Tailoring: encoding inductive biases by optimizing unsupervised objectives at prediction time

From CNNs to attention mechanisms, encoding inductive biases into neural networks has been a fruitful source of improvement in machine learning.