Jim DiCarlo, MIT
Although object and face recognition is fundamental to our behavior and seemingly effortless, it is a remarkably challenging computational problem because the visual system must somehow tolerate tremendous image variation produced by different views of each object (the "invariance" problem). To understand how the primate brain accomplishes this remarkable feat, we must understand how sensory input is transformed from an initial neuronal population representation (a photograph on the retina), to a new, remarkably powerful form of neuronal population representation at the highest level of the primate ventral visual stream (inferior temporal cortex, IT). In this talk, I will review our results on the ability of the IT population representation to support position-, scale- and clutter-tolerant recognition. I will present a geometric perspective for thinking about how this ventral visual stream constructs this representation ("untangling" object manifolds). Finally, I will show our recent neurophysiologic and psychophysical results that suggest that this untangling is driven by the spatiotemporal statistics of unsupervised natural visual experience. Our long term goal is to use the understanding of this biological computation to inspire artificial vision systems, to aid the development of visual prosthetics, to provide guidance to molecular approaches to repair lost brain function, and to obtain deep insight into how the brain represents sensory information in a way that is highly suited for cognition and action.
Xin Chen, Department of Computer Science and Engineering, York University (Postdoctoral Candidate in Bavelier lab)
Joint Attention is the process of following gaze, pointing or gestures of another person and therefore sharing the experience of observing an object or event. It is a triadic coordination of attention between the one, another person, and an object or event. By using a collaborative attention system that fully implemented bi-directional eye gaze communication, I investigated 1. Collaborative strategy during multiple-person visual search; 2. Micro-level real-time behavioral coordination during visual or physical tasks; 3. Spatial reference during time-critical visual tasks. I will also discuss the pivotal of joint attention in autism studies, and the possibilities of video-game training programs that emphasizing on the development of joint attention and "theory of mind" of autistic children.
Reza Shadmehr, Johns Hopkins University
When the brain generates a motor command, it also predicts the sensory consequences of that command via an "internal model". The reliance on a model appears to make the brain able to sense the world better than is possible from the sensors alone. However, this happens only when the models are accurate. To keep the models accurate, the brain must constantly learn from prediction errors. Here I use examples from saccade and reach adaptation to demonstrate that learning is guided by multiple timescales: a fast system that strongly responds to error but rapidly forgets, and a slow system that weakly responds to error but has good retention.
Chris Sims, Postdoctoral Fellow, BCS/CVS, University of Rochester
How are human cognitive, perceptual, and motor processes organized and coordinated towards the efficient achievement of goals in routine interactive behavior? Despite the simplicity of the question, its answer is, at present, poorly understood. The goal of this research is to develop a unifying explanation for the intelligence and behavioral richness inherent to routine human activity. This explanation centers on the capacity to acquire and exploit internal models of embodied dynamics. Embodied dynamics are the recurring interactions between cognitive, perceptual, and motor processes with external tasks and environments during routine interactive behavior. Internal models are formal constructs that have largely been studied in low-level sensorimotor control, but are here extended to include the dynamics of simple cognitive operations. The emphasis of this research is on how these two elements can be combined to generate novel and surprising predictions regarding the capacities of human performance. The reported research demonstrates, through the convergence of three empirical studies, mathematical optimality analysis, and computational cognitive modeling, the human capacity to acquire and exploit such internal predictive models.
Ed Vul, MIT
We investigate cognitive resources and their limits through modeling of the Multiple Object Tracking (MOT) task. Specifically, we define an ideal observer model and test hypotheses about human resource limitations based on the correspondence between human and model performance. Because the ideal agent optimally allocates resources, this approach allows us to estimate what resources limit human performance without making heuristic assumptions about the relationship between resources and performance. The unlimited ideal observer performs substantially better than humans, but still shows many commonly observed phenomena in human MOT: tracking performance suffers as object speed increases, as the number of objects increases, and as the unpredictability of the object trajectories increases. However, the basic model does not account for the characteristic tradeoff between the speed and the number of targets that humans can track. We tested three potential resources that may limit performance: (1) Limited computation (e.g., a central bottleneck that limits the number of state estimates that may be updated); (2) Limited measurement precision. (e.g., a finite amount of attention can be deployed across the visual field to improve the precision of observations); (3) Limited memory fidelity. (e.g., working memory capacity limits the precision of state estimates propagating through time). Our results are most consistent with memory limiting human performance. Further experiments and modeling document the flexibility of human resource allocation during dynamic tracking tasks. I will close with glimpse at other ongoing work applying statistical decision theory to the allocation of cognitive resources.
