MIT algorithm generates extreme event scenarios without rare data
MIT News AI · rss · 2026-08-25
MIT engineers have developed a machine learning algorithm called "Extreme Event Aware" (η-learning) that can generate plausible future extreme scenarios without relying on historical extreme event data.
How it works:
- Combines point statistics (e.g., frequency of max rainfall) with spatial map data.
- Learns patterns from low-res maps and uses statistics to constrain extremes, inferring unprecedented yet plausible events (e.g., a 300mm storm never seen before).
Applications:
- Generates detailed impact maps for extreme weather, floods, and wildfires to aid urban planning and infrastructure reinforcement.
- Extendable to financial market crashes and robotic navigation.
The research was published in Nature Communications.
More from Research
- AI Security Scholar Jingxuan He to Join NTU as Assistant Professor, Previously at NVIDIA Research — xuandongzhao · 2026-08-25
- Traditional ML scholars critique GenAI hype and rebranding — wandb · 2026-08-25
- AI Bot Solves Open Math Problems on MathOverflow — latticecut · 2026-08-25
- Weekly Roundup: Humanoid Robotics Papers, including a football-throwing bot — carlosdponx · 2026-08-25
- Adobe releases LDR: Video world models via latent dynamics — _akhaliq · 2026-08-25
- Astra's Continual Learning Blueprint from GPT's Goblin Problem — imjustnewatai · 2026-08-25