IISc Team Is Training Computers to “See” Objects
Associate professor S P Arun’s laboratory is spearheading the effort to study how computers can be trained to “see” objects, in what is an attempt to replicate the human brain’s visual processing and recognition abilities at the IISc Centre for Neuroscience. It is based on ongoing experiments to better understand the rules by which the brain transforms what humans see into what they perceive.
By gaining deeper understanding on how a human brain recognizes objects, scientists want to know how computers can be trained to “see” better. “I believe to make true progress in understanding how to solve vision, we need to understand the physics and mathematics of vision,” Arun said in a statement issued by Gubbi Labs, a research collective.
Arun is the recipient of the intermediate fellowship awarded by a £160 million initiative funded equally by The Wellcome Trust, UK and Department of Biotechnology, India.
Set up in 2010, Arun’s laboratory tests and validates their theories of vision using electrical recordings from brain cells in monkeys and behavioural experiments with human subjects. Experiments with humans typically involve asking people to pick out specific objects in a group of similar shapes
and measuring the time taken to complete the task. These experiments provide information on how the human brain represents objects perceptually.The Centre for Neuroscience is supporting the Centre for Brain Research, a facility set up through a generous donation of Rs 225 crore by Pratiksha Trust, a charitable trust started by Infosys cofounder Kris Gopalakrishnan and his wife Sudha.
Arun’s laboratory recently found a mechanism by which humans correctly judge the size of objects they see. In the study, Arun and his students found that specific brain cells code for the relative size of objects, enabling the brain to correctly judge the size of an object by comparing it with a nearby object.
Another recent finding was that the human visual system uses a simple combination of features to describe an object. While the features themselves are yet unknown, the group found that they combine in a very specific and simple manner. “These are, I think, some fairly elementary computations being performed by the brain and it’s strange no one has found them so far until we looked,” Arun said.
Notwithstanding the implications for designing better computer vision algorithms, the group’s recent findings on human vision have led the scientists to believe that the brain accumulates systematic knowledge about the external world. Arun’s team hopes it can, in due course, uncover how the brain incorporates and extracts this knowledge.
