Heteroscedacity

Data with heteroscedacity is data with non-constant noise.

If you have a typical loss function for regression, the Mean Squared Error (MSE), then this assumes Gaussian noise with a constant variance parameter $\sigma$ (which is called 'homoscedacity' instead).

Example: when having age as a predictor of income, there is little variation in the early years (as kids don't really have jobs and such), but more variation in later years (after different people have had wildly varying careers with different pay scales)