Bagging ensemble for deep learning based gender recognition using test-time augmentation on large-scale datasets

Abstract

We present a bagging ensemble of convolutional networks in combination with the test-time augmentation technique to improve performance on the cross-dataset gender recognition problem. The bagging ensemble combines the predictions from multiple homogeneous models into the ensemble prediction. Augmentation techniques are often used in the learning phase of the CNNs to improve the generalization ability. On the other hand, test-time augmentation is not a common method used in the testing phase of the learned model. We conducted experiments on models trained using different hyperparameters. We augmented the test data and combine the predictive outputs from these network models. Experiments performed on diverse gender datasets, including Adience, AFAD, CelebA, Gallagher, Genki-4K, IMDb, LFW, Morph, VGGFace2, and Wiki, showed that the use of bagging ensemble of convolutional networks and test-time augmentation outperforms standalone models. We obtained the highest cross-dataset accuracy in the literature on seven out of eleven datasets. For the remaining four datasets we reported the cross-dataset results for the first time.According to our experiments, VGGFace2, IMDb, and CelebA datasets provided the highest cross-dataset classification results for most of the test datasets in the gender recognition problem.

A novel ensemble model for detecting fake news

“…Bagging [14] consists of creating several copies of the same model and then training each copy on a random part of the dataset using a special technique called "bootstrapping" to choose the training data. This assembly method is widely used to resolve classification problems such as sentiment analysis [15], cancer detection [16], gender recognition [17]. As far as fake news classification is concerned, Patel et al [18] used a comprehensive approach for the automatic categorization of news articles.…”

How do patients' perceptions and doctors' images impact patient decisions? Deconstructing online physician selection using multimodal data

“…In this paper, after building the proposed network model using the PyTorch framework, we use the relevant dataset for training. Specifically, we choose GENKI-4K and Eyeglasses-Dataset to train the model to recognize smiles and glasses worn [ 60 , 61 ]. However, the age detection problem is different from the above, and the model is trained using the mega age dataset because the data collected in this study are primarily Asian, and different races bring different sensory perceptions [ 62 ].…”