Genesis Net: Fine Tuned Dense Net Configuration using Reinforcement Learning

Designing neural networks even in the case of relatively simpler fully connected neural networks / dense networks is a time-consuming process since the architecture design is done manually based on intuition and manual tweaking. In this paper, we present “Genesis Net”, a dense net that starts off with a very basic configuration (“seed configuration”), and subsequently tweaks itself via reinforcement learning (RL) to arrive at an optimal configuration for the task at hand. Genesis Net attained a test error within 0.59% of a similar but bigger documented baseline model. Furthermore, our model was able to achieve this using merely 10.11% of trainable weights that the baseline model used. This significantly smaller network was found using Q-Learning combined with a dynamic action space that allowed for fine tuning the network configuration.