We studied extra-receptive field contextual modulation in area V1 of awake, behaving macaque monkeys. Contextual modulation was studied using texture displays in which texture covering the receptive field (RF) was the same in all trials, but the perceptual context of this texture could vary depending on the configuration of extra-RF texture elements. We found robust contextual modulation when disparity, color, luminance, and orientation cues variously defined a textured figure centered on the RF of V1 neurons. We found contextual modulation to have a spatial extent of approximately 8 to 10 degrees diameter parafoveally. Contextual modulation correlated with perceptual experience of both binocularly rivalrous texture displays and of displays with a simple example of surface occlusion. We found contextual modulation in V1 to have a characteristic latency of 80-100 msec after stimulus onset, potentially allowing feedback from extrastriate areas to underlie to this effect.

By means of their small receptive fields (RFs), neurons in primary visual cortex perform highly localized analyses of the visual scene, far removed from our normal unified experience of vision. Local image elements coded by the RF are put into more global context, however, by means of modulation of the responses of the V1 neurons. Contextual modulation has been shown to follow closely the perceptual interpretation of the scene as a whole. This would suggest that some aspects of contextual modulation can be recorded only in awake and perceiving animals. In this study, multi-unit activity was recorded with implanted electrodes from primary visual cortex of awake, fixating monkeys viewing textured displays in which figure and ground regions were segregated by differences in either orientation or motion. Contextual modulation was isolated from local RF processing, by keeping RF stimulation identical across trials while sampling responses for various positions of the RF relative to figure and ground. Contextual modulation was observed to unfold spatially and temporally in a way that closely resembles the figure-ground percept. When recording was repeated, but with the animals anesthetized, the figure-ground related modulatory activity was selectively suppressed. RF tuning properties, however, remained unaffected. The results show that the modulatory activity is functionally distinct from the RF properties. V1 thus hosts distinct regimes of activity that are mediated by separate mechanisms and that depend differentially on the animal being awake or anesthetized.The receptive fields (RFs) of neurons in primary visual cortex (V1) are specialized for conveying information of a low-level nature, such as the orientation and spatial frequency of luminance contrast. Moreover, the small dimensions of each neuron's RF restrict the given cell's analysis to a very limited portion of the visual field (1-4). The elementary analysis performed by individual V1 neurons' RFs is indeed so far removed from our sense of vision that some have proposed activity within primary visual cortex to be strictly preperceptual (5-8). There is mounting evidence, however, that V1 neurons signal, through mechanisms apparently distinct from those shaping the RF, information about the perceptual interpretation of larger parts of the scene than covered by the RF. These mechanisms, expressed as a delayed modulation of stimulus-driven responses, appear to rely not on the particular image features that fall within a cell's RF, but rather on the larger perceptual context of these features (9-18). For example, this kind of modulation may closely reflect perceptual illusions (12) or the percept of figure-ground segregation (16,18).For a wide range of display configurations, we previously have demonstrated that contextual modulation in V1 follows the perceived structure of figure-ground displays (16, 18). These results all were obtained in awake, trained animals. In V1, neural activity also can easily be recorded in anesthetized animals. RF tuning prope...

Abstract& In a backward masking paradigm, a target stimulus is rapidly (<100 msec) followed by a second stimulus. This typically results in a dramatic decrease in the visibility of the target stimulus. It has been shown that masking reduces responses in V1. It is not known, however, which process in V1 is affected by the mask. In the past, we have shown that in V1, modulations of neural activity that are specifically related to figure -ground segregation can be recorded. Here, we recorded from awake macaque monkeys, engaged in a task where they had to detect figures from background in a pattern backward masking paradigm. We show that the V1 figureground signals are selectively and fully suppressed at targetmask intervals that psychophysically result in the target being invisible. Initial response transients, signalling the features that make up the scene, are not affected. As figure -ground modulations depend on feedback from extrastriate areas, these results suggest that masking selectively interrupts the recurrent interactions between V1 and higher visual areas. &

No abstract

Robust driver attention prediction for critical situations is a challenging computer vision problem, yet essential for autonomous driving. Because critical driving moments are so rare, collecting enough data for these situations is difficult with the conventional in-car data collection protocol-tracking eye movements during driving. Here, we first propose a new in-lab driver attention collection protocol and introduce a new driver attention dataset, Berkeley DeepDrive Attention (BDD-A) dataset, which is built upon braking event videos selected from a large-scale, crowd-sourced driving video dataset. We further propose Human Weighted Sampling (HWS) method, which uses human gaze behavior to identify crucial frames of a driving dataset and weights them heavily during model training. With our dataset and HWS, we built a driver attention prediction model that outperforms the state-of-the-art and demonstrates sophisticated behaviors, like attending to crossing pedestrians but not giving false alarms to pedestrians safely walking on the sidewalk. Its prediction results are nearly indistinguishable from ground-truth to humans. Although only being trained with our in-lab attention data, the model also predicts in-car driver attention data of routine driving with state-of-the-art accuracy. This result not only demonstrates the performance of our model but also proves the validity and usefulness of our dataset and data collection protocol.

No abstract

Carbohydrate recognition plays an important role in the development of normal projections of sensory afferent neurons in the leech CNS. Four different carbohydrate epitopes are expressed by sensory afferents on their 130 kDa surface proteins: all sensory afferents share a common carbohydrate epitope (CE0) that helps them to enter and project diffusely across the synaptic neuropil; a restricted expression of three other carbohydrate epitopes (CE1, CE2, and CE3) serves to distinguish three subsets of sensory afferents. We examined the subsets of sensory afferents defined by their subset carbohydrate epitopes in the leech lip, skin, gut, and CNS. We established that the CE1, CE2, and CE3 subset epitopes define disjoint subsets of neurons by double labeling sensory afferents with monoclonal antibodies for different pairs of subset epitopes. We found that CE2 and CE3 afferents populate the lip and skin, but not the gut, and that these two subsets of sensory afferents have convergent projection patterns in the CNS. We found that CE1 afferents populate the gut and skin, but not lips; furthermore, their CNS projections diverge from those of CE2 and CE3 afferents. Our data fit the hypothesis that these carbohydrate epitopes are related to sensory modality of afferent subsets.

Autonomous driving requires operation in different behavioral modes ranging from lane following and intersection crossing to turning and stopping. However, most existing deep learning approaches to autonomous driving do not consider the behavioral mode in the training strategy. This paper describes a technique for learning multiple distinct behavioral modes in a single deep neural network through the use of multi-modal multi-task learning. We study the effectiveness of this approach, denoted MultiNet, using selfdriving model cars for driving in unstructured environments such as sidewalks and unpaved roads. Using labeled data from over one hundred hours of driving our fleet of 1/10th scale model cars, we trained different neural networks to predict the steering angle and driving speed of the vehicle in different behavioral modes. We show that in each case, MultiNet networks outperform networks trained on individual modes while using a fraction of the total number of parameters.