Prof. Tistarelli from the University of Sassary will give an invited talk on Human Face Recognition: Learning from Biological Deep Networks.
Abstract of the talk:
Face Recognition has been extensively studied as a mean to facilitate man-machine interaction in a variety of different applications. Due to the imaging variabilities and to the complex nature of the face shape and dynamics, analyzing and recognizing human faces from digital images is still a very complex task.
In the last decade deep learning techniques have strongly influenced many aspects of computational vision. Many difficult vision tasks can now be performed by deploying a properly tailored and trained deep network. Oxford University’s VGG-face is possibly the first deep convolutional network designed to perform face recognition, obtaining unsurpassed performance at the time it was firstly proposed. The enthusiasm for deep learning is unfortunately paired by the present lack of a clear understanding of how they work and why they provide such brilliant performance. The same applies to Face Recognition.
Over the last years, several and more complex deep convolutional networks, trained on very large, mainly private, datasets, have been proposed still elevating the performance bar also on quite challenging public databases, such as the Janus IJB-A and IJB-B. Despite of the progress in the development of such networks, and the advance in the learning algorithms, the insight on these networks is still very limited. For this reason, in this talk we analyse the neural architecture of the early stages of the human visual system to devise a biologically-inspired model for face recognition. The aim is not pushing the recognition performance further, but to better understand the representation space produced from a deep network and how it may help explaining the process undergoing a real biological neural architecture.
In this talk we analyse an hybrid model network trying to better understand the role of the different layers, including the retino-cortical mapping simulated by a log-polar image resampling. The following issues will be addressed:
- What is the representation space within a deep convolutional network and how this reflects the organization of the human visual cortex.
- How the retino-cortical mapping, implemented in the human visual system, may impact the representation space, hence improving the classification performance.
- The relevance of peripheral vs foveal vision for face recognition.