I’ve been working on an eXplainable AI (XAI) project and today I’ve published the very first version of the SDK that will eventually call the service. It’s also the first time I’ve ever published a package to PyPI; most of my professional work has been proprietary and I usually just release my side projects with permissive licensing on GitHub. So it’s exciting in a way to have an “official” package in some sense.
So what’s ProxyML (“proximal”, “proxy ML”, take your pick)? Quoth the repo:
Most explainability tools require sending your data to a third-party server. ProxyML never sees your training data. You generate synthetic data locally, score it with your own model, and only the surrogate model and summary statistics are uploaded — your data stays yours.
There are any number of explainability tools out there, but most require access not just to your model but to your data as well. That’s a real problem in industries such as health care or finance that can’t / won’t / shouldn’t share data. ProxyML works by looking at the descriptive stats behind your data instead, so your data never leaves your server.
