Reinforcement and Deep learning

Now that I have a plan to realize my original work intentions, I spent the last week reinforcing my technical knowledge of artificial intelligence systems. Throughout my conversations with professionals in the field, two recurring concepts in the area were reinforcement and deep learning. It seemed that these fundamental architectures were at the center of premier AI systems that are bound to power the world. I knew I had to learn more about these architectures to gain a grasp on true, modern AI. Firstly, I was able to analyze the research and explanations provided by DeepSense AI, an organization focused on the business implications and strategies of AI using expert opinion and data science. Using their publications, I was able to discern and elaborate on the practical implications of these two concepts. Ultimately, reinforcement learning uses a reward-penalty system in order to develop the most efficient method of resolving abstract situations. Conversely, deep learning uses a node-based, neural network in order to process multiple inputs through multiple layers and produce an output. Ultimately, these two concepts take very different approaches to embodying the "intelligent" side of artificial intelligence. They are and will be at the forefront of machine learning for the foreseeable future. However, as beneficial as they are, I need to ensure that I take a pragmatic approach to my original work. As thought-provoking and powerful that reinforcement/deep learning may be, the best solution to a problem is often the simplest one. Thus, I knew the supervised learning, while not as glamorous as the other methods, was going to be the best design to meet my needs. Supervised learning employs prediction using known data sets to create an algorithm that is initially trained to generate the missing pieces to a partially completed puzzle. Eventually, when the training is complete, the algorithm should be able to accurately predict data regarding a situation using its prior knowledge. Similar to teaching a child fundamental math and asking him/her to solve problems on a test, the algorithm is put to the test using real data. In the case of my handwriting algorithm, I will need to use handwriting training data from various people in order to create a stronger, self-propagating algorithm that truly "learns". Ultimately, the possibilities are beginning to truly open up and I am excited to see what the next week brings in developments.