Predictive Processes and Tools
How to use our mind's prediction machinery to help your teams navigate the world
There’s nothing quite like a public humiliation in a sporting context to teach you about the brain. It was 2005, and I was 15 a fielder in a practice cricket match at Finchley Cricket Club. The batsman played a gorgeous cover drive which went just wide of me and sped away in the direction of the boundary rope. ‘I can get this’, I thought. ‘I can run really fast. They’ll see. Everyone will be so impressed’.
So I ran at full pelt with my eye on the ball, aiming to reach the ball just before the boundary rope, and therefore saving the day. I reached down, imagining the feeling of the ball in my hand and the satisfaction at catching the ball in time.
I heard the collective groan of the other players before I felt the impact of the metal on my face. I had run headfirst into the metal pole supporting the practice nets on the edge of the field. As I sat with my mangled glasses, wounded pride and a fresh cut in my nose I wondered, what had gone wrong?
From a cognition point of view (and a wounded pride point of view) there is a major question - how did I end up running full speed into a metal pole?
In order to understand this, I’d like to introduce the pre-eminently powerful principle of Predictive Processing, please. In order to put it in context however, we need to start a little further back.
The mind is a camera (it isn’t)
Maybe you think of the mind as a kind of camera. The world is out there, I am in here (inside my skin or skull) and the role of cognition is to use the senses to perceive the stuff out there in different ways and pass the information to my brain. The brain then can process the world and build a kind of image or representation of the sensory information which I use to live and thrive.
Just as the lens of the camera focuses the information about to become a photo onto the film or the camera’s computer, our senses bring information about the world onto our computing brain.
This is what we might call a bottom up process, where in a hierarchy of processes the basic sensory ones sit at the bottom and more complex and abstract processes sit at the top. Sensory information (at the bottom) passively moves up the hierarchy towards complex computational processes (at the top).
This approach comes with problems. First of all, taking all that rich information about the world and processing it all is very energetically intensive - it puts lots of demands on the brain. Secondly, it doesn’t explain why our senses so often get it wrong. This is an oversimplification of the issue, but it sets us up to understand the paradigm that Predictive Processing came to replace.
Quite a Predict-ament
Professor Andy Clark (who has dogs called Borat and Bruno) from the University of Edinburgh offers an alternative which has fast become the dominant integrated theory of how the mind works. It’s being proven and tested in Psychology, Neuroscience, AI and beyond. It goes like this and it’s beautifully simple.
When we interact with the world we don’t start as a blank slate, we go in with a model of how we think it should work. This model informs a prediction about what we expect to find. Our senses then sample the world, and if we experience something unexpected - prediction error - we change the model. Our sensing and acting behaviour then is all geared towards reducing this prediction error so that our model is as accurate as it needs to be.
This model and prediction mechanism allows us to make fast and useful judgements about the world, safe in the knowledge that we’re able to change our model quickly if needed.
For example:
I see a cup of coffee I’ve just made on the table.
Prediction: This cup of coffee is hot and would be nice to drink.
Action: I pick it up.
Sensing: It’s hot on my fingertips.
And then again:
Prediction: This hot coffee I’m holding would be nice to drink.
Action: Put it in my mouth.
Sensing: mmmm, nice hot coffee.
Another phenomenon that Predictive Processing can help us explain is when we sometimes ‘ignore’ a sensory experience. For example, sometimes I leave my coffee out for an hour and it goes cold without me realising. I might pick it up, my fingers sense cold, but I don’t realise and I still believe that it’s a nice hot drink. It’s only when it gets to my mouth that my mind catches up with what my fingers first experienced. In retrospect I can recall that my fingers sensed it was cold, causing me to feel silly.
According to Predictive Processing, we unconsciously prioritise some sensory experiences over others. We pay more attention to the senses that we think are relevant to testing the model. We don’t know what these senses might be, but that’s something our brain learns the more we live our life.
So in the ‘coffee gone cold’ example, it would go like this:
Prediction: This cup has hot coffee in it. That would probably be nice to drink.
Action: I pick up the cup.
Sensing: The cup is cold to my fingers. This isn’t in line with my model, I wasn’t expecting it and it’s not freezing to my fingers, so I down-weight this sensory experience and don’t experience prediction error.
Leading to the next action:
Prediction: This cup I’m now holding has hot coffee in it, that would be nice to drink.
Action: I move it to my mouth.
Sensing: Ok, that’s some cold coffee. Nasty.
I experience a prediction error which causes me to update my model and now I believe that the coffee is cold and I probably wont enjoy it any more, so I can microwave it or throw it away. In this case, I ‘weight’ the sensing that takes place in my mouth more than that in my fingers, causing me to make the grave error of trying to drink cold coffee.
As Andy Clark explains in The Experience Machine, predictions allow for very efficient operating. He uses an example: I ask you to pick me up from the airport next week unless you hear from me. For the whole next week if you don’t hear from me, that’s actually a signal saying ‘please pick me up from the airport’, I don’t need to neurotically text you every day to remind you. Hopefully.
