Closed-loop control of functional electrical stimulation (FES) can restore movement to paralyzed limbs by using recorded nerve signals to modulate stimulation in real-time. Previous work performed state-of-the-art classification of nerve signals using neural networks (NNs). This thesis explored reduced NNs for use in implanted systems and incorporated such NNs into closed-loop stimulation. NNs of varying complexities were evaluated for performance and resource usage when classifying individual naturally-evoked compound action potentials, previously collected from rat sciatic nerves using 7x8-channel cuff electrodes. Acute in vivo experiments used NNs to estimate nerve branch activity in real-time as feedback for neural stimulation, producing desired ankle movement trajectories. Reduced NNs achieved macro F1-scores of 0.70-0.71 while remaining within resource constraints defined by NN implementations in integrated circuit fabrication technologies. We demonstrated closed-loop stimulation in 6 rats. NNs that classify nerve signals can be reduced to fit implantable hardware platforms and used for closed-loop stimulation control.