A Study of Causality

http://www.causality.inf.ethz.ch/images/cause-effect.gif

For many generations it has been a goal of mankind to develop machines to the point where they exhibit human-like intelligence. The field of study tasked with achieving this objective has come to be known as Artificial Intelligence aka A.I. Lots of progress has been made in this regard: Machines can now aid in medical diagnosis (Andrew Ng’s CheXNet), drive automobiles (google, uber), communicate (almost seamlessly) with humans(alexa) and make paintings, to name a few.

Despite the mind boggling achievements in the field, we are still a long way from machines exhibiting human-like intelligence. In order to understand why this is the case we need to be specific about what human-like intelligence is. Judea Pearl does just this in his book, The Book of Why. He talks about three rungs humans have occupied during the process of evolution to get to the level of intelligence we currently possess.

1. SEEING: This is the ability to make observations about the world around you. Animals and humans alike possess this ability. The wildebeest of Africa observe the advent of the rainy season and this spurs the annual Great Migration. Humans here will observe the relationship between opinion polls and election results
2. DOING: This entails predicting the consequence of a deliberate alteration to your environment based on your observations from 1 above, assessing the options of deliberate alterations that will bring about the desired result and proceeding to carry out those alterations to your environment. Humans here will ask what the effect of increasing the advertising budget or altering the campaign schedule will have on the election results. The keyword here is deliberate. The action has to be deliberate with a desired outcome in mind.
3. IMAGINING: This is the point where one imagines states of the world that don’t exist and makes inferences as to why a change in one variable leads to a change in another. Questions here include: what would have happened to the election results had I not altered the campaign schedule or had I not changed the advertising budget? Was it actually the change in campaign schedule or advertising budget that caused the election results or were there other factors that weren’t taken into account?

It is this third level that distinguishes humans from all other animals: the ability to imagine alternate states of the world. For one to imagine alternate states of the world however, he or she has to have some hypotheses about the causal relationships between the various factors that make up the world under study. In order for one to say that a deliberate change in advertising budget caused the election results, he/she would have to imagine a world where the advertising budget remained unaltered and what the election results would be in that imaginary world. This is the essence of human intelligence. This is where machines need to get to before we can say they exhibit human-like intelligence.

The majority of achievements of artificial intelligence have all been at the first level: that of seeing (making associations between different variables). The reason is this: in order for machines to get to the third level (imagining) they must have the ability to make causal inferences about variables in a world. In order for that to happen the machines must be given a set of well structured mathematical instructions for arriving at these causal relationships. This is where the problem lies. For generations a mathematical language for expressing causal relationships was never developed. This is due to the prevailing school of thought that all information about the world can be gotten from data alone. It has now been shown that causal relationships cannot be inferred from data alone but the process that generated the data in question must also be taken into account.

In my next series of posts I will delve into the structured mathematical process for arriving at causal relationships and effects. I will be exploring topics such as confounders (the dreaded lurking variable), causal diagrams (gotten from the data generating process) and the “do” operator.