3 ms·
Epoxy: Interactive Model Iteration with Weak Supervision
- cs702 6y agoNeat. You would use this, for example, if the distribution of your data is prone to change over time and training or finetuning a model to acceptable accuracy quickly enough is impractical. To use it, you need two things: * Noisy labeling functions -- e.g, user-provided heuristics or methods for labeling samples, or existing models trained on subsets of the data. These functions need accurately classify only a subset of possible samples. * Embeddings from a pretrained model -- e.g., a convnet or language model. These embeddings cover practically all possible samples but do not classify them. At a very high level, the authors iteratively extend the coverage of the labeling functions using nearest neighbors search in the embedding space, and then aggregate them using a library called FlyingSquid (by the same authors; there's a link on the repo) for quickly building an aggregate labeling model.