A similar concept exists in image compression. I don’t need to share the colour of every pixel in an image if I can just send one colour representing a whole area, and only tell you about the pixels where the colour changes. Using this method you can infer using predictions where the colours should be, and I can transmit a rich amount of information without actually having to pass on data about every pixel.
To bring it back to Predictive Processing, my models are not ‘correct’ per se, but they are mostly useful enough and they are very fast and efficient. I don’t need to be constantly doubting my reality and gathering rich data if I can just rely on my initial models and trust that prediction error will cause me to course correct if I’ve got it wrong. To quote Clark again:
Cheap, fast, world-exploiting action, rather than the pursuit of truth, optimality, or deductive inference, is now the key organizing principle.
I love this explanation because it also explains our errors of rational judgement in a way that doesn’t treat them as if they were a malfunction. If you’ve read Thinking Fast and Slow, or anything about how irrational we can be, you’ll know that we often make errors of judgement using confirmation bias, or availability heuristic. According to Predictive Processing, these kinds of error are a feature rather than a bug in the system - they are the price we pay for being able to usefully think fast, and are overall efficient and effective for survival.
Let’s go back to Finchley Cricket Club.
Prediction: It’s safe for me to chase this ball, lean down, pick it up, and throw it. If I do that I’ll be a hero.
Action: CHASE THE BALL.
Sensing: ouch ouch ouch, that was not a safe path and I’m no hero.
Prediction error led me to realise that my model was incorrect in an important way.
What does Predictive Processing mean for teams?
Agile On The Mind is in the business of taking principles from Cognitive Science and seeing what happens if we apply them at a team or organisational level. If we imagine a product design paradigm based on Bottom Up sensing we might imagine that it goes like this (starting at the bottom):
You start with a user whom we assume to be out there in the world, learnable about and static. We go and do some user research, model the specifications, and then design a really good solution and build it. This would suffer from the same problems as the Bottom Up processing we mentioned earlier; it’s slow, it doesn’t deal well with evolving user needs and it’s energetically costly to capture all the information you would need about the user. Not a great paradigm for success.
What would it look like if we based the team’s explorations on a Predictive Processing model? It might look more like this (starting at the top):
We start with a model of how this idea/product/feature might work. We then do some actions that allow us to test this model, such as building a prototype, a wireframe, or even a live product. Sensing comes in the form of usability testing with real life users, or through releasing your idea into the real world and keeping track of what happens.
This allows you to experience prediction error if the users don’t use the design in the way you expect or don’t respond to your product launch in the way you expect. If there’s prediction error, the team can then update their model about reality and try again.
This is what Scrum and agility are really about. We can map the sequence above onto the Scrum Ceremonies.
Prediction: At Sprint Planning the team aligns on the current situation and sets a goal based on a prediction about reality. For example, if the sprint goal is ‘release feature x to all users’ we are making a prediction that a) releasing it is technically possible, and b) the feature will result in the desired outcomes, such as increased revenue, or usage.
Action: At Sprint Planning the team coordinates the plan for how to check that model and prediction in reality. At their daily scrum and throughout the Sprint they do the action (and go through smaller Prediction, Action, Sensing loops each day).
Sensing: At the Sprint Review we look at the finished Increment, whatever it may be, and assess whether there is a prediction error or not.
Thinking about Scrum like this can help us move our teams away from thoughtless ‘Zombie Scrum’, in which teams participate in the ceremonies, but don’t inspect and adapt their approach to building the product.
In the day to day cut and thrust of work though this can easily be lost. Often when I’m working with teams they follow a user-centred design process, but they don’t bring it back to the team’s shared model about their reality. I notice this kind of thing when User Experience folks would rather do all the user testing themselves and summarise their findings in a neat powerpoint than find a way to invite the team to participate in user testing themselves, and thereby really understand the user first hand.
Similarly, if the Sprint Backlog is just a mish-mash of tickets that don’t relate to each other or a higher goal, or nobody understands what the cost is of missing the Sprint Goal, that’s a useful diagnostic that the team isn’t leveraging Scrum to help them have better cognition.
If the team isn’t getting this higher level cognition right about learning and navigating their world, then fiddling with processes and details at the ‘lower’ levels won’t be the best way to improve the situation.
We use Predictive Processing unconsciously and constantly. Our team members as individuals do as well. At the whole team level however, it’s not a given that we will choose the most rational and effective cognitive strategies. Unless Predictive Processing is built into the habits, conversations, artefacts and ceremonies - the culture - of our workplaces, it won’t happen.
So that’s the opportunity for you. If you get it right, you can help a team think better and navigate their shared world more effectively and efficiently. Next time on Agile On The Mind I’ll be sharing tools and patterns you can use to help move your team more towards a Predictive Processing cognitive model.