אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

Leveragingthe characteristics of convolutional layers, neural networks are extremelyeffective for pattern recognition tasks. However in some cases,their decisions are based on unintended information leading to high performanceon standard benchmarks but also to a lack of generalization to challengingtesting conditions and unintuitive failures. Recentworkhas termed this “shortcut learning” and addressed its presence in multipledomains. In text recognition, we reveal another such shortcut, whereby recognizersoverly depend on local image statistics. Motivated by this, we suggest anapproach to regulate the reliance on local statisticsthat improves text recognition performance. Ourmethod, termed TextAdaIN, creates local distortions in the feature map whichprevent the network from overfitting to localstatistics. It does so by viewing each feature map as a sequence of elementsand deliberately mismatching fine-grained feature statistics between elementsin a mini-batch. Despite TextAdaIN’s simplicity, extensive experiments show its effectiveness compared to other, morecomplicated methods. TextAdaIN achieves state-of-the-art results on standardhandwritten text recognition benchmarks. Additionally, it generalizes tomultiple architectures and to the domain of scene text recognition. Furthermore, we demonstrate that integrating TextAdaINimproves robustness towards more challenging testing conditions. Short bio: Oren Nurielis an applied computer vision scientist at AWS. He holds an MSc degree inComputer Science from the Tel-Aviv University.