This paper presents an overview of popular retrieval tech- niques based on machine learning for content based multimedia retrieval. Furthermore, we also propose to highlight current gaps and required im- provement in this context. We first introduce common retrieval problems, and the usual models and assumptions made on multimedia data. Thanks to these assumptions, techniques based on machine learning can be used in many application cases. In this scope, we present popular methods for indexing multimedia data, like the ones based on the training of visual dic- tionaries. Then, we present supervised techniques that use labeled data to train and design retrieval components. We show how this last topic could benefit from many improvement from the machine learning community. Finally, this paper presents interesting perspective and new paradigms for multimedia retrieval based on machine learning.