Deep Learning has already proven to be the primary technique to address a number of problems. It holds further promise in solving more challenging problems if we can overcome obstacles, such as the lack of quality training data and poor interpretability. The exploitation of domain knowledge and application semantics can enhance existing deep learning methods by infusing relevant conceptual information into a statistical, data-driven computational approach. This will require resolving the impedance mismatch due to different representational forms and abstractions between symbolic and statistical AI techniques. In this article, we describe a continuum that comprises of three stages for infusion of knowledge into the machine/deep learning architectures. As this continuum progresses across these three stages, it starts with shallow infusion in the form of embeddings, and attention and knowledge-based constraints improve with a semideep infusion. Toward the end reflecting deeper incorporation of knowledge, we articulate the value of incorporating knowledge at different levels of abstractions in the latent layers of neural networks. While shallow infusion is well studied and semideep infusion is in progress, we consider Deep Infusion of Knowledge as a new paradigm that will significantly advance the capabilities and promises of deep learning.