I speculate that there should exist recognition beyond 100%.
Take free will for instance. Since you can classify simple robotic behavior from living things. You can recognize free will.
So the idea is that life started with more robotic behavior and very little free will. This because very little free will is easier to achieve through random events.
The super recognition is then needed because 100% recognition of the early free will is still very crude.
So inspired by this I speculate that we need a ”evolution” score together with the recognition percentage of 0..100%
Then setting the target of a generative process to maximize the score and updating the recognition would give you more than 100% recognition. Compared to the orginal data.
So maybe there should exist something like unsupervised ranking. That is. Putting objects with nearly the same quality into the same group. Here the algorithm could perhaps use two objects with a predefined ranking. An easy problem would be if the two objects represented the highest and the lowest score. Placing the rest into n groups should then be a lot easier.
Then with super recognition you could perhaps generate music that sounds better than the input data. So I speculate that there should exist something like super recognition so that you always can say that object A is better than object B.
I wonder if it could work by enhancing the weight matrix. If the weights are turned into parameters. That is. The parameters are data for a smooth 2D image. The image is then used as data for the actual calculating weight matrix. Much bigger. Then the image could follow ?simple transformation when more input features are added. Maybe then it could be enhanced with post processing filters at same time. Improving the filters would then also improve the network. Filters could include noise reduction, detail or sharpness.