Facebook is using a new technique for marking images used in machine learning, which it called radioactive data. The aim is to help researchers understand whether a particular machine learning model has already been trained with the set of images.
a process that at first glance is simple and useless, but for researchers it can facilitate processes, especially when the machines are being fed with a large data pipeline, on large-scale systems. In this sense, the data can be marked before being placed in the pipeline, which after being stamped can be tested.
This watermark can help researchers and engineers to better understand how colleagues are training their models. The system can even help to detect potential trends in these models.
Facebook calls them radioactive data because they work as an analogy to the radioactive markers used in medicine, for example in certain drugs, such as brio sulphate, in which certain conditions are observed through computed tomography systems. The company says that brands have no impact on models, but remain present in the learning process, being easily detected by neural networks.