David Wozny, UCLA (Postdoctoral candidate in Knill lab)
Integrating information from multiple sensory modalities is an essential process in monitoring the current state of the environment. It would seem natural that evolution has provided us a means for optimally determining which sensory signals correspond to the same event and should be integrated, which signals have come from disparate sources and should be processed separately, and how these signals should learn from each other through experience. I examine the computational rules governing how the nervous system utilizes multiple sources of sensory information by comparing human trisensory perception with a Bayesian observer that allows both integration and segregation. The wide spectrum of multisensory interactions across three different modalities and across behavioral tasks can be well quantified by this Bayesian observer model. Using a Bayesian causal inference model, we quantify behavioral aspects of multisensory learning through a commonly observed, yet not well understood, phenomena of multisensory recalibration in adult humans known as the ventriloquist aftereffect. With this model, we can estimate the likelihoods and priors, and as a result are able to computationally characterize the aftereffect by examining changes in the underlying distributions.
Ben Webb, University of Nottingham, UK
The cortex accumulates evidence from early sensory areas in order to make purposeful decisions. To overcome the inherent ambiguity of local, noisy neural representations in early visual areas and arrive at accurate decisions, cortical pathways combine ('pool') visual information across space and time. Yet we have a poor understanding of the algorithms which govern the pooling process. One popular idea is that the visual system simplifies the combinatorial calculations by deriving summary statistics (e.g. measures of central tendency) of complex visual images in order to guide behaviour. I will present some computational modelling and data from a series of human psychophysical experiments designed to test this hypothesis. The results of this ongoing work suggest that perceptual decision making is an adaptive process that may switch between different strategies as more evidence is accumulated over time.
Krzysztof Palczewski, Department of Pharmacology School of Medicine Case Western Reserve University
Rhodopsin, which absorbs a photon to initiate visual phototransduction, belongs to the superfamily of G protein (guanine nucleotide-binding protein)-coupled receptors (GPCRs), encoded by ~950 genes of the human genome. Mutations in the rhodopsin gene may cause human diseases like retinitis pigmentosa (RP) that usually result in late-onset blindness. Like all GPCRs, rhodopsin has seven transmembrane helices and consists of an apoprotein, the opsin, together with a covalently bound chromophore, 11-cis-retinal.
Mutations in the genes encoding many proteins involved in production of this chromophore have been implicated in causing recessive blinding diseases of humans such as Leber's congenital amaurosis (LCA), Stargardt macular degeneration, congenital cone-rod dystrophy, and retinitis pigmentosa (RP). Promises of therapy for incurable blinding hereditary retinal diseases are more frequently discussed owing to major scientific advances that have increased our understanding of basic disease mechanisms. For instance, mutation in the RPE genes encoding RPE65 is one of several known molecular causes of these blinding diseases, and several therapeutic approaches to treat LCA have been proposed: RPE transplantation, gene replacement therapy, and pharmacological intervention. For example, some of these diseases can be treated by dietary intake of active chromophores or their 9-cis-precursors. We wish to capitalize on recent progress and continue working to understand how vision is triggered by light, is maintained in healthy individuals, and how it can be rescued or regenerated in patients who experience environmental insults on their vision or are prone to vision loss due to their genetic background.
Lindsay Lewis, UCSD (Postdoctoral candidate in Tadin lab)
A number of studies have demonstrated cross-modal responses within visual cortex as a result of blindness. However, little is known about the organizational principles that drive cross-modal plasticity. One possibility is that cross-modal plasticity in visual cortex is pluripotent – cross-modal responses in visual cortex may not show strong selectivity for either modality or task. Alternatively, if cross-modal plasticity is driven by functional specificity, similar activations might be expected for a given task, regardless of modality; whereas if cross-modal plasticity is driven by anatomical connectivity between visual cortex and other sensory cortical areas, similar activations might be expected for a given modality, regardless of task.
