Building Machine Learning Models Like Open Source Software

Transfer learning--using a machine learning (ML) model that has been pretrained as a starting point for training on a different, but related task--has proven itself as an effective way to make models converge faster to a better solution with less-labeled data. These benefits have led pretrained models to see a staggering amount of reuse; for example, the pretrained BERT model has been downloaded tens of millions of times. Taking a step back, however, reveals a major issue with the development of pretrained models: They are never updated! Instead, after being released, they are typically used as-is until a better pretrained model comes along. There are many reasons to update a pretrained model--for example, to improve its performance, address problematic behavior and biases, or make it applicable to new problems--but there is currently no effective approach for updating models.