$($a$)$ How many parameters need to be trained between layer $2$ and layer $3,$ assuming that there are

no bias weights for the units in layer $3?$

$($b$)$ Suppose we change convolutional layer $3$ to a sparsely connected layer by getting rid of

parameter sharing, how many parameters need to be trained between layer $2$ and layer $3?$

Again, assume that there are no bias weights for the units in layer $3.$

$($c$)$ What is the size of layer $4$ if $4-$to$-1$ pooling $(22$ features become $1)$ is used?

$($d$)$ Assuming layers $8$ and $9$ are fully connected, how many parameters need to be trained between

layer $8$ and layer $9?$ Again, assuming no bias weights for the units in layer $9.$