Part of my series of notes from ICLR 2019 in New Orleans.
Introduction
- need actionable information on climate risk
- have vast datasets (unlike in many other fields!)
- we ML people should be able to do something about this
- issues particular to climate change data science
- structured, noisy, non-stationary data
- need for extrapolation & handling of extreme / rare events
- (this is a theme at ICLR so far, think about this)
- healthcheck: monitoring the patient, treating the symptoms, curing the disease
Monitor, Treat, Cure
- monitor – digital online platform
- need interpolate and synthesize / integrate data
- treat – models for evidence-driven decision making
- random fact: typical climate model is >1 million lines of Fortran code
- global models of e.g. surface temps are pretty good, BUT
- need more granularity for the level that decisions are made – e.g. in cities
- current models have systemic bias at local levels
- learn from observed weather to go from coarse- to fine-grained models
- (kinda like image superresolution…)
- account for correlated risks and impact of regulations / policies
- cure – solutions for climate change
- e.g. modeling arctic sea ice - physics at many scales, uncertainties etc.
- but we’ve got data!
- => blend data-driven and physics-based approaches
- but we’ve got data!
- understanding key processes and causal relationships
Conclusion (and Q&A)
- should establish benchmark tasks to drive climate change research (cf. ImageNet)
- another challenge – takes months of work to prepare a dataset for use in this field
- good models are useless if people ignore them…
- getting things to local level is critical – global average surface temperature is abstract and hard for people to engage with
- the trustworthiness of the results is incredibly important for implementation – for many people, for better or for worse, this is linked to interpretability
- (I think these are both really important points to keep in mind)