<style> .reveal { font-size: 32px; } </style> <style> section { text-align: left; } </style> --- # AI for Agriculture # Experiences from livestock disease (and a few other bits) ## I want to mention a few themes via examples: - modelling a system that is a moving target - producing socially feasible outputs - modelling/analysis where data cannot be shared --- ## Modelling a moving target - modelling a system that is **stable** is the easiest <span style="color:green">green</span> - modelling a system that is **changing** is <span style="color:orange">harder</span> - modelling a system that **changes in response** to your model is <span style="color:red">even harder</span>! ## Making socially viable decisions - some outputs are **impractical** - some outputs are socially or politically **infeasible** --- An example that covers both: ## Which farms are highest priority for disease surveillance?  Various forms of this kind of work via EPIC and elsewhere. - a complex output isn't **feasible** - needs to be explainable in a pamphlet - needs to avoid 'computer says no' - simple outputs may seem unfair or be gameable - e.g.: farms of intermediate size have highest risk --- ## Data may not be shareable - commercially sensitive - lack of incentives to share - slow data sharing in an emergency - recent project: DigiVet (https://zenodo.org/communities/digivet/) There are ways to deal with this: - socio-approaches: sharing benefits all - technical approaches: - privacy-preserving approaches federated learning - e.g. https://ioftdatatrustwg.github.io/ Despite these challenges, much opportunity!