Deep generative models for structured outputs : from graphs and motion to language

Abstract

Generative models learn the distribution of data and draw new samples from it. Deep generative models parameterize this distribution with neural networks, usually conditioned on an application-specific input. They are used in many domains, such as drug discovery, chat and coding assistants, robotics, and animation. This dissertation studies how to design and train deep generative models for structured data, whose parts depend on one another. It follows two thrusts. First, when the explicit structure of the output is what we care about, we capture it in the training recipe, in the model design or the training data. Second, when the structure is implicit, hidden within the data, we make assumptions about it and use them to guide the model design. In the first thrust, GraphPNAS generates neural architectures as directed acyclic graphs with a graph neural network that sees the partial graph; under the same search budget, it finds more accurate architectures than a recurrent generator that flattens the graph into a sequence. Referring Transformer gives each phrase its own query, which decodes the phrase's bounding box and segmentation mask in one stage, improving over earlier grounding methods. AoPS rewrites Olympiad-level forum discussions into over 600,000 step-by-step question–answer pairs, which give higher accuracy after fine-tuning than the raw answers, and dates each problem for a contamination-resistant benchmark. In the second thrust, LatentHOI assumes that hand–object motion splits into a global motion and a per-frame grasp linked by a latent code; with scarce motion-capture data, the factorized model generalizes better to unseen objects than joint diffusion, with less penetration per unit of contact. MoME assumes that a token's meaning depends on its context and replaces each token's single memory embedding with several slots chosen by the hidden state; pretrained on up to 100 billion tokens, it improves downstream scores and, in most comparisons, bits per byte over token-indexed memories, and its routing changes with word sense. In these comparisons, building the structure in gave better results with the same data or compute and made the models more interpretable, since their components correspond to parts of the structure.