_Chronicles_Combined

With this experiment, the goal was to show that TurbaNet models would achieve comparable predictive accuracy to PyTorch models while training significantly faster due to simultaneous execution. Model performance was evaluated by comparing prediction errors between individual PyTorch and TurbaNet models, as well as analyzing overall runtime efficiency. Each LSTM model consisted of 32 hidden LSTM cells, a single output node, and was optimized using the Adam optimizer with a learning rate of 1e-5. Mean squared error (MSE) loss was used as the performance metric. Finally, these models were compared on a basis of train time and the average error on the predictions across the test set for each stock.

5.) Results

5.1) Parametric Analysis

To validate that the TurbaNet library functions as intended, a parametric study was conducted to evaluate how runtime scales with increasing swarm size. It was expected that PyTorch would exhibit approximately linear scaling—doubling the swarm size should roughly double the training time. With TurbaNet the expectation is that it should hardly scale with the swarm size until the GPU's memory limit is exhausted at which point we should see it begin to scale as well.

5.1.1) GPU 1 Swarm Sweep Results

Presented below are the results of a sweep of swarm sizes from a single network up to 128 networks being trained in parallel. The networks were trained on a standard two-spiral classification problem where the input is a pair of floats and the output is another pair of floats. The training was done with the same number of epochs (1024) and batch size (128). Each trial was conducted ten times to mitigate variability in training time caused by random system fluctuations, and the results were averaged

136