The challenge of trying to use a deep neural network trained on one domain for an application in another domain is quite well known to us, and the challenges are quite intriguing. Most of us have tried using ImageNet pertained models of GoogLeNet / VGGNet / AlexNet / ResNet / DenseNet for our work, but the challenge is can they directly be used for grayscale images, color images represented in space other than RGB or if their appearance is very different from natural images. This talk would provide an insight to adapting a pertained deep neural network easily across domains, using techniques ranging from partial re-training, to exploiting dropouts. Showcase examples would include it being leveraged for digital retinal angiography where domain adaptation has demonstrated superior performance than training the network from scratch on the target domain (ACPR 2015), in digital pathology where appearance is very different than in natural images (ISBI 2017), and for surgical informatics through video analytics where it has been coupled with an LSTM to achieve adaptation (CVPR 2017).
Debdoot Sheet is an Assistant Professor of Electrical Engineering and Principal Investigator of the Kharagpur Learning, Imaging and Visualisation Group at the Indian Institute of Technology Kharagpur since 2014. His research focus is machine learning, visual analytics, medical imaging, augmented and mixed reality, computing system design. He is an Editor of IEEE Pulse and loves speaking out on his research findings.
Know about speaker more here
Below is the video of online session lecture.Written on June 4th, 2017 by