Talks on Feature Stores and Machine Learning Security
Tuesday 10th May saw our MLOps London meetup return, this time featuring talks on the hot topics of feature stores and security in machine learning. If you haven’t heard of them already, feature stores are a way of storing commonly used features required either for model training or predictions.
Moritz Meister, Head of Feature Store Engineering at Hopsworks, is just about the best person you could ask to come and explain how they work. He kicked us off with a talk about how Hopsworks implemented a high throughput, low latency reverse ETL pipeline. During this talk, he covered the challenges of ensuring schema consistency and consistent stable update rates between data warehouses, an intermediate Kafka Cluster and the operational database RonDB
Next up was Seldon’s very own Alejandro Saucedo, Director of ML Engineering and Chief Scientist for The Institute of Ethical AI. Alejandro had been working on a talk for Pycon DE which had piqued my interest. There is so little research or information around how to do security properly in machine learning that everyone was desperate to find out more…
Alejandro’s talk covered the various steps of the machine learning lifecycle and showed the kind of vulnerabilities that can be introduced at each step. He then gave some advice and best practices on how to mitigate risks in an automated fashion.
You can catch the full recording of Alejandro’s talk here. The code and presentation slides can also be found in the GitHub repository. Our next MLOps London will take place on the 19th July where we’ll have talks on Autonomous Vehicles as well as Graph Analytics in ML. As always, it’ll be broadcast live on our Youtube channel if you aren’t able to come along in person. Hope to catch you there!