Scaling-Aware Data Selection for End-to-End Autonomous Driving Systems

2026-04-09 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

A new data selection framework, Mixture Optimization via Scaling-Aware Iterative Collection (MOSAIC), has been developed to improve training data efficiency for large-scale deep learning models in physical AI applications. MOSAIC addresses the challenge of ambiguous data point effects on various evaluation metrics by partitioning datasets into domains, fitting neural scaling laws from each domain to evaluation metrics, and iteratively optimizing data mixtures. Applied to autonomous driving (AD), MOSAIC was used to train an End-to-End (E2E) planner model, which was evaluated using the Extended Predictive Driver Model Score (EPDMS), an aggregate of driving rule compliance metrics. This framework achieved superior performance on EPDMS, utilizing up to 80% less data compared to diverse baselines.

Key takeaway

For research scientists developing large-scale deep learning models for physical AI, adopting the MOSAIC framework can significantly enhance data efficiency. By strategically selecting data based on its impact on specific evaluation metrics, you can achieve better model performance, such as improved driving rule compliance in autonomous driving, while substantially reducing the required training data volume. Consider integrating MOSAIC's iterative optimization to streamline your data collection and model training processes.

Key insights

MOSAIC optimizes data selection for physical AI by iteratively adding data from domains that maximize metric changes.

Principles

• Partition datasets into distinct domains.
• Fit neural scaling laws per data domain.
• Optimize data mixture iteratively.

Method

MOSAIC partitions data, fits neural scaling laws from domains to evaluation metrics, then iteratively adds data from domains that maximize metric changes to optimize the data mixture.

In practice

• Apply MOSAIC to End-to-End autonomous driving.
• Evaluate E2E planners with EPDMS.
• Reduce training data by up to 80%.

Topics

• Data Selection
• Autonomous Driving
• Deep Learning Models
• MOSAIC Framework
• Neural Scaling Laws

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• MOSAIC: Modular Orchestration for Structured Agentic Intelligence and Composition — MOSAIC structures LLM-based agentic model selection through blueprints, improving automated data science performance and traceability.
• Explaining Data Mixing Scaling Laws — A unified framework explains data mixing scaling laws through capacity competition and noise reduction.
• Drive-KD: Multi-Teacher Distillation for VLMs in Autonomous Driving — Small VLMs can achieve superior autonomous driving performance and efficiency through multi-teacher knowledge distillation and targeted attention signals.
• MIT Proved 90% of Every AI Model Is Dead Weight. It Took 8 Years for the Hardware to Catch Up. — Most neural network weights are redundant; efficient subnetworks can be trained directly with modern hardware and tooling.
• First-Principles Optimizer Matches Adam on CIFAR…No Tuning — A first-principles optimizer dynamically adapts learning rates, matching Adam's performance without hyper-parameter tuning.
• A Mathematical Conflict Framework for Contextual Data Modulation — The framework explicitly models data conflict as an independent, context-sensitive mathematical object, unlike implicit optimization approaches.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.