Transfer learning is a promising area of developing methods to transfer knowledge and features learned in source tasks to improve the efficiency and performance in a related target task. However, the performance of transfer learning algorithms is currently severely limited. One of the causes for poor performances of transfer learning is in the inherent difficulty of capturing the main features of a model expert of the source task. In the evolutionary process, the highly complex structure of the animal brain is compressed and encoded into DNA to be passed to the next generation, yet it still preserves the highly efficient and robust learning behaviors. Inspired by this observation, in this paper, we propose SPAM, a starting point transfer learning method that compresses the neural network itself from the source task to be given as a starting point of the target task. The results are promising, as between OpenAI Gym’s Inverted Pendulum and CartPole environments, we have observed the reduction in time to reach the maximum returns suggesting more efficient learning.