Deep learning techniques are having an undeniable impact on general pattern recognition issues. In this paper, from a developmental robotics perspective, we scrutinize deep learning techniques under the light of their capability to construct a hierarchy of meaningful multimodal representations from the raw sensors of robots. These investigations reveal the differences between the methodological constraints of pattern recognition and those of developmental robotics. In particular, we outline the necessity to rely on unsupervised rather than supervised learning methods and we highlight the need for progress towards the implementation of hierarchical predictive processing capabilities. Based on these new tools, we outline the emergence of a new domain that we call deep developmental learning.