• Print:
  • Share:
Image of eighteen brain slices arranged in three rows

Global Scholar workshops are essential to broadening my view and furthering my career.

Bill Coish
CIFAR Global Scholar Alumni and Assistant Professor, McGill University

Neural Computation & Adaptive Perception Research Progress

A disproportionately large portion of our brains is devoted to visual processing. This suggests that we primarily understand our world through sight, and that studying how we perceive the world visually may be an optimal means for deciphering how the human brain works.

Neural Computation & Adaptive Perception program members are unlocking the mystery of how our brains convert sensory stimuli into information. They are also trying to teach computers to “see.”

Since its inception in 2004 under the leadership of former Program Director Geoff Hinton (2004-2013), the program has made significant progress in three key areas: computational vision (understanding the brain as a computing device), machine learning (teaching computers how to learn), and machine vision (developing methods for automatically interpreting images).

For example, program members are developing “synthetic neural networks,” which are computational vision models that simulate biological neuronic structures and functions. These models can help decipher the vision processes that facilitate recognition of objects, even when they are significantly distorted or observed from a new viewpoint. This work can be applied to tools ranging from handwriting recognition software to automatic translation programs.

The natural progression of this research is to improve machine learning by creating a computational vision model that adapts in response to its environment. Natural biological networks vary their synaptic connectivity in response to the sensory environment. To mimic this dynamism in computational neural networks, program members developed ways to communicate information about both visual features and connectivity simultaneously.

Program members have also developed ways to model attention mathematically – allowing synthetic neural networks to focus on important information, and pay less attention to other sensory input. Other researchers developed an algorithm that allows a computer to select features distinctive to a particular sought object.

Pattern and detail recognition are widely applicable computing capabilities, and the research of this program has contributed to better data compression algorithms, faster and better search engines, object recognition smartphone apps and many other transformative technologies. At least as important, though, are the insights these researchers provide into the nature of human cognition itself.

  • Share:
  • Print: