The basics of deep learning, such as convolutional models, will most likely still exist in some form in five years. Directionally, eager execution and graphs are being combined to make programming feel more natural.