Research Objectives
Sequential Information Gathering in Machines and Animals
Biological organisms make accurate behavioral decisions with extraordinary speed and flexibility in real-world environments, despite incomplete knowledge about the state of the world and the long-term effects of their actions. This ability must be shared by intelligent machines if they are to operate flexibly in similar environments. The main goal of the proposed research is to undertake a detailed interdisciplinary study of sequential decision-making across animals and machines, with a focus on real-time learning and control of information-gathering behaviors in complex cognitive tasks such as scene perception, spatial navigation, and language processing.
Members of the lab are pursuing cognitive processing associated with attention and gaze control, real-world scene perception, object and face recognition, spatial navigation, language comprehension and production, and the integration of vision, language, and action in complex environments. Methods include psychophysical and behavioral experiments, formal mathematical modeling and computer simulation, and development of algorithms for sequential behavior in autonomous intelligent machines.
Research in SIGMA Lab is guided by a comparative approach, combining the study of human cognition, insect decision-making, and computational methods from the study of machine systems. All of these systems provide experimentally tractable test-beds for studying real-time acquisition, use, and generation of information sequentially and dynamically over time.
A moment's consideration makes clear the formidable challenges that an animal or an artificial agent faces in deciding what to do at each point in time. The real world offers a dizzying array of stimuli to which the agent could respond. The space of possible actions may be similarly vast, especially for mobile agents. Compounding the problem, the consequences of particular alternative actions may not become apparent until a sequence of related actions has been taken, and yet the agent may need to estimate the likely payoff of a decision (or a sequence of decisions) in advance. Furthermore, decisions may need to be made in quick succession, leaving little time to sort through the space of all possible actions and associated outcomes. Finally a well-designed agent should be able to learn from its failures as well as its successes, which raises the problem of how to assign credit to the correct decision and how to store all this information in memory.
In light of these complexities, it is deeply puzzling how to explain the apparent fluidity and accuracy of sequential decision-making by organisms, or how to endow machines with similar abilities. To gain insights into this problem, we will study two key issues in the human, insect, and robot systems. First, how do current perceptions and past experience interact to guide behavioral decisions? Second, how can complex tasks be learned from delayed reward in partially observable environments?
A major goal of the research is to show how the design of artificial creatures can be guided by, and serve as a guide for, the study of sequential activity in animals. Understanding the challenges that the designers of intelligent machines face, and the formal frameworks that they have developed to tackle these challenges, leads to novel questions about organismal behavior. Similarly insights gained from organisms in turn helps to suggest ways for improving algorithms for building intelligent artificial agents.
