Automatic screening of Age-related Macular Degeneration (AMD) is important for both patients and ophthalmologists. In this paper, we focus on the task of AMD detection at the very early stage from fundus images. The difficulty of this task is that at the very early stage, the signs, e.g., drusen, are too tiny and subtle to be detected by most of the current methods. To address this issue, we apply deep learning in a multiple instance learning framework to catch these subtle features to detect AMD at the very early stage. The deep networks is able to learn a discriminative representation of the subtle signs of AMD. The multiple instance learning framework helps in two ways. First, It is able to choose the location where AMD happens because it works on image patches instead of the whole image. Second, It works on the image of high resolution instead of down sampling the image which may lead to invisibility of the tiny drusen. The experiments are carried out on a dataset consists of 3596 AMD and 1129 normal fundus images. The final average AUC is 0.79, compared with 0.74 of the same neural network but without multiple instance learning.