In this talk, I address two issues in content-based retrieval using deep learning: accuracy and efficiency. Deep convolutional neural networks (CNNs) have shown impressive performance on image classiﬁcation. However, when the feature representations learned by deep CNNs are applied to image retrieval, the performance is still not as good as they are used for classiﬁcation. We introduce an inter-batch reference learning approach based on the mean-average-precision criterion to handle this task. The learned feature space is better for relevant image search and can improve the retrieval accuracy. We also develop a method that learns binary hash codes from deep networks to speed up the search. Finally, I introduce some recent topics in my laboratory, including generative models and deep embedded vision.