Overcoming catastrophic forgetting in neural networks