Is a neural network sufficiently wide?? #6
Description
Hello,
In ALPHAGO ZERO the residual tower's conv layers contain 256x19x19=92416 values (many times larger than the game state) but only 589824 edges (+bias?)
In SPLENDOR which I am modeling as NON-SPATIAL, the redisual tower's fully connected layers currently contain 587 values (the size of a game state) and 344569 edges (+bias). Is this too few values? Increasing the width of the network drives the number of edges up into crazy territory pretty quickly.