38
Dr. Oliver Hulme Computational models and Education How computational models of development could apply to educational practice Education and Neuroscience workshop Oct 2008

CEN Launch, Oliver Hulme

Embed Size (px)

Citation preview

Page 1: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Computational models and Education

How computational models of development could apply to educational practice

Education and Neuroscience workshop Oct 2008

Page 2: CEN Launch, Oliver Hulme

Computational models and Education

Computational modeling of cognitive development offers a novel framework for thinking about how a child does something (counts, reads, learns etc)

Models make concrete predictions about developmental processes and therefore could be used to predict the outcome of educational interventions

This provides opportunity for formally optimising educational interventions

To delineate how, I will explore a (crude) comparison between medicine and educational neuroscience

Page 3: CEN Launch, Oliver Hulme

Medical and Education?

Commentators have highlighted the abstract similarity between medicine and education (Schlagger, Fischer etc)

Biology

Medical Interventions

Educational Interventions

Medicine Educational neuroscience

Page 4: CEN Launch, Oliver Hulme

Medical and Education?

Biology

Medical Interventions

Educational Interventions

Given that Medicine has translated biological knowledge into real world interventions that work…

How can Educational Neuroscience adapt this paradigm?

Developmental modelling could be critical for this translation

Page 5: CEN Launch, Oliver Hulme

Medical Paradigm Medicine predominantly intervenes through pharmacology

Problem: Millions of candidate drugs but cannot clinically test all due to ethical and resource constraints

Solution: Filter out poor drugs + clinically test few that are most likely to work

Candidate drugs

Lead compound

Clinical trial Market

Page 6: CEN Launch, Oliver Hulme

Drug discovery by screening

Computationally model the potential efficacy of candidates to identify lead compounds (virtual high throughput screening)

Pre-selects drugs for clinical testing

Screening

Candidates Lead compound

Page 7: CEN Launch, Oliver Hulme

Educational Paradigm?Education intervenes through technology and teaching

Problem: Millions of potential interventions but cannot test all due to ethical and resource constraints

Solution: Filter out poor interventions + only empirically test select few that are most likely to work

Candidates interventions

Lead intervention

Pedagogical trial

Schools

Page 8: CEN Launch, Oliver Hulme

Concrete examples?

Suppose we are trying to optimise reading performance for phonological dyslexia through educational intervention

Problem:

The space of possible interventions is infinite

How to define reading performance

Solution:

Test subspaces (i.e. a limited set of dimensions)

Operationalise reading performance with an established index

Page 9: CEN Launch, Oliver Hulme

For exampleOperationalise reading performance by word recognition accuracy

Select subspace of interventions involving simultaneous presentation of graphemes and phonemes and select 3 variables to manipulate

Even this subspace contains a large number of candidate interventionsR

ew

ard

Repetition rate

Age

Intervention subspace

Page 10: CEN Launch, Oliver Hulme

Candidate Interventions?R

ew

ard

Repetition rate

Age

Screen: Use computational models of development to screen the potential efficacy of each intervention on word recognition accuracy

lead interventions (eg Predicted optimum is age 12, repetition rate = 10, reward = 3 hedons)

Models direct us to parts of intervention space that predict educational best results

Screening

Page 11: CEN Launch, Oliver Hulme

Pedagogical trials

Only empirically test the efficacy of the lead interventions in ecologically valid settings

Select educational interventions based on evidence

Page 12: CEN Launch, Oliver Hulme

Advantages of using models to screen

Can explore educational interventions which may be unethical to pilot on humans

Can explore how multiple variables interact to impact on educational performance

Can test parts of intervention space that would be appear too stupid to test in real life and opens up possibility of unexpected results

Can test effects of interventions on whole developmental trajectory not just a discrete timespans

Page 13: CEN Launch, Oliver Hulme

Open questions

Framework applies to any type of computational model. Are connectionist models suitable for this implementation?