Here we present work from two studies. In the first study, fMRI responses to a variety of tasks in auditory and tactile modalities were measured in early blind and sighted subjects. We found cross-modal plasticity (greater fMRI responses in blind than sighted subjects) in visual cortex for all tasks, with many areas showing cross-modal plasticity for all the tasks that we tested – across much of cortex the degree of specialization underlying cross-modal plasticity seems to be relatively weak (pluripotency). However, in ventral regions of visual cortex, we did find evidence for selectivity based on modality (greater response to tactile than auditory tasks), and in occipital-temporal regions of cortex we found modulation of cross-modal response by task.
In our second study we specifically tested the functional specificity hypothesis in visual motion area MT+. To more accurately define MT+, we used two rare sight-recovery subjects. In these subjects MT+ responded to auditory motion, while in visually normal subjects MT+ did not show similar auditory responses. These auditory responses in MT+ were specific to motion compared with other complex auditory stimuli, suggesting that cross-modal plasticity can be influenced by the normal functional specialization of a cortical region. Our results further demonstrate that in the case of sight recovery, cross-modal responses can coexist with regained visual responses within the visual cortex.
Jeffrey Macklis, Harvard University
Co-sponsored with Neurobiology & Anatomy
Given the heterogeneity of CNS neuronal subtypes, and the complexity of their connections, detailed understanding of molecular controls over differentiation, connectivity, and survival of specific neuronal lineages will contribute not only to 1) understanding of the development, evolution, organization, and function of CNS circuitry, but also to 2) support or regeneration of vulnerable populations in neurodegenerative (e.g. ALS, HSP/PLS, HD, PD) or acquired disease (e.g. SCI), to 3) enabling accurate models of neuron type-specific disease (e.g. via ES / iPS cell directed differentiation), to 4) identification of disease genes, and to 5) attempts to functionally repair CNS circuitry. For example, data from our lab demonstrate that new neurons can be added to adult neocortical circuitry via manipulation of transplanted or endogenous precursors in situ (including induction of limited neurogenesis of clinically important corticospinal motor neurons– CSMN– in adult mice), indicating that cellular repair of cortical and cortical output circuitry is possible, if controls over specific lineage differentiation are understood. Using FACS-purified CSMN and other projection neuron populations at critical stages of development in vivo, we have identified both developmentally regulated transcriptional programs of novel and largely uncharacterized genes, and cell-extrinsic controls, that are instructive for development of specific neuron lineages as they develop in vivo (in particular, for CSMN and other projection neuron populations); these control key developmental processes from arealization to subtype-specific differentiation and axonal outgrowth. Loss-of-function and gain-of-function analyses for multiple identified genes and molecules reveal combinatorial molecular-genetic controls over the precise development of key forebrain projection neuron populations that may allow directed control of neural precursors / progenitors / "stem cells" (or ES / iPS cells) toward accurate disease models, neuronal support or regeneration, or functional CNS repair.
Dan Kersten, University of Minnesota
Object size information is critical for appropriate actions across wide range of behaviors, including grasping, navigating around obstacles, deciding on food value, or assessing threat. Yet the computational and neural solutions to size extraction are in large part unknown. The computational problems are two-fold. First, even if the size of an object's image could be extracted from simple measurements of the pattern of two-dimensional retinal intensities, determining the physical size of the object causing the image is confounded by the depth of the object: a big far object and a small near object can project to the same size image. So a solution would seem to require knowledge of three-dimensional depth--a problem that in itself is complex, potentially involving multiple sources of information from the larger context in which the object is seen. Second, determining the size of the retinal image of an object is itself a challenge because it requires determining the spatial bounds of the object's image--a non-trivial computational problem given the typical clutter and occlusions in everyday views of a scene. These two problems suggest that one might see evidence of informational coupling between lower-level cortical areas representing 2D retinotopic spatial information, such as V1, and higher-level regions associated with scene context and depth. In this talk, I will show some visual illusions to illustrate various kinds of information the human visual system uses in determining size and depth. Then I'll describe the results of experiments showing that the spatial extent of human fMRI BOLD activity in V1 shifts anteriorly when an object appears bigger as a consequence of a change in scene context and thus apparent depth, even though the object's retinal size is unchanged. These results suggest that large scale contextual information about depth is brought together with incoming retinal information at the earliest levels of cortical visual processing.
