July 18, 2014
Vision Symposium (Corey Ziemba，Jonathan Winawer, Anna Roe, Ed Connor )
10:00-10:40 Corey Zimba (New York University)
Textural tolerance increases from V1 to V2
10:50-11:30 Jonathan Winawer (New York University)
Spatial integration in the human visual pathways
13:00-14:00 Anna Roe (Vanderbilt University)
High spatial resolution brain imaging and targeted stimulation in nonhuman primates
14:15-15:15 Ed Connor (Johns Hopkins University)
An Efficient Shape Coding Strategy in Visual Area V4
Abstract:
Corey Zimba
Textural tolerance increases from V1 to V2
Visual perception depends on neurons in the visual cortex displaying selectivity for specific visual features as well as tolerance to transformations that maintain the identity of those features. These two aspects of visual coding have been well studied in the context of object-related response properties but have been more difficult to study in mid-level visual areas, such as V2, where we lack a full account of the relevant stimulus dimensions. Our recent work indicates that V2 neurons, but not V1 neurons, are sensitive to particular statistical dependencies found in natural images. Images of natural texture, in particular, are known to exhibit correlations in the output of V1-like filters across orientation, scale, and position. V2 neurons are well driven by stimuli containing these correlations, compared to stimuli that lack them. Such sensitivity suggests a unique form of tolerance; perhaps neurons in V2 maintain their selectivity to the presence of correlations even while other features of the image are randomized. We thus studied the tolerance of V1 and V2 neurons to families of statistically-matched stimuli at both the single neuron and population level. V2 neurons responded more tolerantly than V1 neurons, and this tolerance depended on different receptive field mechanisms in the two areas. By decoding neural population responses, we found that V2 was better than V1 at discriminating families matched for different statistics and worse at discriminating images matched for the same statistics, a pattern of performance that resembles human perceptual experience.
Jonathan Winawer
Spatial integration in the human visual pathways
A major goal in neuroscience is to measure and quantitatively model how the human brain responds to images, and in turn how these neural responses relate to perception. Here, I will describe a series of studies investigating how the human visual system responds to stimuli that vary in spatial position and spatial pattern, using twos kinds of instruments, functional MRI and intracranial EEG, or electrocorticography (‘ECoG’). The first set of studies examines spatial tuning and spatial summation, characterizing how the first several visual field maps summate contrast patterns across the visual field. The fMRI studies show that a general feature of visual cortex is that spatial summation is sub-additive, meaning the response to a large pattern is less than the response to smaller patterns presented separately. These results indicate a failure of linearity. We show that this non-linearity is increasingly important in later stages of the visual hierarchy, and parallels the way in which visual areas become increasingly invariant to changes in stimulus size and position. The ECoG studies show that the same non-linearity is observed in one portion of the ECoG signal, an asynchronous, spectrally broadband signal. A different portion of the signal, the evoked potential, shows spatial summation that is closer to linear, and hence differs from the fMRI signal. Together, the studies help to clarify the circuits and computations involved in spatial summation. A second set of experiments characterizes how the visual system responds to changes in spatial pattern. These studies show that certain spatial patterns give rise to a particular temporal pattern in the ECoG signal, an oscillatory response in the gamma band (peaked around 50 Hz). Gamma oscillations have been previously hypothesized to play an important role in perception and attention, but our studies show that many plainly visible patterns do not elicit gamma oscillations. We conclude that they are not necessary for seeing. Last, we tested whether a new model developed to predict fMRI signals in visual cortex to a large variety of spatial patterns could also predict ECoG signals. We show how this model can be used to predict the time-course of the neurophysiological response during the first 200 ms of seeing an image. Together these studies reflect an effort to bridge together multiple types of measurements of the living human brain, and help us understand the circuits, signals, and computations in the human visual system.
Anna Roe High spatial resolution brain imaging and targeted stimulation in nonhuman primates
One of the greatest challenges in modern science is to understand how the brain produces cognition and behavior. My research focuses on the elemental units of function in the cerebral cortex of nonhuman primates. Using neuroimaging methods in anesthetized and awake, trained monkeys, we have shown that these elemental 200 um sized functional domains encode object features and are modulated by attention. Towards the goal of modulating behavior via external means, we are also developing methods for selective stimulation via electrical microstimulation, pulsed laser stimulation, and optogenetics. These approaches will lead to new understanding and treatment of neurological and mental disease.
Ed Connor
An Efficient Shape Coding Strategy in Visual Area V4
Efficient coding is widely recognized as a primary goal of image transformation in the visual system. Efficient coding is particularly important for maximizing storage capacity in visual memory. efficient coding in early visual cortex is achieved by abstracting local oriented spatial frequencies and by excitatory/inhibitory surround modulation. Little is known about strategies for efficient coding in intermediate and higher-level visual cortex. I will describe an efficient shape coding strategy in primate area V4, based on convergent results from neural recording and simulated V4 populations trained on natural object contours. Neural responses were strongly biased toward acute convex and concave curvature. Simulated V4 populations showed a similar bias specifically when an efficiency (sparse coding) constraint was imposed. These results suggest that efficient coding is achieved in intermediate visual cortex by emphasizing representation of uncommon but diagnostic regions of acute contour curvature.