The Cola Scenario: 2-point pattern recognition

This article examines the common phenomenon of recognising patterns based on only 2 points of information. If you are a human you do this. And you'd benefit by being aware of it. 

Introduction

Pattern recognition is one of the most powerful cognitive behaviours demonstrated by homo-sapiens. It's the mechanism through which we categorise and make sense of the world we're in. It's what allows you to have a pretty good idea what the result would be of eating a platypus or cartwheeling off the top of a skyscraper, even though you've probably done neither of those things. And pattern recognition is the very basis of all the decisions we make — like why we consistently choose to keep our cartwheels near sea-level, and cook platypus meat before we eat it. 

Since pattern-recognition is so fundamental, what is the minimum input we can get away with? Is 2-points of data really enough? Let's find out! 

The following scenario involves a person acquiring a specific taste preference and using correlated data based on only two experiences to inform a decision which results in an unexpected and unwanted result. 

Cola Scenario

Step 1 
A person receives a glass of Coca-cola to drink for the first time, and enjoys the flavour. The drink has these properties: colour:black, texture:carbonated, temperature:chilled. Neurologically, an association is formed correlating the attributes of the drink, the pleasure derived from the taste, and the brand name. 

The data gained looks like this: 

name:Coca-Cola | object-type:beverage, sub-type:cola | purpose:drink | colour:black, temperature:chilled, texture:carbonated, taste:sweet | usage-result:pleasure 

In template form, this model is: 

NAME | PURPOSE | TYPE | CLASSIFICATION | ATTRIBUTES | USAGE RESULT 

Note this model is modular — additional data can be added to it at any point, e.g. OWNERSHIP owner:me | SOURCE source:vending machine | VALUE price:$1.00, production-cost:$0.10 | COMPONENTS ingredients:water, sugar, caramel, flavouring 456 |...and so on, as more data is collected. 
 
Step 2

Later, the same person receives a glass of Pepsi to drink and enjoys that drink too. The data set is affected, as the brand-name correlation has been expanded beyond Coca-cola to include Pepsi as well. The data set is now: 

name:Coca-Cola, Pepsi | object-type:beverage, sub-type:cola | purpose:drink | colour:black, temperature:chilled, texture:carbonated, taste:sweet | usage-result:pleasure 

This is where something subtle but significant happens. In order to retain the information as a pattern, rather than as knowledge about a particular subject, the data set is collapsed into a simplified form which is easier to retain by dislodging the NAME data: 

object-type:beverage, sub-type:cola | purpose:drink | colour:black, temperature:chilled, texture:carbonated, taste:sweet | usage-result:pleasure 

This enables this set of ATTRIBUTES to be used as a set. This means that if an object encountered in future contains this group of attributes then it fits the CATEGORY for cola, and can be identified as such. 

Step 3
On another occasion the subject sees a chilled bottle of black, carbonated liquid and buys it to drink. Using the data obtained by past experience, our subject observes that the drink is black, cold, and fizzy, reasoning from that observation that the drink is cola and will taste similar to the similar beverages enjoyed previously. (This assumption could even have prompted the purchase, which is an interesting additional point: we make financial decisions with incomplete data: assumptions). Note that our subject has omitted the NAME attribute (i.e. this drink is not called "Coca-cola" or "Pepsi"), and is basing the decision on a broader set of data -- the actual attributes of the drink, correlated with the pleasurable USAGE RESULT. 

On tasting the drink, the subject determines that not only is the expected sweetness attribute not present, the drink actually tastes bitter. The expected pleasurable USAGE RESULT doesn't occur! Because this is a bottle of chilled, carbonated coffee. To make sense of this, a new data set is created to cater to the new CLASSIFICATION: 

object-type:beverage, sub-type:cola | purpose:drink | colour:black, temperature:chilled, texture:carbonated, taste:sweet | usage-result:pleasure 

object-type:beverage, sub-type:coffee | purpose:drink | colour:black, temperature:chilled, texture:carbonated, taste:bitter | usage-result:displeasure 

Summary

Based on 2 experiences the subject formed an incomplete knowledge-pattern. Assuming the knowledge-pattern to be adequate for decision-making, the result was an unexpected and unwanted result (a bitter taste instead of sweet) and some wasted money. 

Key findings

1. Even though the result was unwanted, the subject's logic was legitimate. 
2. The human's assumption that the knowledge-pattern was complete was reasonable: without data they didn't have, the existence of non-cola drinks with the black, chilled, carbonated attributes couldn't have been foreseen. 
3. The 2-point pattern recognition phenomenon is not inherently flawed: all knowledge-patterns must be formed one datum at a time. An incomplete data-set is not necessarily a wrong data set. 
4. The value of 2-point pattern recognition is that it provides a natural framework for experimentation. 
5. The problem with knowledge-patterns based on 2 points of data is the overemphasis humans habitually give to them, forming entrenched convictions that we later can't change. It's only by being able to update or replace these patterns with new, better-informed patterns whenever new information is encountered, that we gain any value. In Step 3 the subject achieved this – but it's equally possible that they may have rejected the new information. This typically happens when the effort of assimilating new information outweighs the reward of doing so. 
6. Another (lesser) danger of 2-point pattern recognition is the risk of unexpected unwanted results when basing decisions on very little information, as in the Cola Scenario. The risks of making decisions on incomplete data will be the topic of a subsequent blog post. 

But what's the answer to the question "is 2 points of data ENOUGH?"
The answer is yes.
Two points of information is a legitimate basis for forming a knowledge pattern. But such a pattern will not be perfect. It must be continuously updated when new information is gained – failure to do so will result in irrational beliefs. If in The Cola Scenario, the subject had not updated the data set after discovering chilled coffee because they already had formed the belief that all drinks with the black, chilled, carbonated attributes were cola, they would continue making the same mistake and receiving the same unexpected, unwanted result with chilled coffee. Therein lies the danger of forming a belief or full-blown conviction based on a 2-point pattern (rather than simply forming a more malleable idea, notion, opinion, or hypothesis). 

Meta-cognition exercise

1. Think of a time you recognised a pattern based on 2 points of data. How did it benefit you?
    
2. Will you be able to recognise it happening the next time your brain does it? 
   ​
   
3. If you do recognise it as it happens, will you re-evaluate the strength of the correlation between the shared attributes and the conclusion you derive from it? 

The exercise above is a simple type of meta-cognition (being aware of your thoughts as you think them, and evaluating their logic). Given how interesting our thoughts are, their real-time analysis is often easy to overlook! But even without meta-cognition, if you are simply willing to update your understanding of a thing when faced with new information about it, you will avoid the danger of forming beliefs and convictions that cause you to repeat the same mistakes indefinitely. 

Thanks for learning! 
Michael