Modeling Human Development: Effects of Blurred Vision on Category Learning in CNNs

William Charles, Daniel Leeds

Keywords: CNNs, computer vision

Abstract: Recently, training convolutional neural networks (CNNs) using blurry images has been identified as a potential means to produce more robust models for facial recognition (Vogelsang et al. 2018). This method of training is intended to mimic biological visual development, as human visual acuity develops from near-blindness to normal acuity in the first three to four months of life (Kugelberg 1992). Object recognition develops in tandem during this time, and this developmental period has been shown to be critical for many visual tasks in later childhood and adulthood. We explore the effects of training CNNs on images with different levels of applied blur, including training regimens with progressively less blurry training sets. Using subsets of ImageNet (Russakovsky 2015), CNN performance is evaluated for both broad object recognition and fine-grained classification tasks. Results for AlexNet (Krizhevsky et al. 2012) and the more compact SqueezeNet (Iandola et al. 2016) are compared. Using blurry images for training on their own or as part of a training sequence increases classification accuracy across collections of images with different resolutions. At the same time, blurry training data causes little change to training convergence time and false positive classification certainty. Our findings support the utility of learning from sequences of blurry images for more robust image recognition.

One-sentence Summary: Mimicking human development by using blurry images to train CNNs for object recognition can make them more robust to image degradation.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): https://openreview.net/references/pdf?id=gQWN-jpl2