Research Areas - Computational Neuroscience Research in the Computer Science Department at Carnegie Mellon

CSD faculty: Tai-Sing Lee, Mike Lewicki, Tom Mitchell , Jay McClelland (Psych), Dave Touretzky

Carnegie Mellon has long been active in research at the intersection between AI and Cognitive Science, with research on computational models of human cognition such as Anderson’s ACTR system, Newell’s SOAR, and the development of several important connectionist learning procedures, and recently more broadly with the efforts of the Center for the Neural Basis of Cognition that involve a number of CSD faculty (Lee, Lewicki, Mitchell and Touretzky). As an interdisciplinary field, computational neuroscience is closely aligned with computer science in that it goes beyond applying computational techniques to biology but is in fact focused on the study of the computational principles of the brain itself. Today, CSD computational neuroscience faculty as well as many other CSD faculty are making closer contact with neuroscience in their quest for a better understanding of the computational principles of neural computation and natural intelligence.

1 Faculty

CSD is the home department for three computational neuroscientists: David Touretzky, TaiSing Lee, and Michael Lewicki. Michael Lewicki and Tai-Sing Lee have been hired in recent years in collaboration with the Center for the Neural Basis of Cognition (CNBC) in a strategic plan to foster the growth of new research in neural cognition. Both hold tenure-track appointments in CSD and have extensive training in neurophysiology. Their computational research makes close contact with real neural systems: Lee in particular maintains a primate neurophysiology laboratory for carrying out computationally motivated neurophysiology experiments. Recently, Tom Mitchell also joined the field by investigating the use of machine learning techniques to analyze and model human fMRI signals. In computer graphics, Jessica Hodgins and Nancy Pollard are beginning to integrate computational neuroscience into their studies of movement control and perception. Finally, from the psychology department, David Plaut, John Anderson, and James McClelland, who held courtesy appointments in CSD, have also participated in supervision of CSD graduate students.

2 Research themes

Computational neuroscience research in CSD is distinguished by three key traits: (1) principled approaches with an emphasis on computational theory, (2) rigorous investigation of real neural phenomena and/or implementation of real computational systems, and (3) synergistic integration with robotics, AI and computer science research. These traits are characteristic of the tradition of interdepartmental collaboration found in most computer science research at Carnegie Mellon. One long term goal of our research is to understand neural processing in terms of fundamental computational principles. Two areas in which this approach have been particularly successful are neural coding and perceptual inference. These theories furnish elegant explanations of the observed data describing receptive fields of cochlea neurons and of simple cells in the primary visual cortex. They additionally make predictions about optimal neuronal adaptation to natural signals and hierarchical computation in the visual cortex. Current efforts focus on the nature of the encoding of higher order sensory information. Because of their related research interests, the Lee and Lewicki groups have regular meetings together and are currently exploring synergistic collaboration.

Another major research goal is to understand the nature of learning and computation in interactive, hierarchical neural systems. McClelland and Touretzky have studied the hippocampus in particular, developing important new theories of learning and memory and spatial reasoning. Some of Touretzky’s studies concern spatial representations in the rodent brain, such as place cells in the hippocampus and head direction cells in related areas. His recent work includes an attractor neural network model of deformation of the hippocampal cognitive map. Changes in this map may reflect an animal’s “reconceptualization” of its environment, and research here may shed light on how abstract mental representations are formed in the brain. Another research area is the modeling of reward prediction in the dopamine system. Meanwhile, McClelland, Lewicki and Lee have achieved important insights into hierarchical interactive systems for perceptual inference. These works have led to novel ideas and frameworks for how the cortex works as a distributed interactive system. With the support of neurophysiological data from the laboratories of Lee and others, these ideas have started to bring about a paradigm shift in the scientific conception of the functional roles of primary visual cortex.

One hallmark of research at Carnegie Mellon is its emphasis on practical application of theoretical ideas. Computational neuroscience research in CSD is no exception; indeed, it is distinguished by its close contact with contemporary robotics and computer vision research efforts. David Touretzky is continuing to develop Tekkotsu, an open-source, high level programming system for the Sony AIBO robot dog, drawing inspirations from ideas in cognitive science. Tai-Sing Lee, meanwhile, is presently developing computer vision programs for perceptual organization and 3-D scene inference. Michael Lewicki is pursuing the implementation of his ideas to improve cochlea implant technology and compact image and sound coding. John Anderson brings his computer tutor systems, developed based on his cognitive research, to high school classroom. Recently Tom Mitchell has developed new approaches to analyzing human brain activity captured by functional Magnetic Resonance Imaging (fMRI), by using machine learning methods. For example, his group has successfullytrained their machine learning systems to distinguish whether a human subject is reading words about tools (e.g., hammer, saw) versus buildings (e.g., palace, hut), with over 80in fMRI images. His group has been extending these algorithms to track multiple overlapping cognitive processes from fMRI time series data, and is working in collaboration with Marcel Just in the Psychology department to apply these approaches to study human language processing. Altogether, this group of computational neuroscientists capitalize on the artificial intelligence, robotics, and computer vision traditions and strengths of Carnegie Mellon to contribute to the study of brain functions and make their research relevant to the real world.

3 Education

Carnegie Mellon’s computational neuroscience affiliated faculty are also dedicated to instruction. CSD faculty alone offer four courses on computational neuroscience, including David Touretzky’s ”Artificial Neural Networks” and ”Computational Models of Neural Systems”, Lewicki’s ”Computational Perception and Scene Analysis”, and Lee’s ”Computational Neuroscience”. These courses have attracted both undergraduate and graduate students over the years and have become an integral part of the computer science culture at Carnegie Mellon. The roster will be expanding soon with Touretzky’s AIBO-based “Cognitive Robotics” course. More generally, CSD has introduced a CNBC track within its Ph.D. graduate program to usher graduate students into the tradition of interdisciplinary computational neuroscience research, and a matching computational neuroscience track in the undergraduate computer science program is under consideration. CNBC’s NSF IGERT training program provided additional funding support to encourage computer science students to receive training in neurophysiology and neuroscience students to receive training in computational methods, which helps to establish a vigorous interdisciplinary research community beyond the walls of the department. A Brain Science Seminar has recently been organized in Wean Hall, bringing students and faculty from different laboratories and research areas together to brainstorm.