Alexandre Pouget - US grants

DESCRIPTION (Adapted from applicant's abstract): The goal of this proposal is to develop a model of hemineglect - the neurological syndrome resulting from lesions in the parietal lobe - based on the response of single neurons in the parietal cortex. Such a model is an important step toward understanding the neural structure of spatial representations and their role in the control of behavior. Previous models have not attempted to relate the syndrome to the neurophysiological data obtained in behaving monkeys and, conversely, none of the neuronal models of spatial representations have been applied to hemineglect. The approach is based on the basis function theory of spatial representations. This theory, which has been recently developed, was motivated by both theoretical considerations regarding the role of parietal cortex in sensorimotor transformations and the way parietal neurons integrate sensory and posture signals. The hemineglect model combines the basis function theory with classical models of spatial attention and with anatomical constraints on the distribution of neurons across hemispheres. The long-term objective is to show that these three components are sufficient to account for the behavior of neglect patients in a wide range of tasks. Simulations focus on five issues: the line bisection test, the line cancellation experiment, the frame of reference of the deficit, object-centered neglect and recovery. It is argued that understanding spatial representations at the neural level can shed new light on some aspects of the syndrome which cannot be easily explained in psychological terms. This applies, in particular, to the temporary recovery after caloric stimulation of the vestibular system or the fact that multiple frames of reference are affected simultaneously.

The nervous system can sense the world through a variety of sensory modalities providing complementary as well as redundant information about various aspects of the real world. Through a process known as multisensory integration, these sensory signals are merged and used to perform a variety of tasks, such as locating objects in space or identifying the words uttered by a speaker. With NSF funding, Dr. Alexandre Pouget is conducting a strongly theory-driven research program, combining modeling and human studies to investigate the neural basis of this process. The experiments involve asking human subjects to locate their arm in space using multisensory cues. The researchers seek to learn whether subjects can integrate over time three types of cues: visual, auditory and proprioceptive. The optimal strategy in this situation is to compute a weighted sum of those cues, with weights inversely proportional to the reliability of the cues. Intuitively, this makes sense: The system should put greater weights on cues that are more reliable. The funded experiments will determine whether subjects perform this weighted average, and whether those weights adapt when the reliability of the cues is changed from trial to trial. The models will be used to predict human performance and to develop a theory of optimal multisensory integration in neural circuits. The approach relies on a general theory of computation known as Bayesian inferences. This theory has been applied to a variety of problems ranging from sensory processing, high level reasoning (like medical diagnosis), and motor control. Therefore, understanding the neural basis of Bayesian inferences in the context of multisensory integration will have wide implications for general theories of how the brain works, with applications to computer vision and robotics.

The Broader Impacts of this project include the development of a course in computational neuroscience at the University of Rochester and opportunities for undergraduate and graduate students to participate in research. This project will also foster collaboration with a laboratory in France directed by Dr. Jean-Rene Duhamel.

The nervous system can sense the world through a variety of sensory modalities providing complementary as well as redundant information about various aspects of the real world. Through a process known as multisensory integration, those sensory signals are merged and used to perform a variety of tasks such as locating objects in space or identifying the words uttered by a speaker. With funding from the National Science Foundation, Dr. Alexandre Pouget is conducting a strongly theory-driven research program combining modeling, human psychophysics and non-human primate neurophysiologal studies to investigate the neural basis of the process known as multisensory integration. Two factors make multisensory integration difficult: First, the sensory modalities are often in different formats (e.g., the sound and image of the same object are not directly comparable); and Second, the sensory modalities are not equally reliable (e.g., it is typically much easier to tell which word is uttered by a speaker based on sound than on lip movements). Using the framework of Bayesian inference, the research plan is to first develop a neural theory
of multisensory integration that can solve both problems optimally. Then the plan is to perform recording in multisensory areas of awake monkeys to test the validity of the approach. While intuition would suggest that a neuron should respond to the same location in space regardless of the modality, a preliminary model suggests otherwise. In the computational network, the visual and tactile receptive fields of a given neuron do not occupy the exact same location. Experiments will involve recordings from cortical visuo-tactile neurons to test whether their receptive fields behave as predicted by simulations. The temporal aspect of the theory will also be tested through psychophysics experiments in humans. Subjects will be asked to perform sequences of eye movements in the presence of artificial motor error. The theory of optimal integration predicts that
subjects will attempt to correct for their errors in proportion to the reliability of the visual feedback.

