Précis of Active Perception by Nicholas

Machine perception technologies have the potential to revolutionize the fields of health, individual and public safety, marketing, education, and art. But for all their potential, existing machine perception technologies are insufficient. The aim of Active Perception is to contribute to the scientific understanding of perceptual intelligence in humans, and its synthesis in machines. Attempts to synthesize computational intelligence have typically succeeded in inverse proportion to the perceived level of intelligence. For example, winning complex games is difficult for humans but relatively easy for computers; meanwhile, " seeing " is effortless for humans but is astoundingly difficult for computers. This disparity highlights a gap in our understanding of the fundamental nature of human intelligence. Perception is an active process at many levels: neurons grow synapses and modulate their firing properties; attention sweeps its spotlight from place to place; eyes move; infants explore the effects that they cause; scientists design experiments to understand the world; groups of individuals gather and coordinate information to form a shared understanding. Although there are unique peculiarities at each of these levels, there are also basic commonalities. Active Perception contains six related research studies that span multiple levels of cognitive science. This breadth is possible because a common computational framework focuses on the invariants of all levels. Study (1) considers a population of active neurons that modify internal and synaptic parameters to maximize their capacity as information channels. We show that these model neurons exhibit firing rate patterns and receptive field patterns that have been previously reported in empirical studies, and solve non-linear independent components problems. The neural adaptation mechanism is simple and local in time and space, which are necessary for a plausible biological mechanism. This work suggests new algorithms for unsupervised visual feature extraction, a basic building block of visual object recognition systems. Study (2) explores an information theoretic account for human visual attention, and proposes an efficient computational implementation of attentional systems. This enables deployment of a real-time algorithms. We tested these in a robot, and found that the robot's attention was directed to regions of its surroundings that contained people even though it had no specific knowledge of the appearance of humans. Our system shares many algorithmic similarities to the popular SIFT key point detection algorithm, while operating at a rate that is about eighteen times faster. Thus it has the potential to significantly increase the speed and reduce the power requirements of …