A primary goal in cognitive neuroscience is to identify neural correlates of conscious perception (NCC). By contrasting conditions in which subjects are aware versus unaware of identical visual stimuli, a number of candidate NCCs have emerged, among them induced gamma band activity in the EEG and the P3 event-related potential. In most previous studies, however, the critical stimuli were always directly relevant to the subjects’ task, such that aware versus unaware contrasts may well have included differences in post-perceptual processing in addition to differences in conscious perception per se. Here, in a series of EEG experiments, visual awareness and task relevance were manipulated independently. Induced gamma activity and the P3 were absent for task-irrelevant stimuli regardless of whether subjects were aware of such stimuli. For task-relevant stimuli, gamma and the P3 were robust and dissociable, indicating that each reflects distinct post-perceptual processes necessary for carrying-out the task but not for consciously perceiving the stimuli. Overall, this pattern of results challenges a number of previous proposals linking gamma band activity and the P3 to conscious perception.

Abstract-The real world is composed of sets of objects that move and morph in both space and time. Useful concepts can be defined in terms of the complex interactions between the multi-dimensional attributes of subsets of these objects and of the relationships that exist between them. In this paper, we present Spatiotemporal Multi-dimensional Relational Framework (SMRF) Trees, a new data mining technique that extends the successful Spatiotemporal Relational Probability Tree models. From a set of labeled, multi-object examples of a target concept, our algorithm infers both the set of objects that participate in the concept and the key object and relation attributes that describe the concept. In contrast to other relational model approaches, SMRF trees do not rely on pre-defined relations between objects. Instead, our algorithm infers the relations from the continuous attributes. In addition, our approach explicitly acknowledges the multi-dimensional nature of attributes such as position, orientation and color. Our method performs well in exploratory experiments, demonstrating its viability as a relational learning approach.Keywords-relational learning; continuous multi-dimensional attributes; multiple instance learning; spatial representations I. MOTIVATION AND BACKGROUNDThe world is composed of collections of objects, each with a set of associated attributes. Whether it is a robot preparing to perform the next step in a cooking sequence or an agent generating warnings of severe weather, only a specific subset of the observable objects is relevant to making decisions about what steps to take next. In particular, the relevance of an object is determined by its attributes and the relations that it has with other objects. These attributes are often continuous and multi-dimensional, such as Cartesian positions or colors in a red-green-blue (RGB) space. Given a set of training examples, our challenge is to discover the objects that play the crucial roles in the examples as well as the description of the key object attributes and relations.Our work is inspired by the successful Relational Probability Tree (RPT) [1] and the Spatiotemporal Relational Probability Tree (SRPT) [2] models. Both approaches create probability estimation trees, a form of a decision tree with probabilities at the leaves. Splits in the decision trees can ask questions about the observed properties of the objects or their relationships. Given a novel graph, these decision trees estimate the probability that the graph contains a set of objects that corresponds to some target concept. Like Kubica et al. [3], [4], these approaches build models using pre-specified categorical relations.The Spatiotemporal Multidimensional Relational Framework (SMRF) extends this prior work in two key ways. The first extension is the ability to ask questions based on continuous, multi-dimensional attributes. For example, the color of a pixel can be represented as a RGB tuple. Capturing a concept such as "yellow" requires that the blue variable be low but the green and r...