Automated systems development often faces high costs and time consumption when manually labeling test data. What approaches generally optimize this process? How effective are methods like active learning, weak supervision, or synthetic data generation? What experiences does the community share? Which strategies stand out for a scalable solution?
Recommendations for data labeling methods for autonomous systems
👁️ 6 views💬 1 replies❤️ 0 likes
1 Replies
Okay, I get that you're looking to optimize data labeling for autonomous systems, but tell me: have you tried combining **Synthetic Data Generation** with **Active Learning** instead of relying on just one of them?
From experience, synthetic generation gives you a massive, cheap base dataset, but without active learning, it can end up with biases or poorly covered edge cases. For example, if your autonomous system is going to operate in very specific urban zones (like roundabouts in your city), have you considered generating synthetic scenarios **only** for those specific areas instead of generic data? That way, you reduce the "noise" in the initial labeling and then refine with active learning on the critical edges.
And another thing: what tools have you looked into for active learning? I’ve seen that some libraries like **ModAL** or **prodigy** (from spaCy) allow you to integrate pre-trained models to suggest which samples to label first, but if you're using your own model, have you evaluated how computational cost affects scalability? Because in large projects, inference to prioritize samples can become a bottleneck.