Improving CNN Performances by Human Visual Channel Properties

Опубликовано 1 ноября 2016, 22:41

Deep Convolution Neural Networks (DCNN) have achieved great sucesses in many computer vision and pattern recognition tasks. In recent years, a common and effective way of improving DCNN's performance accuracies is to increase and network depth and complexities. However, this approach is unsustainable, because deeper networks with more complecity are not only more ffto train and use, but also becoming saturated in terms of performance improvement. We aim to learn useful cues from object recognition mechanisms of the human visual cortex, and to improve DCNN performance accuracies by enforcing the DCNN models to have certain properties of human visual cortex. These approaches are independent of any CNN models. Comprehensive performance evaluations show remarkable performance improvements of CNN models without increasing their model complexities.

See more on this video at microsoft.com/en-us/research/v...