This research will involve postdocs, undergraduate and graduate students, and the results will be presented at major multidisciplinary conferences. The results will also have implications well beyond multisensory integration. Indeed, multisensory integration is a subcase of the general problem of Bayesian inference, which is believed to be at the heart of numerous cognitive processes such as object recognition, visual perception, motor control and abstract reasoning.

[unreadable] DESCRIPTION (provided by applicant): Decision making is one of the most critical functions of the human brain, impacting all aspects of our life on time scales extending from tenths of seconds to days. Over the past decade, experimental studies have begun to reveal the neural mechanisms that underlie simple decisions, such as deciding whether an object is moving to right or to the left. For binary decisions about motion perception, the brain's strategy is consistent with a class of race (or diffusion) models: evidence is integrated to form a decision variable, which terminates the process when it reaches a criterion level. In parallel to this work, several groups have developed neural theories of optimal Bayesian inference, focusing more particularly on how neurons represent probability distributions and how they update these distributions over time in a Bayes optimal way. The goal of this proposal is to bring together this theoretical work on optimal Bayesian inference with the experimental data on decision making. We propose an interactive program of theoretical and experimental studies to examine the validity and limitations of a theoretical framework known as probabilistic population codes (or PPC for short). Our studies combine modeling of existing data and design of new experiments that test this PPC theories. We will first develop a neural network model with integrate-and-fire spiking neurons for decision making over a continuous variable (direction of motion) using PPCs. Next, we will simulate the model in a psychophysics experiment to determine its performance and reaction time when tested on N-forced-choice motion discrimination in which the model has to decide between N4, 8 or an infinite number of directions. These predictions will then be compared with the performance on monkeys and humans in the same experiments. Finally, we will use the model to generate predictions about the response of LIP neurons which will then be tested through single and multielectrode recordings while monkeys perform a motion discrimination task. We will also consider an alternative class of models for Bayesian inference, which we call the log probability model. This contrasting hypothesis lies at the heart of several recent studies of Bayesian inference in neural circuits. Understanding the brain mechanisms of decision making will ultimately benefit patients with mental and neurological disorders affecting diverse cognitive functions such as demential, neglect, apraxia and addiction. [unreadable] [unreadable] [unreadable]

The ultimate goal of this project is to provide a novel model of the cognitive and neural basis of numerical cognition, and to use this knowledge to guide the development of new training methods that could improve mathematical abilities in children. The project is a collaboration among investigators at the University of Rochester, Johns Hopkins University, and Cold Spring Harbor Laboratories. Recent research suggests that acuity of numerosity judgments is predictive of success in formal mathematics education, and that similar cognitive processes can be trained by specific kinds of domain-general experience. The core idea is that the firing of neurons encodes a probability distribution, thereby representing simultaneously the most probable sample from the distribution and the variance (i.e., confidence) of the estimate.

This project will develop and test a formal Bayesian model that has the unique feature of naturally accounting for a number of metacognitive factors, a critical but undertested factor in the acquisition of expertise. The primary advantages of this Bayesian approach are its ability to provide a natural description of: 1) how the confidence of a learner relates to the precision of their number knowledge; 2) how a learner can combine information from multiple sources of information about number; 3) how intuitive preferences (also known as prior belief) predict learners' errors; and 4) how improvements in probabilistic inference may benefit the precision of the number sense.