Strategic Freezing

Convolutional neural networks (CNNs) are notoriously data-intensive, requiring significantly large datasets for training accurately in an appropriate runtime. Recent approaches aiming to reduce this requirement focus on removal of low-quality samples in the data or unimportant filters, leaving a vast majority of the training set and model in tact. We propose Strategic Freezing, a new training strategy which strategically freezes features in order to maintain class retention. Preliminary results of our approach are demonstrated on the Imagenette dataset using ResNet34.