Brain System Theory

Artificial Intuition

Artificial Intuition

The most prominent example of Artificial Intelligence is probably the machine learning. It is based on a theory of artificial neural networks - ANN. They are loosely inspired by the network structure of the animal brain. The word “intelligence” though, is misleading. The aim of the following article is to make a point that machine learning is descendant more of human intuition than human intelligence. The similarities between the ANN and human intuition can shed a bit of light to understand our unconscious behavior.
We start with the description of what the artificial neural network is. To make it straight, the ANN is nothing more then a mathematical function that comprises tens or hundreds of simple functions like ‘Y = A X + B’. In the terminology of artificial neural networks, these are called nodes. Output of one function is used as an input to the next one. Each node collects input not only from the previous one, rather it collects inputs from multiple ones. The function of such node looks like the following one: ‘Y = A1 X1 + A2 * X2 +… + B’. This chaining of the functions forms a web of functions - nodes.
The A1, A2 constants are weights describing how much each of the previous output influence the current node. The goal of the ANN is to adjust these weights by training. Data scientists give the neural network samples of data. The training algorithm then adjust the weights by iterating through each of the data example. After the training, the ANN is ready to solve the problems.
This is the core structure of artificial neural networks.
The design is both impressive and clever. The learning is a purely mechanical process. It requires no human interaction, and it involves no intellectual decision. Because of the mechanical nature of the learning, we can create complex meshes of nodes with a virtually unlimited number of weights. The algorithm used to update the weights works the same way, regardless of the number of nodes forming the network. Since learning is a time-consuming process, the only real limiting factor of training the ANN is time. It takes several hours or days to train the simple neural network. 

The quality of the output of the ANN depends on a number of nodes in the artificial neural network and the number of training samples. By increasing each of them, we can in theory increase the precision of the ANN. This of course proportionally increases time to train the neural network.

There is no more to ANN. These are the fundamental concepts behind the artificial neural networks.
The knowledge of the ANN sits in the adjusted weights which come with a result. We cannot explain the reasons that drive solutions made by the ANN. We cannot answer the simple question of “WHY?” just by looking into the weight map of the ANN. There is no way we can validate the outcomes of the ANN. The trust is the only tool we have when dealing with the decisions made by the ANN’s.

This is the point where human intelligence differs from the ANN. We base our knowledge on the ability to explain our decisions and answer a series of “why?” questions. We build the trust in the society on predictability. The ability to explain and understand is the prerequisite of predictability. Our knowledge is descriptive. The way we approach the knowledge allows us to search for reasons and intentions. Our intelligence lives within our consciousness. It is the explicit intelligence.

There is however another, more powerful quality that drives our lives. It is the unconsciousness. It is helpful when we choose our life partners, or look for our children. The unconscious behavior helps us to ride a bike without paying explicit attention to the road. Some abilities like bike riding we learn consciously, intentionally. Yet much more of our knowledge is comes from the experience. Unfortunately, in such a case we can rarely explain our decisions. In fact, many of our decisions are unconscious. We make them intuitively and later we rationalize them. This kind of intelligence is woven in between the neurons of our brain. This is unconscious - implicit intelligence; the intuition. 

The intuition is the quality of the human we often devalue. Like in case of ANN, the intuition does not give us responses. It is self-asserting. It does not posses the quality we need to support our decisions. We cannot validate our intuitive decisions.  Should we then not listen to our impulses? 

We understand how easily we can misrepresent our intuition. Someone who grows up in a toxic or abusive environment is a sad example of a human with a distorted perception of reality. We manipulate our perception just by getting unbalanced information or omitting important facts. Sometimes our unconsciousness is skewed to the extent we need to search for a qualified therapist. The therapist is a person with whom tries to compensate our biases by offering the alternative view on our life issues. The therapist is there to helps to heal our unconscious biases.

But what about ANN? Can the ANN have biases?

In 2015 Amazon had to shut down the experiment to search for candidates by using the ANN to scan through candidates CVs. The algorithm the Amazon used, penalized women candidates. They trained the ANN by using the male candidate profiles, as most of the candidates in the past were males. 

Now imagine what would be our reaction if they would train the network with a balanced set of CV’s and the outcome would be the same. Would we accept and respect the mathematical precision of the choice made by the ANN?

It looks like the ANN can be prone to biases much like the humans are. The ANN exhibits the same issues as human unconsciousness. It can have biases originating from the quality of training dataset we create. 

Compared to human unconsciousness, the ANN is primitive. We will eventually build complex ANN’s and train them with detailed information. With the progress of the ANN’s comes in hand collecting of enormous amounts of data about many aspects of our lives. We build a network of sensors that collect data we were not interested in the past. The more complex ANN’s we built and the more data use for training the ANN, the more complicated will be to change the decision made by the ANN.

While the ANN’s are simple, dropping biased ANN is cheap. Yet as will the complexity of the ANN grow, and we will use giant datasets to train the ANN, dropping a trained ANN will not be a viable option. We like zero down-times of our systems. There will be no time to train the ANN from the scratch. The only possibility in such a case would be to use compensation data samples prepared by the ANN data specialist. And as crazy as it sounds this is the role of the ANN therapist.

The convergent evolution, illustrates why the evolution of different species creates  analogous features. We should pay close attention to similarities between the live organism and the way IT is evolving. These two areas have more in common that we think.