Abstract: Target encoding plays a central role when learning Convolutional Neural Networks. In this realm, one-hot encoding is the most prevalent strategy due to its simplicity. However, this so widespread encoding schema assumes a flat label space, thus ignoring rich relationships existing among labels that can be exploited during training. In large-scale datasets, data does not span the full label space, but instead lies in a low-dimensional output manifold. Following this observation, we embed the targets into a low-dimensional space, drastically improving convergence speed while preserving accuracy. Our contribution is two fold: (i) We show that random projections of the label space are a valid tool to find such lower dimensional embeddings, boosting dramatically convergence rates at zero computational cost; and (ii) we propose a normalized eigenrepresentation of the class manifold that encodes the targets with minimal information loss, improving the accuracy of random projections encoding while enjoying the same convergence rates. Experiments on CIFAR-100, CUB200-2011, Imagenet, and MIT Places demonstrate that the proposed approach drastically improves convergence speed while reaching very competitive accuracy rates.

Abstract: Face analysis in images in the wild still pose a challenge for automatic age and gender recognition tasks, mainly due to their high variability in resolution, deformation, and occlusion. Although the performance has highly increased thanks to Convolutional Neural Networks (CNNs), it is still far from optimal when compared to other image recognition tasks, mainly because of the high sensitiveness of CNNs to facial variations. In this paper, inspired by biology and the recent success of attention mechanisms on visual question answering and fine-grained recognition, we propose a novel feedforward attention mechanism that is able to discover the most informative and reliable parts of a given face for improving age and gender classification. In particular, given a downsampled facial image, the proposed model is trained based on a novel end-to-end learning framework to extract the most discriminative patches from the original high-resolution image. Experimental validation on the standard Adience, Images of Groups, and MORPH II benchmarks show that including attention mechanisms enhances the performance of CNNs in terms of robustness and accuracy.