Title: Measure the generality of convolutional layers with projection correlation
Authors: Yuan Ke - University of Georgia (United States) [presenting]
Abstract: The transfer learning problem is studied in image classification applications. A phase transition phenomenon has been empirically validated: the convolutional layer shifts from general to specific with respect to the target task as its depth increases. It is suggested that measuring the generality of convolutional layers through an easy to compute and tuning free quantity named projection correlation. The non-asymptotic upper bounds for the estimation error of the proposed generality measure has been provided. Based on this generality measure, a forward adding layer selection algorithm to select generable layers is proposed. The algorithm aims to find a cut-off in the pre-trained model according to where the phase transition from general to specific happens. Then, we propose to transfer only the generable layers as specific layers can cause overfitting issues and hence hurt the prediction performance. The proposed algorithm is computationally efficient and can consistently estimate the true location of phase transition under mild conditions. Its superior empirical performance has been justified by various numerical experiments.