Large-margin representation learning for small-size datasets

Abstract

This work presents a novel approach combining convolutional layers (CLs) and large-margin metric learning for training supervised models on small-size datasets. This approach has four main components: i) a set of CLs with a global average pooling; ii) an instance selection method to choose anchors and instances of interest; iii) a loss function to induce the weight update of the CLs; iv) a large-margin discriminant. Images are provided to the CLs to generate a latent representation and then used to train a large-margin discriminant. The information of the discriminant (support vectors and predictions) aids in selecting anchors and instances of interest to build a loss function that aims to minimize the distance between chosen samples. The weight update using the loss function seeks to change the latent representation to increase the margin between classes. The proposed method’s advantage is that it can train a filter bank with a small amount of data due to the reduced number of parameters. It also has reduced training costs since only a subset of instances is used in the backpropagation algorithm. Experimental results with a synthetic dataset, a texture image dataset, and three histopathologic image datasets (with textural characteristics) showed that the proposed method converges within a few epochs. It also has a low computational cost, produces close or superior accuracy results than equivalent methods, deals well with imbalanced data, and can adapt the filters to different textures.

Date	11 Jul 2023
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Alessandro Lameiras Koerich (Supervisor) & Alceu Jr. de Souza Britto (Co-supervisor)