| Recent machine learning advances in computer vision and speech recognition have been largely driven by the application of supervised neural networks on large labeled datasets, leveraging effective regularization techniques and architectural design. Using more data and computational resources, performance is likely to continue to improve in the future. Despite these nice properties, supervised neural networks are sometimes criticized because their internal representations are opaque and lack the kind of interpretability that seems evident in human perception. For example, detecting a dog hidden in the bushes by looking at its exposed tail is a task that is not yet solved by discriminative neural networks. Another class of challenging tasks is one-shot learning, where only one training example for a new concept is available to the model for training. It is widely believed that learned prior knowledge must be utilized in order to tackle this problem.;My dissertation tries to address some of these concerns by introducing domain-specific knowledge to standard deep learning models. This domain-specific knowledge is used to specify meaningful latent representation with structure, which forces the model to generalize better under certain scenarios. For example, a generative model with latent gating variables that will "switch off" noisy pixels should perform better when encountering noise at test time. A generative model with latent variables representing 3D surface normal vectors should do better at modeling illumination variations. These are the kinds of relatively simple domain-specific modifications that we explore in this thesis. It is true that we should not rely too much on manual engineering and learn as much as possible. This principle is taken seriously and we strike a balance between using too much laborious engineering on the one hand and learning everything from scratch on the other. Adding structure to deep generative models is not only helpful for computer vision applications, but it is also very effective for unsupervised density learning tasks. |
