Computer Vision News - January 2018

(A) Representative reCAPTCHA parses showing top two solutions, their segmentations, and labels by two different Amazon Mechanical Turk workers. (B) Word accuracy rates of RCN and CNN on the control CAPTCHA data set. CNN is brittle and RCN is robust when character spacing is changed. (C) Accuracies for different CAPTCHA styles. (D) Representative BotDetect parses and segmentations (indicated by the different colors). Comparisons against the compositional patch model (CPM) and CNN (Fig above) demonstrate RCN’s advantage. RCN’s recognition performance was 76.6% versus 68.9% for CPM and 54.2% for VGG-fc6. RCN proved robust to various types of clutter introduced for testing, without special training to learn those transformations. CNNs’ generalization performance drops significantly when such out-of-sample test examples are introduced (B in the above figure). A lesion study was conducted in order to isolate the contributions of the forward and backward lateral connections: results, summarized in (C in the above figure), show that these lateral connections significantly contribute to RCN’s performance. Summary RCN’s performance suggests that incorporating insights from neuroscience can lead to highly data-efficient, generalizable and robust machine-learning models. Research Computer Vision News Research 9