INESC TEC researchers win prizes for their work on convolutional neural networks
Isabel Rio-Torto and Ricardo Cruz, researchers at INESC TEC’s Centre for Telecommunications and Multimedia (CTM), received three prizes awarded by Fraunhofer Portugal and the Portuguese Association for Pattern Recognition (APRP) – the latter, within the scope of 26th edition of RECPAD – Portuguese Conference on Pattern Recognition.
23rd November 2020
Isabel Rio-Torto won two awards in the category of Best Master’s Thesis, thanks to her work “Producing Decisions and Explanations: A Joint Approach Towards Explainable CNNs”, developed within the scope of the master’s degree in Electrical and Computers Engineering, at the Faculty of Engineering of the University of Porto (UP-FEUP). The researcher aimed to develop a convolutional neuronal network (CNN) for medical image analysis, namely concerning cervical cancer, capable of distinguishing between different cases and presenting a visual explanation. This way, the system can be applied in real clinical settings, helping doctors in the diagnosis process.
According to the researcher, “the innovative element of the work was the production of visual explanations, carried out with the decisions of the neural network, unlike the other current methods, which only produce the decisions and use another algorithm to generate explanations”.
As said by Isabel Rio-Torto, “it was a great privilege to witness my work getting acknowledged, and all my dedication and effort awarded, thus demonstrating the importance and potential of Explainable Deep Learning in clinical context. It is something that encourages me even more to continue researching, now that I have started my career as a PhD student and researcher at INESC TEC”.
Ricardo Cruz won the Best Oral Presentation at RECPAD, and his work “Training Convolutional Neural Networks to be Background Invariant” was acknowledged among the 53 poster presentations.
The paper focused on a convolutional neural network for the recognition of objects, and some issues related to its performance – a neural network has problems recognising objects when the background is different from the usual. In this sense, Ricardo Cruz aimed to explore a way to train CNN, so that it remains invariant to the background of the images, by using adversarial training between two different CNNs, thus contributing to greater resilience of the original CNN.
According to the researcher, “this is the first work that we know of whose objective is to make a CNN more robust, depending on the background of the images. Although there is already a lot of work that resorts to adversarial dynamics to generate realistic photographs or other media content, there is less work that demonstrates how this type of dynamics can be explored to improve the recognition models themselves”.
With regard to the award, Ricardo Cruz said that “since I am about to present the PhD thesis, after four years of hard work, it was rewarding to see one of the products of said work being acknowledged this way”.
RECPAD aims to promote the collaboration between the Portuguese scientific community in areas such as Pattern Recognition, Image Analysis and Processing, Computing, etc.
With regard to Fraunhofer Portugal, its major goal is the production of knowledge with potential impact for customers and partners, exploring technological innovations towards economic growth and the improvement of social conditions and the quality of life of end users.
The INESC TEC researchers mentioned in this news piece are associated with UP-FEUP and FCT.