Dipartimento di Scienze di Base e Applicate per l'Ingegneria

Abstract:

Convolutional Neural Networks have proved valuable tools in a wide range of domains since their introduction. In particular, due to their ability to extract high level features from raw data, they demonstrated excellent capabilities in image classification, image segmentation and object detection. The notable technological improvements and the availability of well established benchmark datasets such as ImageNet, MNIST or CIFAR10 paved the way to the development of deep CNN architectures that are characterized by multiple layers and perform computationally expensive operations, which require the possibility to access to powerful devices (e.g., servers endowed with GPUs). This poses questions about their applicability in real-world scenarios, where limited resources or time pressures might represent key constraints. In this seminar, I will discuss my personal experiences during my collaboration with the Department of Basic and Applied Sciences to Engineering (SBAI). Specifically, I will present our previous and ongoing work on image classification with CNNs and I will illustrate the main findings of our research and some possible research directions.