My dissertation research focused on the perception of rhythm. It addressed the question: How do listeners perceive temporal structure in auditory sequences despite significant temporal fluctuations that distort the duration of temporal intervals? The answer I proposed was that internal networks of adaptive, nonlinear oscillations entrain to the complex rhythms that we encounter in the natural world. Such systems have the ability to recognize temporal patterns, to sustain coordinated patterns in a the face of a fluctuating signal, and to reestablish coordination after significant changes, or perturbations. These features enable robust recognition and tracking of complex rhythms, such as those found in improvised music and fluent speech. I designed, implemented, and tested a computational model based on this dynamical systems approach (Large & Kolen, Connection Science , 1995). The contribution of this work is the concept of modeling beat and meter perception as the attractor states of a temporally stable dynamical system; this has implications for many aspects of perception and attention, as I describe below.
The dynamic approach to rhythm perception developed from an investigation of perceptual constancy in the face of surface variation. An example of this is the recognition of a familiar melody in a variation on a musical theme. I conducted a study of music performance in which pianists were instructed to perform notated melodies and improvise variations on them. Next I designed and implemented a neural network, and trained it to produce reduced memory representations for these melodies. Comparison of the network's representation of musical sequences with relative importance measures based on analysis of the music performances supported a reductionist (Shenkerian) account of the perception of musical variation. The primary contribution of this work is its finding concerning the relative importance of events in mental representations for musical sequences (Large, Palmer, & Pollack, Proc. Cog. Sci. Soc. , 1993; Large, Palmer, & Pollack, Cognitive Science , 1995). In addition, these performances and improvisations displayed intricate rhythmic patterning and often extreme rubato, providing excellent examples of temporal complexity and robust tests for the rhythm perception model.
The next focus of my research program has been to extend the dynamic approach to rhythm perception toward a general theory of attention and perception of complex, temporally structured events. We have proposed an attentional theory and evaluated its predictions in a series of psychophysical experiments on time discrimination (Large & Jones, Psychological Review , submitted). Briefly, the theory contends that attentional selectivity in the time domain is rhythmical, and that attentional rhythms entrain to periodicities in naturally occurring events such as speech and music. Empirically we have shown that the ability to detect small deviations in auditory patterns depends strongly upon the rhythmic structure of the pattern, and the context in which the pattern occurs. Our theory explains these and other deviations from Weber's law revealed by our experiments, as well as a number of similar results that have been reported in the psychological literature on temporal acuity. In an independent experimental program I am directly testing predictions regarding attentional selectivity in time, measuring reaction times and discrimination thresholds for events that occur at expected versus unexpected times. A second facet of this work includes further investigations in the domain of music perception. This is a excellent area for the study of time perception because music is a highly temporally structured, natural form of communication that operates without explicit referential semantic content. This aspect of the research program involves collecting musical performances and generating predictions about the perception of time in these very complex patterns using the theoretical model. Predictions are evaluated against listener judgements of phrase structure, auditory stream segregation, and meter perception (Large, Proc. Cog. Sci. Soc. , 1996; Large & Palmer, JASA , 1996; Large & Palmer, JASA , in prep.). This research program links higher-level responses to natural communication sounds (e.g. perception of meter and phase structure) with lower-level responses to carefully controlled laboratory patterns (e.g. temporal acuity and discrimination thresholds) through a model that is capable of predicting both.
Recently I have focused on the possibility that similar temporal processes may play a role at different time scales in auditory perception. We now know that information about the temporal structure of acoustic signals is faithfully preserved at different time scales at virtually every level of the auditory system. Temporal processes clearly play a role in pitch perception, formant tracking, stream segregation, for example. But how does this work? In collaboration with John Crawford's auditory neurophysiology group here at Penn, I am developing a mathematical characterization of neural mechanisms for the processing of communication sounds in the vertebrate auditory system. We focus on a fish (Pollimyrus: Mormyridae) that has a simple auditory system and, most significantly, lacks a cochlea. Because there is no peripheral structure that can be expected to yield a precise mechanical frequency analysis, analysis of communication sounds must occur primarily through temporal processes in the nervous system. Single unit recordings that are being made from the auditory nerve, medulla, and midbrain of this animal are providing a compelling picture of the representation of auditory information as it ascends the auditory pathway. I am working toward a mathematical characterization of the temporal transformation that takes place from primary afferent axons through the auditory midbrain. This research will improve our understanding of the fundamental nature of temporal processes in the auditory system, and it has already generated predictions for physiological experiments that we are designing in the lab.
At the robotics (GRASP) Laboratory, I am investigating the adaptive generation of action sequences, especially as it relates to multi-agent cooperation. The basic approach is to integrate the planning and control of action using nonlinear dynamics by erecting a vector field that governs behavioral variables, such as movement direction and velocity. I have investigated the scalability of this approach, and in particular the ability to incorporate multiple behavioral constraints in a nonlinear vector field governing movement. I have proposed a solution that involves competitive dynamics operating in the space of behavioral constraints and shown that this methodology can be used to design and analyze systems that display complex behaviors. I have implemented this approach in a two-robot system that performs a cooperative navigation task, and I have demonstrated the ability of the individual agents to make decisions and autonomously generate non-trivial behavioral sequences (Large, Christensen & Bajcsy, International Journal of Robotics Research., accepted; Large, Proc. IEEE ISIE ., 1997; Large, Christensen & Bajcsy, Proc. IJCAI , 1997; Large, Christensen & Bajcsy, Proc. IEEE ICRA ., 1997). I have also experimentally evaluated the performance of vision based navigation systems (Venetianer, Large, & Bajcsy, Machine Vision and Applications , 1997). This framework presents numerous possibilities for the design of systems that must interact with humans in increasingly naturalistic physical and/or virtual environments. Finally, in a new research initiative in the GRASP Lab, we are beginning to investigate the use of auditory perception in intelligent robotics, focusing on dynamic approaches for segregation, identification, and localization of sounds in the environment. This research integrates dynamic modeling of auditory perception with the dynamic generation of action sequences to address feedback mechanisms.