One thing I don't understand is why it would be smarter to augment the inputs with some of the missing board information - particularly the availability of castling. Even though this network is a quick-and-dirty evaluation, seems like there's room for a few additional bits in the input and being able to castle very much changes the evaluation of some common positions.
This network is just to signal to the final search algorithm how good the board looks in a given spot on tree of possible moves.
A few examples come to mind where, for instance, white has a bishop and knight attacking f7 and it's black to move. If black castles nothing interesting happens and white is likely overextended. But if black cannot castle white will win a rook.
I can make up a similar situation with en passant, and this does come up, but a lot less frequently.
So anyway, surprised you wouldn't toss a few bits in there, at least the 2 for ability to castle by each player.