Given that connectionist models ‘are not intended to be neural models, but rather cognitive information processing models’ (Mareschal and Thomas 2007)

The question is whether they are sufficiently grounded in biology to yield the accuracy required for predicting the outcomes of educational intervention

Page 14: CEN Launch, Oliver Hulme
Page 15: CEN Launch, Oliver Hulme

Problems…

How to map from model to reality

How generalisable are the abstract tasks the model performs

Difficulty of mapping existing interventions onto model parameters and vice versa

Page 16: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Page 17: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

What will the computational guys think

What will the educators think

What will the neuroscience guys think

Page 18: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Title

Biology

Medical Interventions

Educational Interventions

Page 19: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Title

B

Biology

Medical Interventions

Educational Interventions

Page 20: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Title

B

Biology

Medical Interventions

Educational Interventions

Page 21: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Format

Page 22: CEN Launch, Oliver Hulme

Most developmental psychologists describe what children can do not how they they do it

Knowing how a child does something, count, read, reason, offers the opportunity for principled intervention to improve or remediate that faculty. This would be the long term aim of educational neuroscience, having a physical theory of childrens development, and its interaction with educational interventions. Through these models one could facilitate optimal trajectories for each child.

Page 23: CEN Launch, Oliver Hulme

Computational Modelling• A standard in the physical sciences

• They are tools for exploring causal mechanisms of development

• Can track HOW learning mechanisms interact with the characteristics of the environment to produce observed behaviors

Equally they would be tools for exploring the causal mechanisms of educational intervention and child development.This opens up the space of possible interventions and allows one more efficiently to identify subspaces which offer optimal developmental trajectories.

How do the learning mechanisms interact with the characteristics of the educational environment to produce observable educational outcomes.

Page 24: CEN Launch, Oliver Hulme

Computational Modelling

Of course, all models involve approximations!

Making the right approximation depends on the nature of the target problem and the of our understanding of the problem.

This involves collaboration between the psychologist, pedagogues, pedagogists, computational neuroscientists,

Page 25: CEN Launch, Oliver Hulme

Building a model is NOT the same as building a baby!

"The art of model building is the exclusion of real but irrelevant parts of the problem, and entails hazards for the builder and the reader. The builder may leave out something genuinely relevant; the reader, armed with too sophisticated an experimental probe… may take literally a schematized model whose main aim is to be a demonstration of possibility."

The question is whether in designing complex interventions the accuracy of the models trajectory may depend on what is left out.

Models are tools for reasoning

Page 26: CEN Launch, Oliver Hulme

Connectionist Models

• Cognitive models loosely based on neural information processing

• Develop internal representation as part of learning

• Not tabula rasa systems.

I want to relate the mechanisms of neural information processing to behaviors characteristic of cognition

Page 27: CEN Launch, Oliver Hulme

An illustrative example…

The What and Where of infant object-directed behaviors

•Some old stuff

•Some new stuff

•Some future stuff

Page 28: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Page 29: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Page 30: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Emphasis is on interacting with the environment.

Education is an environmental intervention and therefore these models are suited to testing different interventions without ethical concerns of experimenting with a child’s developmental trajectory.

Page 31: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Page 32: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Full developmental model should also respond in the same way as a normal child to environmental interventions such as education.

Page 33: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Page 34: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Page 35: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

PlanWhat are the parameters of an intervention?

Reward schedule, magnitude, valence, frequency

Optimal stage in developmental trajectory

Longitudinal frequency?

Personalised developmental models?

Given a set of data pertaining to the infants pscyhology, set up a individualised model, against which one can test your interventions for a full blown individualised learning scheme

Model full classrooms?

Macroscopic forecast modelling of educational policy?

Page 36: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

How does it change now?

What will it change in 5 years time?

In 25 years time

Page 37: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Review of Posners book by Bradley Schlaggar

Page 38: CEN Launch, Oliver Hulme

Dr. Oliver Hulme

Plan

Kurt Fischer and others in the 1st issue of MBE highlight the analogy between MBE and medicine, both using knowledge to inform practice.

Medical applications of biological knowledge require independent empirical tests.

Educational applications of biological knowledge also require independent empirical tests.

Neuroscience and modelling of neuroscience data require judicious interpretation followed up by research that tests their application to the classroom