Skip to main content


Biomimetic computer vision system competes well with the best "engineered" systems


Object recognition is a task performed by the human visual system countless times every day, and is an area of intense research in the field of computer vision, and spills over into areas like robotics, automated navigation, image analysis, high-throughput biological assays, and national security. In work of COB trainee Kevin Jarrett, done jointly with his advisor Yann LeCun and others, it is demonstrated that simple hierarchical architectures loosely modelled after the ventral pathway of the visual cortex can categorize objects with a performance similar with the best “engineered” systems from computer vision. Modern object recognition systems typically have feature extraction stages composed of a filter bank, a nonlinear transformation, and a feature pooling layer. Most either use one stage of feature extraction with hard-wired filters, or two stages where the filters in one or both stages are “learned”. One question that Jarrett asked was how the nonlinearities affected the accuracy of recognition. Among other things, he showed that using non-linearities that include rectification and local contrast normalization is the single most important ingredient for good accuracy on object recognition benchmarks.

Address Goals

Computer vision and its tools are enabling technologies in robotics, automatic navigation, image analysis, high-throughput biological assays, and is central to national security. It is an area where advances might be found by understanding the solutions to object recognition that evolution has engineered in biological systems.