Cosyne 2005 Workshops
March 21-22, 2005
Snowbird, Utah
Workshop Title
Invariant Representations in Vision
Organizer(s)
Bruno Olshausen and Dileep George
Abstract
Finding out how our brains make sense of the visual world is one of the goals of computational and systems neuroscience research. There is considerable psychophysical evidence that the brain uses invariant representations. For example, when you move towards an object while still looking at it, the images that fall on your retina vary from one instant to another whereas your perception of that object remains stable. It is hypothesized that the brain makes sense of this highly variable world by forming invariant representations of structure in the environment. In addition, there is physiological evidence for neurons that maintain object/shape selectivity independent of large changes in position and size of an object.
The precise nature of invariant representaions and the computational mechanisms by which they are learned remains a mystery. While some researchers show evidence for near-perfect invariant recognition of objects, some other psychophysics researchers show that the learning is far from perfect for novel stimuli. Computational modeling attempts of the invariance property range from modeling of geometric transformations to modeling of invariant feature detectors. While earlier researchers thought of invariance (and hence vision) as a purely spatial phenomenon, many recent researchers have started incorporating time into their models. Anatomists and physiologists provide vital clues in to the brain's solution to the invariance problem and their interaction with theorists could result in key experiments that tease out various aspects of the brain mechanisms involved in solving the Invariance Problem.
This workshop will bring together researchers from the areas of computational modeling (learning), psychophysic and physiology to explore the problem of invariant representations in vision from these different perspectives. We list here some of the themes we wish to address:
1: Psychophysics: How invariant are the invariant representations? It is important to recognize that invariance is a brain-defined problem while attempting to model it mathematically. Psychophysics research reveals the properties of brains invariant recognition mechanisms. Some pertinent questions in this area include: Do invariances learned at one position automatically transfer to other? How good are we at recognizing flashed images? Are eye movements necessary? Do they help? What deficits do patients with damaged dorsal/ventral streams exhibit?
2: The role of time in learning invariant representations. When we move about in this world, the scene around us change in a continuos fashion. This continuity in time could provide a vital cue to the brain for learning invariances. Can invariances be learned without the use of time? How does the brain use the temporal information?
3: What happens to the variant part (the "where" information)? In many cases, the formation of invariant representation involves losing information about the variant part - for example the location of an object in the scene. However, knowledge about the location of an object is important for navigation. What is the nature of information lost in forming invariant representations? What happens to where information? Is it represented elsewhere? What are the roles of ventral and dorsal streams?
4: Physiology. Knowing the physiological responses of neurons in the visual cortex would provide vital clues regarding the nature of the representaitons.Nature of the neurons and their receptive fields in the two streams. Do dorsal and ventral streams interact?Where/How is "where" information represented?
5: Sensori-motor Contingencies. It could be that the representations are intimately tied to the motor interactions we have with the world.What are the differente ways in which motor interactions manifest in invariant representations?
