Teaching Robots 101

https://www.youtube.com/watch?v=nuSkcLY0YQg

Supposing that we will be the creators of superintelligence, we also have to act as instructors, as teachers. In my mind the training of robots should not be much different than the one already being used in humans. It seems rather realistic to me that they will have to learn through interaction with others and with the environment. What seems rather unrealistic is the existence of a school or a university for robots. We will have to use our creativity on this matter.

The truth is we have already found some ways of teaching them. There is the backpropagation algorithm, mainly used in artificial neural networks. Each neural network is composed by nodes, which act like neurons. Each node is connected to other nodes and this connection is characterized by its weight. The bigger the weight the stronger the connection is between two nodes. The way this algorithm works is by calculating and adjusting the weights backwards (the whole procedure is executed and if the result is not the desirable then changes are being made) in order to improve the network until it is able to perform the task for which it is being trained. The problems with this method are that it is a form of supervised learning, meaning that we have to categorize the variables we use as input data and the fact that it does not scale well in bigger networks. It works really well with small networks but the time it takes to train a network tends to increase exponentially with the number of neurons it contains.

Another learning mode is the Hebbian algorithm. It could described by the following sentence : ‘neurons that fire together wire together’. The more often two neurons are activated together, the stronger the relationship between them. It doesn’t require supervision and it operates well in larger networks but it has yet to be explained how it could produce by itself all the forms of learning and adaptation the human mind is capable of doing.

Here is where the interesting part begins. Maybe we don’t have to understand exactly the way we will be teaching them. After all we have not yet fully understood how our brain is able to store so many information and process them again and again. It just does it. And we are fine with that. We keep teaching and learning. The important thing is the result. What if the same could happen with robots? What if we could somehow teach them and start seeing proof that they are learning? Well, guess what. This method does already exist and it is called machine learning. Named originally by Arthur Samuel in 1959 this algorithm does not contain explicit instructions on how the desired result should be obtained. It relies on patterns and inference. It uses some input data, known as training data, in order to make predictions or decisions without being explicitly programmed to perform the task. It is like a black box where no one really knows what happens, but it has results.

So in the end, we might have been using the wrong approach since the beginning and perhaps we have just started following the right direction. Which means we might be much closer to the Singularity than we originally thought. However is it only me that gets a bit scared of the fact that they have the ability of learning without us understanding how? It deprives us from the ability to intervene at any point. Let's hope that it is for the best. 

PS : Lately I have been reading many articles about artificial intelligence, thanks to this blog. The majority of these articles are 10 to 20 years old. So it is really interesting and funny at the same time reading older predictions and realizing how far from the present reality they were. It makes you feel that you have the right to mock the author since he is proposing something so audacious. Of course we get to answer from a safe spot because we already know how the story is going to unfold.

Aldo do you know the future? If so please correct me, so that I don’t look like a fool in a couple of years.

‘Yes but I won’t tell you, just to keep you guessing. I too enjoy reading that this or that will happen in the next 30 years. Always this number, as if there is no other one.’