Inductive biases are any assumptions that learners utilize to learn the world and predict the output that reduce the amount of data needed to fit the model while constraining the model’s flexibility.
Inductive biases are any assumptions that learners utilize to learn the world and predict the output. These assumptions can either be related to the date-generation process, the functional family space chosen, or the space of the solutions. Inductive biases could be a regularization term added for better generalization or preventing overfitting, and they could also be embedded in the model function family or model architecture (e.g. CNN vs MLP). Inductive biases generally reduce the amount of data needed to fit the model while constraining the model’s flexibility. Ideally, inductive biases will improve training efficiency (small data points, small gradient steps, faster optimization), while maintaining performance and generalizing well. However, if inductive biases are mismatched with the problem domain, this will lead our solutions to be sub-optimal. Another way to think about this is through the lens of the bias-variance tradeoff: inductive biases induce a large bias term in approximation error.