Coast to Coast Seminar Series 2015: "Visual Perception: The Ultimate Big Data Problem"

Abstract

The human retina contains over 60M photoreceptors. If we ballpark temporal sampling at 60Hz this implies a data rate of about 3.6GBps. At 25 years of age you have already processed on the order of 700 petabytes (6e18 bits) of data. This is big!

How do human and machine visual systems extract value from this onslaught of information? One common factor is learning: Human perceptual development is strongly data-driven and requires extensive experience; in fact it is now known that some visual function does not fully develop until the late teens. In the last few decades we have witnessed a revolution in machine vision research through a congruent integration of machine learning methods that take advantage of big data.

Despite this commonality, there are profound differences between the human visual system and machine vision systems. Biological systems are inherently general-purpose in nature, called on to solve a diversity of problems (scene layout, object recognition, navigation…), often concurrently. This encourages the development of efficient, generative, general-purpose models that serve multiple functions. Machine vision algorithms, on the other hand, tend to target narrowly defined problems associated with specialized datasets, leading to specific models that can limit representations to capture only the most relevant features.

Despite this divergence, there are also opportunities for convergence, particularly as we try to make machine vision systems more general in nature. I will discuss research problems in perceptual organization, shape perception, linear perspective and spatial attention which can contribute to the development of general-purpose visual systems, and where research in biological and computer vision has already been synergistic.