Abstract: Deep Learning has shown clear success in applications involving audio, video, images and, more recently, NLP. When considering Tabular Data, especially diverse or heterogeneous datasets, Deep Learning is often dismissed or else shown to fall short of more popular approaches such as XGBoost. The difficulty of training deep learning models is often cited as a prime reason to avoid making use of the technique, which likely stems from the reality that default settings of your favorite framework likely won't produce a good model. Recently there have been efforts to mitigate this issue with libraries such as FastAI and significant research on hyper optimization applied to neural networks, but on many datasets, accuracy still falls short, training time is computationally unreasonable, or a combination of the two. But don't lose hope. By taking a disciplined approach to tuning hyperparameters, leveraging some recent techniques, and building some intuition, Deep Learning can be a useful approach to learning heterogenous tabular data.
Bio: Jason McGhee is a machine learning engineer at DataRobot, primarily focused on neural networks and deep learning. Jason has a background in leading and building top quality products, previously co-founding Cursor, acquired by DataRobot, and serving as a Senior Software Engineer at Pandora Media. He has a computer science degree from UC Berkeley.