Deep learning (DL) is one of the most emerging types of contemporary machine learning techniques that mimic the cognitive patterns of animal visual cortex to learn the new abstract features automatically by deep and hierarchical layers. DL is believed to be a suitable tool so far for extracting insights from very huge volume of so-called big data. Nevertheless, one of the three “V” or big data is velocity that implies the learning has to be incremental as data are accumulating up rapidly. DL must be fast and accurate. By the technical design of DL, it is extended from feed-forward artificial neural network with many multi-hidden layers of neurons called deep neural network (DNN). In the training process of DNN, it has certain inefficiency due to very long training time required. Obtaining the most accurate DNN within a reasonable run-time is a challenge, given there are potentially many parameters in the DNN model configuration and high dimensionality of the feature space in the training dataset. Meta-heuristic has a history of optimizing machine learning models successfully. How well meta-heuristic could be used to optimize DL in the context of big data analytics is a thematic topic which we pondered on in this paper. As a position paper, we review the recent advances of applying meta-heuristics on DL, discuss about their pros and cons and point out some feasible research directions for bridging the gaps between meta-heuristics and DL.