Michael Shadlen, University of Washington
Studies of decision-making in such varied fields as psychology, economics, statistics, political- and computer science must consider the flexible and nuanced way that information bears on the choices that agents make. In neuroscience, the study of decision-making opens a window on the neural basis of many other higher cognitive capacities which also use information in a contingent fashion and in a flexible time frame — free from the immediacy of sensory events or the need to control a body in real time. Recent experiments have begun to expose the neural mechanisms that underlie simple forms of decision-making, in particular those in which a choice must be made based on evidence acquired through the senses. The decisions that arise in these tasks can be thought of as a form of statistical inference: what is the (unknown) state of the world, given the noisy data provided by the sensory systems? The neurobiology inspires us to reformulate many inference/reasoning problems with explicit incorporation of stopping rules and, remarkably, an interplay between time and probability. These insights were anticipated by Alan Turing in his code-breaking work during World War II, and they were developed by Abraham Wald into the field of Sequential Analysis. Besides its mathematical elegance and strategic importance, this computational mechanism may be essential for higher brain function. If so, the principles revealed by the study of decision-making may one day lead to new treatments for neurological disorders affecting our most cherished cognitive abilities.
Babak Razavi, University of Rochester (Advisors: Gary Paige and William O'Neill)
Sathyasri Narasimhan, Department of Optometry, University of Bradford, United Kingdom (postdoctoral candidate in Bavelier lab)
Despite the perceivable variety in one's environment, the visual system is capable of attending to only a few objects at any instance while completely ignoring (or "being blind to") the rest. The term, "Multiple Object Tracking" (MOT), refers to the ability of observers to distinguish and track multiple moving target items among identical items which act as 'distractors'. Since previous studies have already established that human observers can track multiple items simultaneously (e.g. Pylyshyn & Storm, 1988), the next logical questions that arise regarding MOT are as follows: How effectively can human observers detect changes in an object while tracking multiple objects simultaneously? What are the factors that contribute to observers' performance while tracking multiple items? The goal of the current study is to provide answers to these questions. The stimulus used in the current study consisted of several dots moving along linear, non-parallel, left-to-right trajectories. Halfway through the trajectories, one of the dots, the target, deviated clockwise or counter-clockwise. The observers' task was to identify the direction of deviation. Our study consisted of eight experiments using afore-mentioned stimulus. Experiments 1-4 were performed using markedly supra-threshold deviations. The number of trajectories tracked ranged from 4-5 trajectories for +76° deviations to only 1-2 trajectories for +19° deviations. Experiments 5-6 were performed using deviations that were close to threshold. Thresholds rose steeply when the number of distractors was increased from 0 (typical threshold ~ 2°) to 3 (typical threshold > 20°). When all the 'trajectories' were presented statically in a single frame 'deviation' thresholds remained around 4°, even with 10 trajectories in the stimulus (Experiment 7). When a temporal delay was introduced at the trajectory midpoint, thresholds increased steeply with increase of this delay (Experiment 8). A sensory memory hypothesis based on the decay of trajectory-traces is proposed to explain all of these results.
