Neural Network Tuning Using Activation Occurrence Maps

Problem

Deep learning can deliver impressive results. However, finding the optimal architecture for a problem is hard. In contrast to hyperparameters, which can be tuned using various methods, finding the right number of layers and neurons per layer of a network for a given problem is often a trial-and-error process.

Proposal

We use visual analytics to improve the performance of a given pre-trained network (for a given problem). For this, we propose several views that support the user in deciding which parts (of which layers) of a network are to be kept or adapted, to increase the model's performance.