A Quantitative Approach to Predicting Representational Learning and Performance in Neural Networks