Yang Liu, University of Texas at Austin (postdoctoral candidate in Knill lab)
We studied the empirical distributions of luminance, range and disparity coefficients using a coregistered database of luminance and range images. The marginal distributions of range and disparity have high peaks and heavy tails, similar to the well known properties of luminance wavelet coefficients, but the kurtosis of range/disparity coefficients is significantly larger than that of luminance coefficients. We used generalized Gaussian models to fit the empirical marginal distributions. We found that the marginal distribution of luminance coefficients have the shape parameter p≈0.5, while range and disparity coefficients have p≈0.2, corresponding to a much higher peak. We examined the conditional distributions of luminance, range and disparity coefficients. The magnitudes of luminance and range coefficients have a clear positive correlation. However, this positive correlation is not observed between luminance and disparity coefficients. We used generalized Gaussians to model the conditional distributions of luminance and range coefficients. The luminance coefficients at rough surfaces (large magnitudes of range coefficients) have a significantly different shape parameter (p≈0.8) than that (p≈0.46) of the luminance coefficients at smooth surfaces.
CVS Undergraduate Fellowship Poster Session: Meliora Hall, 2nd Floor, 9:00 am-12:00 pm
CVS Picnic: Genesee Valley Park, 12:00-5:00 pm
Mark Albert, Cornell University (postdoctoral candidate in Knill's lab)
Traditionally, visual development is thought to occur in two distinct stages, an innate stage which occurs before experience and a learning stage in which experience shapes development. These two stages are more interrelated than this traditional view suggests. Spontaneous neural activity is necessary prior to experience for proper development of the neural code. The goal of this work was to demonstrate how this activity is not only necessary, but also sufficient and instructive to guide development. The early, endogenous activity appears to have the appropriate statistical structure to be instructive in the same way that natural visual inputs are instructive in refining the young visual system. The visual system can use the same learning method both before experience on spontaneous neural activity patterns and after experience on natural, visual signals.
Receptive fields in primary visual cortex have the characteristic appearance of 2D gabor wavelets. Efficient coding techniques, such as ICA and sparse coding, have produced these filters by efficiently encoding images of our natural, visual environment. The details of how the visual code develops and how the efficient coding strategy is implemented are unclear. However, we can apply the same general computational principles to development. The same high-level, efficient coding algorithms which help to explain adult visual coding can also be applied to spontaneous patterns of activity to produce analogous visual codes. The spontaneous activity models will be simple and abstract, but based on known retinal, LGN, and V1 physiology. The initial monocular model will provide the first evidence of this approach, and a recent extention to a binocular model will be presented. These examples are intented to promote this 'innate learning' approach as a way to improve our understanding of sensory development.
Michael Stryker, UCSF
The topographic representation of visual space is preserved from retina to thalamus to cortex anf from retina to superior colliculus. Topographic mapping is also preserved in the connections between cortex and colliculus. How is near-perfect precision in each of the connections achieved? We have previously shown that precise mapping of thalamocortical projections requires both molecular cues and structured retinal activity presen before eye-opening. To probe the interaction between these two mechanisms, we studied mice deficient in both ephrin-As and retinal waves. Functional and anatomical cortical maps in these mice were nearly abolished along the nasotemporal (azimuth) axis of the visual space. Both the structure of single-cell receptive fields and large-scale topography were severely distorted. These results demonstrate that ephrin-As and structured neuronal activity are two distinct pathways that mediate map formation in the visual cortex and together account almost completely for the formation of the azimuth map. Despite the dramatic disruption of azimuthal topography, the dorsoventral (elevation) map was relatively normal, indicating that the two axes of the cortical map are organized by separate mechanisms. The outcome of the interaction between neural activity and ephrin-A signaling is somewhat different in the maps of the retinal projection to the superior colliculus, and shows a striking similarity to the predictions of the Koulkov model (2006). The map in the lateral geniculate nucleus is different still, demonstrating that the same mechanisms can result in different outcomes depending on details of timing and intrinsic factors that affect connectivity. Recent results in animals in which the visual map in the superior colliculus is duplicated but fully orderly now demonstrate that the cortical projection to the colliculus is guided by the retinal input if the neurons in the visual system have normal activity.
Ralph Freeman, UC Berkeley
Experiments will be described that are aimed at providing basic information concerning the relationship of the BOLD signal in fMRI and neural activity in the cerebral cortex. The vision system is used for this work. Additionally, studies will be covered on attempts to alter neural organization by application of electrical stimulation with transcranial magnetic stimulation (TMS).
