Learning, truth and data science

It is commonplace these days, to hear expressed the view that humankind, taking all things into account, is not the blessing to our planet we might like to think.

This is a view you agree with or not. A conclusion that is difficult to argue with though, is that our species has come to dominate our planet, and to shape its environments, and the lot of its other inhabitants, like no other species.

This means, in evolutionary terms at least, that humankind is a spectacular success.

But what is at the root of this success? Is it innate?  Are we born with the skills and knowledge to succeed?

Far from it.  At birth, a calf on the savanna has inherited the skills required to run; useful on the savanna if it is to survive for very long. By contrast, a human baby is, by itself, helpless. Relative to other species, humans are born knowing nothing and able to do nothing. We also have the longest childhoods of any species.

But ironically, this relative ‘blank slate’ state is a product of our successful evolutionary journey.  Evolutionary biologists call this neurological plasticity. In simple terms, the advantage of plasticity is that it means we are less set in stone at birth and better able to learn and adapt to the conditions in which we find ourselves (Haying & Weinstein, 2021).  It is our unique ability to learn that is key to our success.

We are learning machines.  We enjoy marvelling at the problem-solving ingenuity of our animal cousins, but we are blasé about our own technology and all that our species has achieved.  And not only are we good at learning, but learning is good for us too (“Lifelong learning is the secret to happiness in old age” The Guardian, May 2011).

But how do we learn?  In part, we learn from one another.  More than other species, our mastery of language means that we can pass knowledge to each other, to our children, within our tribes and beyond, face-to-face, through stories, in books and now via the worldwide web.

But even this familiar process may not be as it seems.  In his 2001 book “Closure: A Story of Everything” philosopher Hilary Lawson explores whether the relationship between language and the reality it is used to describe, works in the way we think.  Central among the ideas in the book is that human beings make sense of the world through a process he calls closure which he defines as “holding that which is different as one and the same”.

So, an infant learns about a cup by experiencing it through her senses.  She holds it, drops it, puts it into her mouth and experiments holding it in different orientations with varying results.  She learns what a cup is and what a cup does, and eventually, as her language skills develop, she learns the word that describes that experience.

But, Lawson argues, every cup, person or circumstance is unique and could combine in an infinite number of possibilities. The concept of ‘a cup’ is therefore an example of a ‘closure’ – in the right context we are holding a universe of different things, to be the same thing – a cup – because it is useful to do so.   The word ‘cup’ is a proxy or a tag for our experience and very useful tool for making interventions in the world – if the intervention we want to make must involve needing a cup.

And this, Lawson continues, is the same for everything.  Our language is a kind of shorthand for experiences we have had and expect to have in the future, rather than a description of an objective reality.

One of the implications of this concept is that in order to maximise our understanding and interaction with our world, we first have first close off all that it could possibly be and learn to live with a more manageable and useable set of possibilities.  This compromise, moving from openness towards closure, is fascinating for philosophers like Hilary, but also has parallels in how we are now understanding our evolutionary and physiological development.  For example, our new-born human has the capability of learning any and possibly all languages at the moment of her birth, but that capacity degrades as she grows, not just because she likely only hears the single language of her parents as she develops but also, studies now reveal, because the parts of her brain that offered this possibility are repurposed for more essential tasks, certainly by the time she reaches her teens. Pragmatism rules when gene survival is at stake.

But following Lawson’s theory, it becomes quite easy to see how the pragmatic compromise of closure can also create difficulties, particularly when we need to come together with other people to achieve things.

This is because we can only reference our personal experiences when defining our ‘closures’, not anyone else’s.  A cup is a fairly straightforward thing to understand and talk about because is represents a simple and universal experience.  But what about a business process?  If we have worked in a business of any modest size, we will know that processes are important and will have a view of how they work.  But is your understanding of a process the same as mine?  Is your understanding of the words you are using to describe it the same as mine?

In Lawson’s world, it can’t be because your language is based on a different experience than mine.  And if you don’t think that this has any material consequence, try gathering the senior team of your business together and asking them each to describe your business management system.  You may find some diverging opinions.

So what about predictive models?  If Lawson is right and our language at least, cannot define an objective current reality, is it reasonable to expect a mathematical model to predict the future?  And if isn’t, what use are they?

While data science is a relatively new sector, mathematical modelling is not new.  The COVID-19 pandemic has however, thrust modelling into the spotlight, and some modellers are now household names to a degree that they would never have imagined or possibly wished for.  And largely, this has not gone well.

Maybe the COVID modelling teams bear some culpability for this (there is modelling and good modelling just as there is maths and good maths) but for certain the actions of news media and politicians from all sides keen to be seen to be ‘following the science’ has resulted in a repeating cycle: dramatic headlines focused (usually) on modelled worst-case scenarios, with no acknowledgement when those predictions (happily) proved wide of the mark.  This has left the public at large in no doubt about modelling’s ability to predict the future.

What has been largely missed throughout was the opportunity to explain to the public how mathematical models work, what they can and can’t do and why and how they are useful.  A rare exception to this was a comment by Sunetra Gupta, Professor of Theoretical Epidemiology at the Department of Zoology, University of Oxford (Spectator TV, August 2021) where she stated:

Models are absolute crucial and invaluable as tools for generating testable hypotheses, [but] they should never be used to predict. And while they should be used a basis for formulating policy, they should not be treated as truth in any sense”.

So can predictive models give us future truths.  No, not least because, if Hilary Lawson is right, we would have no language to communicate such a thing if it existed.  But if not, are they useful at all?

Absolutely.  Humans, as we have discussed, are learning machines and like our cup-wielding infant, we learn by experience, by doing new things and assessing the results that came from those interventions.  Learning through experience can be expensive and risky.  But it is the only way in which we learn anything genuinely new.  Ever.  Our ancestors benefited from the wisdom of their elders not because it meant they didn’t need to try new things – they had new problems to solve after all – but because their elders advice made learning by doing more efficient and less risky, by providing testable hypotheses of how new problems could be solved.  New things still had to be tried, but we could focus on those approaches which experience had taught us were most likely to work and avoid those that we had previously learnt didn’t work.

Similarly, predictive models are useful because they guide us in further exploration or considered risk-taking and bet placing, helping move us to action and new learnings more safely and more efficiently.

As data scientists we should not be defensive about the fact that our models can’t tell us with certainty what will happen in the future.  Instead we should explain how they work and how they can best be used.  And as clients and project sponsors, we should explain to our colleagues that we can’t learn anything new until we do something new, and that models are useful tools in helping us to take those first new steps.

By Kevin Cornwell
Categorized as blog

Leave a comment

Your email address will not be published. Required fields are marked *