One of my ideas about the ARC challenge

The ARC challenge offers a new way to measure intelligence and to be more specific the abstraction side of the intelligence.

My approach is the following. I think that the transformations needed to solve the whole arc challenge are some what simple and numerous. But how I can discover them ?

I will use the idea from my previous blog post and try to “decompose” the transformations with an Auto Encoder and use the context layer in order to get some information on the functions there and after that to use this information on genetic evolution algorithm in order to get(guess?) the correct function.My idea is that if I process the input and output with an auto encoder I can get the “finger print” of the functions that have been used in the abstract transformation and with a little search I can find the correct function.And since a picture is worth a thousand words here is my arch :

Here we have and auto encoder(variational autoencoder) and when it is trained we can use the context layer in order to feed some seed (when it is passed trough some sort of data manipulation like neural network) and feed it to the code generator(most likely some system that will use genetic evolution algorithm and it will use the input and output data for validation of the algorithm). I am not sure about how the input and output data for the auto encoder will be formatted but I think that it will be a good idea to feed the auto encoder with the input and the output data together concatenated