Chess NNs In Humans And Programs

[towforce]

In my mind, I see shapes on the board where pieces could move to.

In Leela's NNs, I'm guessing (please correct me if I'm wrong) that there's no concept of straight line movement, diagonal movement, or, indeed, any movement at all. I think that all Leela's NNs know about is patterns of pieces relative to each other and the edges of the board.

In this sense, although Leela has NNs, it doesn't really look at a position in the same way a human does.

Per any NN, the neuron and connection (synapse) map and weightings could be translated into a (big) numerical expression.

[Daniel Shawul]

There is the same concept of shapes, straight/diagonal movement in NNs too.
The way residual neural networks work is that the top layers figure out basic shapes and movements and the following
layers figure out increasingly complex concepts. I am sure I have seen visualization of a chess ResNet that shows a heat map
of rook/bishop movements. In other image recognition to, e.g. human face, first lines and curves defining a face are understood,
then eyes and noes, then a face, then a particular face of someone.

[towforce]

Very interesting!

One of the biggest problems for a human is that the huge cerebrum, which is much bigger than in other animals, slows everything down. This is thought to be why chimps consistently beat us easily at video games that they are able to learn.

In terms of adaptation, the cerebrum gives us many layers of abstract thought, which can be used in multiple fields of expertise. At the moment, this is giving us an adaptive advantage over other animals (combined with other features), but it does mean that we're slower on particular tasks that other animals can manage.

Relating this to chess, it means that the human NN is "a bit of a mess", though when you reach GM level of play, you will have developed some rapid chess processing paths that will guide you to good moves (and help you avoid bad ones) much more quickly than a beginner could. So it might be fair to say that a human GM chess brain is more like Leela's NNs than a beginner's chess brain is.