MAST: A Tool for Visualizing CNN Model Architecture Searches
Published at Debugging Machine Learning Models at ICLR in New Orleans, LA 2019
Any automated search over a model space uses a large amount of resources to ultimately discover a small set of performant models. It also produces large amounts of data, including the training curves and model information for thousands of models. This expensive process may be wasteful if the automated search fails to find better models over time, or if promising models are prematurely disposed of during the search. In this work, we describe a visual analytics tool used to explore the rich data that is generated during a search for feed forward convolutional neural network model architectures. A visual overview of the training process lets the user verify assumptions about the architecture search, such as an expected improvement in sampled model performance over time. Users can select subsets of architectures from the model overview and compare their architectures visually to identify patterns in layer subsequences that may be causing poor performance. Lastly, by viewing loss and training curves, and by comparing the number of parameters of subselected architectures, users can interactively select a model with more control than provided by an automated metalearning algorithm. We present screenshots of our tool on three different metalearning algorithms on CIFAR-10, and outline future directions for applying visual analytics to architecture search.