There is a lot of buzz around AI (Artificial Intelligence). People speak about
AI and suggest all kind of areas or problems to which AI will be applied to in
the future. However, is this realistic? Marketing campaigns add to this believe,
because many tools or systems now claim to have AI. But what does that mean?
Where does the AI come in?
Personally I believe that AI will help us solve a lot of problems that we humans
otherwise could not find a solution to. The reason for that is basically the
Moravec’s paradox. This paradox simply says that (1) the computer is good at
things, we are not good at, e.g. massive computations or repetitive tasks. If
further says that (2) this is also true vice versa, that we are good at what the
computer is not good at, e.g. creative problem-solving. In terms of pattern
recognition at the moment it is probably a draw and it depends on the area of
application. So AI will definitely have a place in areas that need, require or
enable massive computations.
But What Is AI?
But what is AI? AI can be seen as a set of methodologies, that perform all kinds
of cognitive tasks and then make appropriate decisions based on them. There are
many solutions in the area of AI but a very prominent one is a neural network.
You can see an
example of a neural network
in the picture above. This is a screenshot from a neural network that was built
with Tensorflow. Tensorflow is open
source software and can be found on
GitHub. That means
the software can be downloaded by everybody for building neural networks.
Tensorflow is one of the most prominent AI solutions in the world and often also
the basis for other project such as natural language processing. But of course
there are other renowned toolkits for building neural networks such as
It is easy to spot that the neural network above looks similar to a regression
model from statistics with some exogeneous and some endogeneous variables — only
with many more layers in between. In my PhD thesis from 1997 about
Statistical Neural Networks
I have, indeed, shown, that neural networks are similar to nonlinear
nonparametric regression models.
So is the application also similar? Just from looking at the neural network
above you can easily understand the following points:
In order for a neural network to give an answer to a difficult question you
need to exactly specify this question. Otherwise the neural network cannot
produce a fitting answer.
You need to train the network for producing an answer. Train means to
determine the strength of the connections between the individual neurons. For
this you need to have a lot of data that fit to the question. It is obvious,
the more difficult the question and the more complex the network is (i.e. the
more layers and neurons it has) the more data you need. Also the data need to
be of good quality and cleaned for outliers or missing values. Otherwise the
result may be distorted.
Building and dimensioning the neural network for the question is a lot of
work. Finding, cleaning and storing the data is also a lot of work. It
requires a lot of expertise, technical knowledge and IT infrastructure.
… And You May Get an Answer
At the moment artificial intelligence can only answer the specific questions it
has been trained for. Artificial intelligence therefore will not replace human
intelligence any time soon when it comes down to understanding, emotional
intelligence, critical thinking or problem-solving.
But two more things are also clear: (1) AI can or will only be applied to
questions that have massive amounts of relevant data. Neural networks have
beaten experts in games such as backgammon because it is easy to produce any
amount of training, test and validation data. The same is true for all kinds of
pattern recognition. (2) AI will with a high priority be applied in areas that
do have large scale effects so that an efficiency gain can justify the high
investments necessary in knowledge, skills, data and IT-infrastructure.
Understand the Buzz
So, if you come across the buzz word AI in ads or in marketing campaings or if
you wonder whether a certain application of AI is reasonable you may want to
consider the following points:
- Which question should an artificial intelligence or a neural network exactly
- Is this a widespread question and is the answer widely applicable? Will the
answer differ depending on the domain (country, language, segment, legal or
regulatory issues, etc.)
- Where does the necessary data come from? If the answer is different per
domain, is there enough domain specific data?
- Is it realistic to get this data?
- In which database are the data going to be administered? Is it mainly
historical data or is it a frequent flow of data?
- Who collects the data and where is it collected?
- Who controls and programs the artificial intelligence?
- Who monitors the artificial intelligence and carries out the quality
- Does the necessary knowledge exist?
- Can the inputs, costs and efforts be justified by the expected output, income
- Is it realistic that the efforts will be made or will the knowledge rather be
applied in a different area?