EJ Chichilnisky, The Salk Institute
Retinal ganglion cells assemble inputs from photoreceptors via the retinal circuitry, and send a processed visual image to the brain. Classical, coarse-grained receptive field analysis has yielded a great deal of information about retinal computations performed on photoreceptor signals. However, some aspects of the neural code of the retina require measurements at higher resolution. We have examined how the ganglion cell receptive field is assembled from the inputs of individual cones. Using large-scale multi-electrode recordings from primate retina, combined with high-resolution receptive field mapping, we show how all the cones over a large retinal area provide input to complete, independent mosaics of midget and parasol retinal ganglion cells, which provide the numerically dominant visual input to the brain. These measurements revealed high redundancy in the cone inputs to different ganglion cell types, and low redundancy among ganglion cells of the same type. Surprisingly, parasol cells and ON-midget cells received little input from short-wavelength sensitive cones, while OFF-midget cells received significant input. ON- and OFF-midget cells sampled from long- and middle-wavelength sensitive cones in a non-random fashion. This selective sampling could not be explained by clumping of different cone types, and resulted in a greater degree of color opponency in ganglion cell signals transmitted to the brain.
Josh Gold, University of Pennsylvania
It has been recognized for over 100 years that training can cause long-lasting improvements in the ability to detect, discriminate, or identify sensory stimuli. The prevalence of this phenomenon, called perceptual learning, implies that plasticity is a central feature of normally functioning cortex, even in adults. However, our understanding of the plasticity that underlies perceptual learning is incomplete, especially for vision. My talk will describe recent behavioral, physiological, and computational studies from my laboratory indicating that at least some forms of visual perceptual learning involve changes in how information is read out from sensory cortex to form decisions that guide behavior.
Duje Tadin, University of Rochester
First documented by Aristotle, the motion after-effect (MAE) is defined as an illusory sensation of motion resulting from a PROLONGED adaptation to a moving stimulus. This requirement for prolonged adaptation is in stark contrast with the time courses of neural adaptation, which can occur in tens of milliseconds. If neural adaptation of motion selective neurons gives raise to the perceptual experience associated with the MAE, then why are adaptation durations required to yield MAE orders of magnitude longer than rapid neural adaptation? One possibility is that MAE is a perceptual correlate of slower adaptation processes occurring on the order of seconds. This, however, would greatly diminish a possible functional utility of perceptual adaptation. In our dynamic visual world, moving stimuli typically occupy a given retinal region for only a fraction of a second. Thus, to be functionally useful, motion adaptation should occur on a timescale that is shorter than average eye fixation.
In a series of experiments, we demonstrate that as little as 25ms of adaptation can yield perceivable MAEs, even when the adaptation duration is far below that required for above chance motion discriminations of the adapter. In turn, this indicates that MAE can be observed following an adaptation period that is considerably shorter than previously thought. Evidently, the MAE is not merely a perceptual illusion that follows prolonged exposure to a moving stimulus, but rather a process that can occur essentially every time we experience motion.
In parallel, we also measured neural adaptation for a population of MT neurons in alert macaques. The results revealed that MT neurons exhibit directional selectivity to static stimuli following very brief periods of motion adaptation. The observed directional selectivity to static stimuli was in anti-preferred direction, suggesting a possible neural correlate for our psychophysical results.
Hany Farid, Dartmouth College
In an attempt to quell rumors regarding the health of North Korea's leader Kim Jong-Il, the North Korean government released a series of photographs showing a healthy and active Kim Jong-Il. Shortly after their release the BBC claimed that the photographs were doctored. The article pointed to purported visual incongruities which were claimed to be the result of photo tampering. The BBC was wrong.
Because judgments of photo authenticity are often made by eye, we wondered how reliable the human visual system is in detecting discrepancies that might arise from photo tampering. We describe three experiments that show that the human visual is remarkably inept at detecting simple geometric inconsistencies in shadows, reflections, and planar perspective distortions. We also describe computational methods that can be applied to detect the inconsistencies that seem to elude the human visual system.
