The Good and the Bad of AI

The Good and the Bad of AI

The Good and the Bad of AI

The internet likes lists, so here are five reasons to use Artificial Intelligence, five why you might want to avoid using it and five cautionary tales. For these examples it is assumed that AI is being used to create predictive models.

First the Good:

  1. Neural Networks (NN) are powerful. They tend to be more powerful than many of the classical methods such as Logistic Regression, especially where the selection of available variables is large.

 

  1. NN can be less affected by diminishing returns compared with classical models. With classical models such as Logistic Regression the majority of the variation is often explained by only a few variables and adding additional ones often does not add much more power to the model.

 

  1. NN are often better at finding latent patterns in the data than classical methods. The structure of the solution allows the NN to diverge from the main effect of the model.

 

  1. There are a number of free AI software solutions. Analysts can use R or Python to create viable solutions if you fancy developing your own models. There as also pre-packaged fully fledged AI solutions, but these can leave you with a couple of missing limbs.

 

  1. Neural Networks are used by real data analysts and data scientists. The successful use of artificial intelligence makes you attractive to talk to at dinner parties.

Artificial Intelligence

Artificial Intelligence

Next the not so Good:

  1. Artificial Intelligence is not intelligent. AI is good at finding patterns in the data, but just because there is a pattern does not mean that the solution is viable (e.g. Nicholas Cage films and deaths in swimming pools: http://www.tylervigen.com/spurious-correlations ).

 

  1. The models can be difficult to interpret. Classical methods tend to be better at displaying the solution in an understandable way; whereas, AI tend to be harder to understand due the conditionality between the variables and various node values.

 

  1. Neural Network models can be slower than classical methods. A simple Multi Layer Perceptron (MLP) needs a number of passes through the data to create a viable model, a logistic regression model can build a model much more quickly.

 

  1. Using the Artificial Intelligence Software solutions can be daunting. Getting to grips with the myriad of R AI libraries and often labyrinthine code can make one’s heart skip a beat (or multiple beats).

 

  1. They can be difficult to implement in third party software. It is fairly trivial to deploy a logistic regression model in a database and use it for selection purposes, it is far more difficult to do the same for an artificial intelligence solution with similar ease.

 

Finally The Ugly

  1. Offending software? In the film “Minority Report”, AI could be used to predict the chance that an alleged offender would commit a crime. Unfortunately there was racial bias and black offenders were marked as being at a higher risk of reoffending.

 

  1. We know who you are! It is possible to use AI to determine if a person is likely to be gay or criminal by examining their face. Facebook regularly gets confused when looking at photos of my son and myself and will often mistake my son for me (to his chagrin); he is in his twenties and I am a bearded old codger.

 

  1. Accident prone. There have been a number of deaths and accidents connected with AI controlled vehicles. The most recent of these connected with Uber, but this is not the only make that has had accidents. Maybe AI is too structured and can’t cope with the chaos that is humanity?

 

  1. It is what you say! Microsoft’s chatbot “Tay” had to be shut down after uttering racist, sexist and homophobic slurs. It was supposed to be based around a 13 year old teenage girl but within a day had turned into a “Hitler-loving, feminist-bashing troll”.

 

  1. Watch this space…….. There are many more to come.

That’s about it for now.

 

Leave a Comment

Enjoy this blog? Please spread the word :)