A Conceptual Framework for Deep Learning Algorithms and Their Applications

Deep learning plays an important role in our lives. It has already had a huge impact in areas such as cancer diagnosis, personalized medicine, self-driving cars, predictive analytics and speech recognition. Intuitive hand-crafted features used in traditional learning, classification, and model recognition systems are valuable for large data sets. In many cases, depending on the complexity of the problem, DL can also overcome the limitations of sparse networks in the past that prevent effective training and spatial representation of high-dimensional training data. A deep network uses many layers (deep) of units with sophisticated algorithms and architecture. This paper reviews several optimization methods to improve training accuracy and reduce training time. We delve into the math behind the training algorithms used in the latest deep networks. We describe current failures, improvements, and implementations. The review also covers different types of deep architectures, such as deep convolution networks, deep networks, regular networks, reinforcement learning, differential autoencoders, etc.