Deep learning for attribute prediction and small sample size problems

The first part of this work focuses on facial attribute prediction using a novel deep learning formulation, termed as R-Codean autoencoder. The work presents Cosine similarity based loss function in an autoencoder which is then incorporated into the Euclidean distance based autoencoder to formulate R-Codean. The proposed loss function thus aims to incorporate both magnitude and direction of image vectors during feature learning. Inspired by the utility of shortcut connections in deep models to facilitate learning of optimal parameters, without incurring the problem of vanishing gradient, the proposed formulation is extended to incorporate shortcut connections in the architecture. The proposed R-Codean autoencoder is utilized in facial attribute prediction framework which incorporates patch-based weighting mechanism for assigning higher weights to relevant patches for each attribute. The experimental results on publicly available CelebA and LFWA datasets demonstrate the efficacy of the proposed approach in addressing this challenging problem. The second part of this work focuses on understanding the limits of data augmentation by synthesized images by Generative Adversarial Networks. Using extensive experiments, we investigate the question of ’how much augmentation is good augmentation’, in terms of the ultimate aim of classification. Experiments with multiple GAN architectures and classifiers across datasets further substantiate our findings. In the final part,we present new architectural and training procedures for Generative Adversarial Networks(GAN)[16] .We first determine whether the probability distribution of data generated by GANs is equal(or close enough) to the probability distribution of the dataset. For this purpose we train Deep Convolutional Network Classifiers[32, 37] on the generated images and compare them with those trained on the dataset. We show that the probability distributions in both the cases differ.We conduct this analysis on the MNIST[36], CIFAR10[31] and Adience[12] datasets. Having shown the above, we suggest changes in the current GAN training algorithm using the subclass information present in datasets. We show improved results using the CIFAR100[31] and Adience(each of these datasets have the subclass information present) and follow the standard protocol defined on these datasets for training the GAN.