Convolution layer(s) with fewer parameters

[accosmin]

Reduce the number of parameters from O(#outputs \times #inputs) to O(#outputs).

Some ideas:

• random weighted convolutions: O(o) = sum(i, w_o,i * (C_o @ I_i)), w_o,i being fixed
• weighted convolutions: same as above, but the weights should be some normalized versions of parameters (e.g. w_o,i = x_o,i / sqrt(1 + x_o,i * x_o,i))

[accosmin]

Also new variation of the linear layer: use normalized weights for parameters (like described above)