Understanding complex relations between neuronal activity and animal behavior is central question in neuroscience. Rapid advancements in Artificial Intelligence (AI) methods offer powerful tools to investigate highly non-linear mapping between motor cortex activity and body movements. Here, we developed a Generative Adversarial Network (GAN) that showed that detailed videos of behaving rats can be recreated from activity of just few selected neurons. This analysis also revealed that the predictability of behavior from neuronal activity (and vice versa) initially increases as a rat learns a new task. However, after the animal performance on the motor task achieves the required accuracy, then coupling between neuronal activity and behavior decreases, without degrading task performance. A plausible interpretation is that, as training progresses from Early to Mid training days, more neurons become engaged, forming a denser, broadly distributed representation, which then in the Late training days evolves into a sparse and more energy-efficient representation, with only a small subset of tuned neurons. Neuronal network simulations showed that such changes in coding strategy may be explained by neurons minimizing their energy use. Thus, our approach reveals a non-linear relationship between learning stages and neural-behavioral coupling, which is likely driven by energy efficiency.