Closing the data gap: human-in-the-loop strategies for keeping a model relevant

Codon has built a content moderation system for a Swedish social media app. This talk will describe how we approached and overcame some of the many problems involved, such as scarce and heavily skewed data, concept drift and expectation management. The model, now two years old, is still live and continuously updated and we will also discuss how that is done.

• Data augmentation for textual data
• Metrics and expectation management
• Prediction confidence estimation

————————————————————————————————————————————————————

Mikael Huss – Founder & Data Scientist | Codon Consulting

Mikael is currently co-founder & principal data scientist at Codon Consulting, a company which builds machine learning systems for a variety of clients in life science, agriculture, finance and other sectors. After a research career in Singapore and Stockholm, which resulted in a PhD and an associate professor title in bioinformatics from KTH, he decided to pursue his long-standing interest in machine learning and in industry. This led him to work as a senior data scientist at IBM and Peltarion. After co-founding Codon, he has devoted his attention to things like transformer models in biology, content moderation with language models, and crop yield forecasting with computer vision. Mikael has also previously been an active participant in Stockholm’s meetup ecosystem as a meetup organizer and speaker for Stockholm AI and SRUG (Stockholm R useR group). He’s excited about large language models and the convergence of AI and biology.