Learning to see across domains and modalities

Deep learning has recently raised hopes and expectations as a general solution for many applications (computer vision, natural language processing, speech recognition, etc.); indeed it has proven effective, but it also showed a strong dependence on large quantities of data. Generally speaking, deep learning models are especially susceptible to overfitting, due to their large number of internal parameters. Luckily, it has also been shown that, even when data is scarce, a successful model can be trained by reusing prior knowledge. Thus, developing techniques for \textit{transfer learning} (as this process is known), in its broadest definition, is a crucial element towards the deployment of effective and accurate intelligent systems into the real world. This thesis will focus on a family of transfer learning methods applied to the task of visual object recognition, specifically image classification. The visual recognition problem is central to computer vision research: many desired applications, from robotics to information retrieval, demand the ability to correctly identify categories, places, and objects. Transfer learning is a general term, and specific settings have been given specific names: when the learner has access to only unlabeled data from the target domain (where the model should perform) and labeled data from a different domain (the source), the problem is called unsupervised domain adaptation (DA). The first part of this thesis will focus on three methods for this setting. The three presented techniques for domain adaptation are fully distinct: the first one proposes the use of Domain Alignment layers to structurally align the distributions of the source and target domains in feature space. While the general idea of aligning feature distribution is not novel, we distinguish our method by being one of the very few that do so without adding losses. The second method is based on GANs: we propose a bidirectional architecture that jointly learns how to map the source images into the target visual style and vice-versa, thus alleviating the domain shift at the pixel level. The third method features an adversarial learning process that transforms both the images and the features of both domains in order to map them to a common, agnostic, space. While the first part of the thesis presented general purpose DA methods, the second part will focus on the real life issues of robotic perception, specifically RGB-D recognition. Robotic platforms are usually not limited to color perception; very often they also carry a Depth camera. Unfortunately, the depth modality is rarely used for visual recognition due to the lack of pretrained models from which to transfer and little data to train one on from scratch. We will first explore the use of synthetic data as proxy for real images by training a Convolutional Neural Network (CNN) on virtual depth maps, rendered from 3D CAD models, and then testing it on real robotic datasets. The second approach leverages the existence of RGB pretrained models, by learning how to map the depth data into the most discriminative RGB representation and then using existing models for recognition. This second technique is actually a pretty generic Transfer Learning method which can be applied to share knowledge across modalities.