Categorical Distributions are Effective Neural Network Outputs for Event Prediction

Kevin Doran, Tom Baden

Keywords: point processes, temporal data, datasets, neural networks, spike prediction, neuroscience

TL;DR: Outputting a categorical distribution can be used effectively for event prediction. Secondary: we explain performance of neural TPP models wrt model size, model output structure and dataset properties.

Abstract: We demonstrate the effectiveness of using a categorical distribution as a neural network output for the task of next event prediction. We find that training set sizes help explain performance differences between models: when training sets are increased, performance differences largely disappear. We introduce 3 new datasets which provide informative ways to explore model performance; they demonstrate cases where larger models and the use of the categorical output are effective.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 7466