Effect of scale on catastrophic forgetting in neural networks