Physically Unclonable Functions (PUFs) have been addressed nowadays as a potential solution to improve the security in authentication and encryption process of Cyber Physical Systems. The research on PUF is actively growing due to its potential of being secure, easily implementable and expandable, using considerably less energy. Depending on the application, the size of a PUF Challenge-Response Pair (CRP) set can be different. Applications that demand frequent use of PUF, require enrollment of a very large set of CRPs per PUF unit. This for resource constraint ecosystems, especially in IoT edge device authentication, can become a challenge. In this work our aim is to put spotlight on the prediction power of trained Neural Network models, constructed using Deep Learning techniques, to replace the traditional usage of CRP set tables. Potentially, the trained Neural Networks for that purpose require very small set of CRPs per PUF unit for enrollment (training) and can predict a considerably larger set of CRPs for a given PUF. Different implementation of Neural Network based PUF authentication potentially exist, to which we point out and explain the pros and cons. In addition, we will also discuss other benefits of conducting Deep Learning for enrollment, such as being resilient to instability of PUF CRP, and a Deep Learning based PUF enrollment can be utilized to implement robust key generation for encryption.