Looking ahead in the context of my future being purely a programmer, it honestly seems pretty dreary. I can’t see myself happy coding and fixing my code for 6 to 7 hours a day working for a huge company. Coding has never been too attractive to me, just as English never really has. But none the less, I still enjoy writing. Maybe I will learn to look at coding in the same perspective, as a medium to get things done rather than just a tedious task. I want to do, research about what fascinates me in this world and getting a tangible reward is important. Creating intelligence, creating deep neural networks fascinates me because I see how all of the back end math comes together to make systems that have intelligence. Learning how a machine could learn some basic patterns in data and apply them with a function approximation is so interesting because it is so close to what happens in us. This same mystery exists within us, and as we question things about creating intelligence in machines, we in turn explore our own creation and design. It attracts me not because I want it to, I have no external push towards being a DL or ML engineer because of any parental pressures or the salary that comes along with it. I’m attracted to the connections to other fields of study, how this branch of computer science, deep learning, has so many implications and connections into philosophy, neurology, psychology, control theory and much more. Building a good understanding for deep learning will allow me to think about things in a different broader scope.
My reinforcement learning journey has been great, being able to understand code through the context of psychology and training agents to solve environments. I think what keeps me on track for reinforcement learning specifically is its connection to learning based on experience, just like we do. Trying to make connections between the way I believe our internal reward structures are designed and the reward structures in algorithms interests me deeply. One of the core problems in RL is called the Credit Assignment Problem. And what this problem is getting at is, to what actions and to what states can we blame for a given state. When we look at all of the actions and states that an agent visited over the course of an episode, it is not obvious. Moreover, this psychological perspective is actually beneficial to have when designing learning algorithms, whether that be value or policy based. The difference between what I study and the typical comp sci student would be where these ideas come from. Computers are analytical, and are not inspired by the nature of humans. On the other hand, ANNs come from biological systems, and everything in the study of DL or DRL has come from this idea of copying from nature and using the mechanisms we observe in nature to better our systems.