Model-agnostic meta-learning (MAML) and itsvariations have demonstrated remarkable performancein few-shot learning scenarios, wherelimited labeled data is available during modeltraining. However, owing to inefficient learningand limited cross-domain generalization ability,MAML based methods face severe challengeof failure, when there are significant gapsin the distribution between training and testingtasks. Inspired by classical proportional-integralderivative(PID) control theory, such issues can bealleviated through efficient parameter optimizationmethods. With the motivation to improve theconvergence of the network and accelerate the optimizationprocess, this work designed a MAMLbased layer-adaptive PID optimizer (LA-PID).Specifically, the generated network of LA-PIDis devised to dynamically generate task-specificPID control gains per layer. Simultaneously theoptimal convergence condition initialization is analyzedfrom the first principles perspective. Aseries of experiments on four benchmark datasetsvalidated the effectiveness: LA-PID achievedstate-of-the-art optimization in few-shot classificationand cross domain tasks with fewer trainingsteps.