In my last post, I had made a case for expertise – and why it is important to have the necessary domain expertise in high performing teams. And not surprisingly, I got several responses arguing that experts are over-rated. Which I tend to agree with – experts are over-rated, but expertise is definitely not. The real question then is, how do you build learning systems that gain expertise?
[Sidebar]: I am currently reading a fascinating book on Physics called Fundamentals: Ten Keys to Reality by Frank Wilczek, a Nobel prize winning Physicist. The beauty about the book is that he has managed to distill the most fundamental concepts into a language that makes it possible to engage laypeople in contemporary physics (it is not an easy read but undoubtedly engaging). The reason I bring it up – in one of the chapters on elementary particles, he comes up with four basic principles that beautifully capture the laws which govern the elementary particles. And that got me thinking – can we put together a few principles that govern all learning systems? Here is my attempt – and I am sure that there are better thinkers out there who have thought more deeply about learning systems. For the rest of this blog, I will use ‘system’ in a very informal, loose way – to refer to a business function within an organization. This could just as well apply to individuals or social groups et al
#1. Learning starts with a priori hypotheses: There is a growing body of research that points that babies are not, as was previously assumed, born with a ‘blank slate’ – we are all born with a mental model of the world, our initial hypotheses. Similarly, you should be able to describe a system with a set of initial hypotheses of two types: a set of states (which describe ‘what there is’) and a set of rules (which describe how a ‘change in states’ can happen). These hypotheses can act as powerful accelerators in building a learning system – which is why it is a critical first step in setting up a learning system.
Once you get past the meta question of ‘why must a system exist’ and trying to define a pre-ordained end goal, it becomes apparent that learning is all about incremental growth – enabled by these four principles:
#2. Learning is exploring a space of probabilities: Often the most effective learning systems focus on determining the ‘next best action’ by looking at a series of options, each with some kind of a probability and executing the one with the highest probability of success. And so, it follows that the effectiveness of learning comes from the ability to defining options (more the better) and assigning probabilities to each one of them.
#3. Learning is restricting the search space and is local: Exploration without constraints will end up draining resources – that much is obvious. What this implies is that an effective learning system must have the ability to restrict the search space and be comfortable with optimizing for a local optima, as opposed to chasing a global optimization function. This is important on two counts – drives towards a bias for action with a mindset of incrementality; not settling for the local optima but continuing to attempt to seek a better state.
#4. Learning is optimizing a reward function: The idea of seeking a local optima requires the definition of a reward function. In some cases, you are able to define a pre-defined outcome: e.g. release a product by a specific date, reduce costs by a specific factor. Real-life situations tend to be far more ambiguous: and here’s where supervised learning models can teach us a thing: start with a simple model and measure the error. And through a process of trial-and-error, learn from the error (formally known as back propagation) by tweaking the parameters to do incrementally better than the previous iteration. And repeat until it cannot get better any more. This is, in essence, the ‘gradient descent algorithm’ which incrementally reduces the probability of errors until it cannot get any better.
#5. Learning is adjusting parameters based on feedback: Again, we can take inspiration from neural networks: they start with an initial set of parameters (formally weights in each layer) and based on error rates obtained in each iteration, adjust the weights. And it naturally follows that with more iterations, the system keeps on getting better – there is nothing like a feedback loop. The best learning systems go one step further and experiment to engineer a feedback loop that can test and learn on an ongoing basis.
And finally, executing on these principles requires a critical ingredient. As the saying goes, what you cannot measure, you cannot improve. It is important to clearly and objectively define a measurement mechanism that can actually be computed using data. And then comes the harder part – the discipline of measuring the changes over time and see if the errors are on a descending gradient. In other words, measure if the system is actually learning and improving over time.
This list is obviously not exhaustive – it is a good starting point to help structure and build an execution framework that actively builds a learning mindset into the organization.
How does this last stack up with your organization’s learning system? Would love to hear views.
Fundamentals: Ten keys to Reality by Frank Wilzcek : highly recommended, especially if you want to re-visit what you learnt in Physics in high school