It is now common practice to try and solve machine learning problems by starting with a complex existing model or architecture, and fine-tuning/adapting it to the task at hand. However, outliers, errors or even just sloppiness in training data often lead to drastic drops in performance.
We investigate a simple generic approach to correct for this, motivated by a classic statistical idea: trimmed loss. This advocates jointly (a) selecting which training samples to ignore, and (b) fitting a model on the remaining samples. As such this is computationally infeasible even for linear regression. We propose and study the natural iterative variant that alternates between these two steps (a) and (b) - each of which individually can be easily accomplished in pretty much any statistical setting. We also study the batch-SGD variant of this idea. We demonstrate both theoretically (for generalized linear models) and empirically (for moderate-sized neural network models) that this effectively recovers accuracy in the presence of bad training data.
This work is joint with Yanyao Shen and Vatsal Shah and appears in NeurIPS 2019, ICML 2019 and AISTATS 2020.