Bill Warren, Brown University
How do people generate paths of locomotion through a complex changing environment? Behavioral dynamics studies how stable patterns of behavior emerge from the interaction between an agent and its environment, which is typically non-stationary and unfolds over time. In this talk, I will describe our effort to model the dynamics of visually-guided steering, obstacle avoidance, interception, pursuit-evasion, and shadowing, based on experiments in an ambulatory virtual environment. By combining these components, we seek to predict paths of locomotion in more complex situations, and ultimately to model the collective behavior of crowds. The results demonstrate that locomotor paths can be understood as emerging on-line from the agent-environment interaction, making explicit path planning unnecessary.
Chris Pack, McGill University
Many neurons in the primate visual cortex are highly selective for complex stimuli such as faces and motion patterns. These neurons are typically found in brain regions that are many synapses away from the retina, suggesting that the brain derives its selectivity via successive transformations of the visual image. However, the nature of these transformations is poorly understood, with one important exception: The simple and complex cells discovered by Hubel & Wiesel have been successfully modeled based largely on combinations of feedforward inputs with specific spatial and temporal properties.
In this talk I will describe a combination of experimental and computational approaches that permit the characterization of neurons in the dorsal pathway that extends beyond the primary visual cortex. In particular I will show how the medial superior temporal (MST) area, where neurons exhibit a baffling array of selectivities to optic flow stimuli, can be modeled based on transformations similar to those used to characterize simple and complex cells in the primary visual cortex. This suggests that the brain uses a style of computation that is conserved across visual areas and possibly across sensory modalities.
Jennifer Hunter, University of Rochester
The visualization of individual cells in the living retina has proven invaluable in the study of normal and diseased retinal structure and function. Many challenges, including image degradation from ocular aberrations, scatter, and eye movements, have been overcome, allowing images of blood cells in the smallest capillaries and single cones in the photoreceptor mosaic. Other cell layers in the retina are either transparent or opaque and are thus more difficult to visualize. However, the addition of fluorescence imaging capabilities to a scanning laser ophthalmoscope equipped with adaptive optics makes it possible to successfully image the retinal pigment epithelium, a single layer of cells behind the cones that provide critical support for photoreceptors. This imaging technology has allowed us to assess light damage in vivo at a cellular spatial scale and to establish that cellular damage occurs with yellow light exposures that are below current safety standards. Recent results with two-photon fluorescence imaging in the living eye provide a new method to safely image not only retinal structure, but also to assess retinal function.
Vik Rao, University of Rochester
What are the computational mechanisms that underlie perceptual and cognitive behavior? Any answer to this question must start with the observation that the brain has to work with uncertain information at every level of analysis. The presence of uncertainty has the consequence that the problem of computation in the brain becomes one of probabilistic inference. Indeed, we can recast all cognitive processing as comprising sequential stages of probabilistic inference, performed over data of varying abstraction. In this framework, the goal of processing at a particular level is to infer the variable of interest given the input information and the goal of learning at a particular level is to improve the quality of the inference that is being carried out.
In this thesis we explore and computationally characterize the inference that underlies cognitive processing at multiple levels, using multiple research methodologies. At the neural level, we derive a simple analytic expression that allows for the relation of network properties to the quality of the inference being carried out during neural representation and transmission. This derivation provides an important tool that can be used to elucidate mechanisms leading to efficient inference. We then use this expression to explore the neural mechanisms that underlie the improvements in behavioral performance, observed during perceptual learning. We report that perceptual learning can be neurally mediated through an improvement in the inference process in early sensory areas. Importantly, this model, in addition to accounting for the training induced changes in behavioral performance, also captures the training induced changes in neural properties.
Finally, at the behavioral level, we show that human multi-sensory integration during categorical speech perception is well described by a normative model for optimal inference, thereby providing behavioral evidence for efficient inference in the brain. As opposed to previous studies, the study described here computationally and experimentally probes cue integration in categorical tasks, thereby representing an important extension of previous work since most real-world perceptual tasks involve judgments over categorical dimensions.