Abstract: Three sources of evidence have been cited for face-specific processing in human object recognition. First, there are neuropsychological case studies where brain-injured subjects appear to be disproportionately impaired at face recognition as compared to the recognition of common non-face objects (prosopagnosia). Second, functional brain imaging studies (fMRI and PET) appear to show a dedicated neural substrate -- a portion of inferior temporal lobe (IT; part of the fusiform and inferior temporal gyri) -- in humans that is more active when viewing faces as compared to when viewing common non-face objects. Third, psychophysical studies have revealed a range of putatively face-specific behavioral effects -- in particular, extreme sensitivity to the configural aspects of stimuli across a variety of recognition tasks.
It is our contention, however, that face recognition should be considered a particular case of within-category/item-specific recognition, in which members of a visually homogeneous class must be distinguished by experts. Using this framework we tested claims regarding face-exclusive processing for all three sources of evidence. First, across multiple experiments we found that prosopagnosic subjects are disproportionally sensitive to increasingly specific levels of categorization with both familiar and novel non-face objects (rather than faces _per se_). Second, in an fMRI study we found that the same region of IT associated with faces is engaged by the subordinate-level recognition of familiar non-face objects. Third, in behavioral studies we found that experts, but not novices, with a novel class of non-face objects ("Greebles") showed configural sensitivity similar to that obtained with faces. Finally, we combined the latter two techniques and found individual subjects' "face areas" were active during Greeble processing, but only following expertise training. Together, these results suggest that the level of visual categorization and the degree of expertise, not just stimulus class membership, may be important mediating factors in dissociations generally found between face and object recognition in the human ventral pathway. At a more general level, these results suggest that models of visual recognition cannot simply assume a restricted domain of explanation -- the human visual recognition system is highly flexible and theories must be able to account for a wide range of recognition behaviors. At a minimum, our results provide additional controls for proposed of face-specific mechanisms and neural substrates.
Abstract: Partial differential equations have become a very popular vehicle for the implementation of anisotropic diffusion on greyscale images. A number of different diffusion techniques have been developed which tend to smooth away noise in an image, while maintaining the fidelity of edges and other salient image features. One category of such filters is based on gradient descent along energy functionals, many of which cannot be formulated in a straightforward manner for vector-valued data. Another category of filters are derived from geometry driven diffusions which require the data to be represented as a hypersurface. Here again, there is a problem in applying these methods to vector-valued imagery since a greyscale image is naturally represented as a hypersurface while a vector-valued image is not.
In this talk, we begin by deriving a well-posed partial differential equation for diffusing two dimensional greyscale images. The equation will implement a projected form of mean curvature motion on the underlying image that tends to remove noise will maintaining sharp edges. We will then show how this filter can be extended into an entire class of vector-valued filters in a very straight-forward manner. The performance of the filter will be shown on a variety of color and greyscale images.
Abstract: Many features of surfaces and curves which are important in applications can be described by means of the contact between the curve or surface and some elementary geometrical object like a line, plane, circle or sphere. Much is now known about the behaviour of these contacts, for example under deformations of the curve or surface. I shall talk about some of these, bringing in the ideas of parabolic curve, ridge, and other things.
Abstract: As processor architectures have increased their reliance on speculative execution to improve performance, the importance of accurate prediction of what to execute speculatively has increased. Furthermore, the types of values predicted have expanded from the ubiquitous branch and call/return targets to the prediction of indirect jump targets, cache ways and data values. In general, the prediction process is one of identifying the current state of the system, and making a prediction for some as yet uncomputed value based on that state. Prediction accuracy is improved by learning what is a good prediction for that state using a feedback process at the time the predicted value is actually computed. While there have been a number of efforts to formally characterize this process, we have taken the approach of providing a simple algebraic-style notation that allows one to express this state identification and feedback process. This notation allows one to describe a wide variety of predictors in a uniform way. It also facilitates the use of an efficient search technique called genetic programming, which is loosely modeled on the natural evolutionary process, to explore the design space. In this paper we describe our notation and the results of the application of genetic programming to the design of branch and indirect jump predictors.
Dr. Joel S. Emer is a Senior Consulting Engineer in Digital's Semiconductor Engineering Group. He holds a Ph.D. in Electrical Engineering from the University of Illinois, and M.S.E.E. and B.S.E.E. degrees from Purdue University. He is a 18 year Digital employee, where he has worked on processor performance analysis and performance modeling methodologies for a number of VAX and Alpha CPUs, as well as researched heterogeneous distributed systems and networked file systems. His current research interests include multithreaded processor organizations, techniques for increased instruction level paralellism, instruction and data cache organizations, branch prediction schemes and data prefetch strategies for future Alpha processors.
Abstract: The currently available techniques to analyze data to detect if deterministic chaos is present in a time series require relatively long and noise free data records. For example, most methods become unreliable in distinguishing the presence of deterministic chaos when the noise variance becomes greater than 2% of the signal variance. The consequence of this is that experimental noise can cause a false representation of the presence of deterministic chaos when in fact the system is not chaotic. The algorithm we present in this paper reduces the requirement of relatively long and noise free data records. Our algorithm achieves this accurate detection of deterministic chaos via the use of the stochastic nonlinear autoregressive model followed by the estimation of Lyapunov exponents. Using the well known nonlinear deterministic chaotic continuous (Rossler and Lorenz) and discrete (Hennon and logistic) systems, we show that our algorithm remains accurate for both types of chaotic systems for small number of data as well as when the variance of noise approximately 2.5 times greater than the variance of the signal.
Abstract: The presentation is focused on the idea that modern imaging opto-electronic systems (optics + sensors) and image processing algorithms should be considered as an integrated system in which the imaging system structure is dependent on the processing algorithm, and the algorithm itself is designed taking into account projected imaging hardware advances. The following topics will be considered:
1. Active optics and image post-processing
2. Adaptive imaging
2.1 Image quality metrics for adaptive wavefront correction
2.2 Control algorithms
2.3 Adaptive system implementation
3. Synthetic imaging
3.1 Non-adaptive anisoplanatic image correction
3.2 Depth-estimation with a single camera
3.3 System integration
4. Nonlinear optical image processing
4.1 "On-the-fly" edge detection
4.2 Moving object tracking
4.3 Self-organized image mapping
Back to the LEMS homepage.