Many organizations fail to leverage their existing ML pipelines to perform the same predictive tasks on new data modalities (e.g. video, audio), which can significantly delay model development. Suri et al. demonstrate how Google utilizes “organizational resources” like knowledge bases and aggregate statistics, to connect data points across modalities. Using organizational resources, they create common features (i.e. by transforming data to representations shared across modalities). Then, they generate labeled data by leveraging these shared features and weak supervision and employ multimodal techniques for model training on multi-modal data from disparate label sources. They demonstrate how this approach is used for various production-scale classification tasks at Google to drastically reduce development time.