4TH International Congress on Technology - Engineering & SCIENCE - Kuala Lumpur - Malaysia (2017-08-05)

Regularization In Deep Structural Networks On Pathology Images
Overfitting is a major concern in deep structural networks, especially when a small dataset is considered as input. This problem usually occurs once a complex architecture along with many parameters to tune is performed. Availability of big datasets is a way to tackle this problem, however in some applications, such as medicine, it is hard to access to a huge amount of labelled data. Hence, to prevent overfitting, an accurate regularization technique is required. In this paper, the impact of utilizing different techniques of regularization, rather than Drop-Out, is investigated on a breast cancer pathology dataset. The experimental results show that L2 norm performs properly on such a small and complex medical benchmark, resulting in a robust classification system.
Amin Khatami, Faramarz Samsami, Abbas Khosravi, Saeid Nahavandi