Jim Grey and the Fourth Paradigm

Experimental Science

Theoretical Science

Computational Science

The Fourth Paradigm

• Computer infers the model from the data
  • "Machine learning" - nonlinear curve fitting
  • Optimise fit for training data

Read the book

Why it works

• Theoretical advances in the mathematics of classification and regression
• Computers that are good at global optimisation in large spaces (GPUs)
• Disproportionate effectiveness of machine learning with huge datasets

Data Fundamentalism and the problems of Big Modelling

Weaknesses of the Third Paradigm

• Huge parameter spaces
  • Glorified curve fitting
  • Without statistical rigour
  • Labour intensive parameter management

Weaknesses of the Third Paradigm

• Computational expense
• Large result sets
  • Unwieldy meta-analysis
  • In-situ data reduction

Weaknesses of the Fourth Paradigm

• Does not produce insight
  • Lack of traceability
  • Need for defensible and auditable machine decisions
• Problems with extrapolation: 'what if'.
• Prefers labelled examples (supervision)
  • Crowd sourcing?

The Détente

Use what you know

• Machine learned models with physically informed structure

  • Biological systems and the soft-switch

• Physical understanding as a prior in function space?

Machine assisted parameterisation

• Use machine assistance to explore:
  • Parameterisation literature
  • DBs

Do inference with your models

• Statistically rigorous approaches to parameter fitting and model selection
• Modelling with uncertainty
  • Stochastic programming

Do inference with your models

• Fit simpler structures to your mechanistic models to save time
  • Need a rigorous approach to model run management