Our ability to quickly and continually make sense of the complex and changing world around us is astounding. It is central to our intelligence and creativity. We do not necessarily process each and every thing we see or hear, but instead we decide what the relevant tidbits of information mean to us. To establish meaning from just a few bits and pieces of information, we make a host of assumptions about the world. Many of these assumptions are hidden and outside of our conscious awareness. The set of assumptions we use and how we use them are largely unknown. My research goal is to uncover the computational nature of these hidden assumptions and how we make and break them (see Research Statement for more details).
My research is guided by three questions:
- How can these hidden assumptions be represented (in computational terms) such that they can be learned flexibly in a variety of ways (incl. under uncertainty) and invoked at a moment's notice?
- What is the role of these hidden assumptions in sense-making? How are these assumptions used to resolve perceptual (linguistic and visual) ambiguity and assist with decision making and action-selection?
- How do we know when these assumptions need to be broken and how do we break them?
AI has seen rapid progress in recent years, especially with the development of new deep neural network architectures. However, despite these efforts, there are many open questions about how to computationalize core human intelligence capabilities of compositionality, non-monotonic reasoning with symbolic knowledge, social intelligence and open-world discovery of new knowledge. Central to my research has been the development and application of techniques in knowledge representation and reasoning as well as in uncertainty processing that have shown promise. These techniques help us elucidate the more abstract computational principles -- associated with making and breaking assumptions -- that are at play during sense-making.
1. Language Understanding through Sense-Making
Traditionally, natural language processing (NLP) has been thought of as a set of processes arranged in a pipeline where syntactic properties are extracted before semantics, which in turn are extracted before pragmatics. However, there has been some evidence in psycholinguistics as well as in cognitive neuroscience of language to suggest that these processes are not linear and sequential. I have begun exploring how language understanding (and particularly resolving linguistic ambiguity) can be achieved by a more holistic constraint satisfaction process - asking if a particular grounding "makes sense" given some background knowledge. I have shown various proof of concepts to suggest that this approach can help resolve pronouns and indirect speech acts.
2. Social Intelligence: Representing and Reasoning with Social Norms
Human behavior is frequently guided by social and moral norms; in fact, no societies, no social groups could exist with- out norms. But, what are these social norms, how are they represented and how are they used? I have explored the role of social norms in deciding how to use objects (e.g., how to grab a knife when using it versus when passing it to someone -- a situation where social safety norms apply). I showed how this form of social intelligence can be computationalized and implemented in an embodied robotic system. More generally, normative aspects are highly pervasive in many different forms of human-human interaction. I showed that centuries old legal precedent can help us categorize and organize our thinking about norms. Legal principles can provide a deeper understanding the normative fabric in society and provide guidance for designing more natural human-robot interactions. At the core of social intelligence is the questions of how norms are represented. I showed that human norms are context sensitive and I have developed computational techniques for learning them under uncertainty.
3. Creative Problem Solving
From time to time, we can break our own assumptions giving us room to assimilate new knowledge and make new discoveries. The processes underlying creative cognition remains largely unknown. There has been over a century of psychology and neuroscience research in insight problem-solving, attempting to understand how humans reframe problems or reconstruct their mental representations. Modern AI systems are able to generate novel works of art and music, but they cannot solve problems that require out-of-the-box thinking. I have begun making some early progress towards formalizing creative problem solving computationally. Doing so has a two-fold benefit: (1) allows us construct potentially solvable computer science problems, and (2) provides a theoretical foundation for constructing stimuli for human-subject experiments.