Learning Spatial and Temporal Inference Machines
Abstract
Reasoning about space, time, and structure lies at the heart of intelligence—both natural and artificial. Nonetheless, AI systems still struggle when they need to reason about space and time, or propagate information across complex latent structures, particularly if they simultaneously need to learn to recognize new concepts or work with uncertainty. So, we propose to develop new representations, languages, algorithms, and theory for structured, spatiotemporal, uncertain reasoning, leading to a qualitative change in our ability to design, learn, and scale AI systems for real problems. We will base our approach on probabilistic languages, particularly templated probabilistic graphical models: these are the de facto standard for structured models of uncertain processes. A key problem, though, is that current uses of these probabilistic languages suffer serious practical and theoretical difficulties. Approximate inference and latent state make model learning intractable: good parameters exist but are impossible to find. As a result, in practice, AI reasoning systems can only put together short chains of inference on learned concepts before performance suffers unacceptably. To overcome this problem, we propose a radical break from traditional methods for learning and using models in probabilistic languages: we claim that clean separation of modeling and inference, often touted as a benefit of these languages, is in fact the obstacle that prevents scaling to hard AI problems. Instead, we propose to learn inference procedures directly: the model is the inference procedure. We will demonstrate that this idea lets us learn accurate high-level models of complex, structured, uncertain, spatiotemporal processes, and solve reasoning tasks that require propagating latent information across these models. Example reasoning tasks include looking at a photograph of a scene and answering “what is about to happen?”; looking at a diagram of an interconnected system and answering “how will a change in this input affect the overall behavior?”; or following a human as he or she solves an exam problem and answering “what is the next step in the solution?” Central to our approach is the notion that prediction can be a lingua franca for AI: every internal representation—whether inference step or latent variable—can be interpreted as a prediction about some combination of observables. It is this viewpoint that provides compositionality: we can divide a reasoning system into components, and as long as each component can understand predictions about the variables it is concerned with, it can communicate with other components. Prediction binds together inference and learning, enables efficient learning of latent state, and leads to good performance on the array of problems we seek to tackle. Our preliminary work proves our ideas’ promise and achieves state-of-the-art performance on several benchmark tasks. Much work remains: we propose to develop a language for building inference machines, methodologies for discovering latent structure, a practical theory that enables effective composition of chunks of learned inference, and larger-scale demonstrations of the approach. We will demonstrate learning and reasoning first on simple simulated systems, and then on systems of growing complexity.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 08, 2016
- Source ID
- N000141512365
Entities
People
- Geoffrey Gordon
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy