a$)$

$1.$ describe the overall iterative learning process for artificial neural networks discussing, among the other the things, the chain rule $($e$.$g$.$ how it works$),$ what it entails and what possible risks are involved;

$2.$ given the following definition of a convolutional neural network $($comments about each line of Python code highlighted in yellow before the line$),$

model $=$ Sequential$()$

#input as a $(1$x$28$x$28)$ tensor, $24(3$x$3)$ kernels, stride$=1,$ padding$=$"same" to keep input size, relu activation model.add$($Conv$2$D $(24,$ kernel$\_$size$=(3,3),$ padding$=$"same", activation$=$"relu", input$\_$shape$=(28,28,1))$

# max pooling with pooling size $(2$x$2)$ applied as in the lecture's example model.add$($MaxPooling$2$D $($pool$\_$size$=(2,2)))$

#input as previous layer's feature maps, no padded convolution here model.add$($Conv$2$D $(48,$ kernel$\_$size$=(3,3),$ padding$=$"valid", activation$=$"relu"$))$

# max pooling with pooling size $(2$x$2)$ applied as in the lecture's example model.add$($MaxPooling$2$D $($pool$\_$size$=(2,2)))$

# input as previous layer's feature maps, see description in the first layer for the rest model.add$($Conv$2$D $(64,$ kernel$\_$size$=(3,3),$ padding$=$"same", activation$=$"relu"$))$

#max pooling with pooling size $(2$x$2)$ applied as in the lecture's example model.add$($MaxPooling$2$D $($pool$\_$size$=(2,2)))$

#input as previous layer's feature maps, no impact on the number of the network's parameters model.add$($Flatten$())$

# $128-$wide dense layer with input as flattened vector from previous step model.add$($Dense $(128,$ activation$=$"relu"$))$

# no impact on the number of the network's parameters

model.add$($Dropout $(0\mathrm{.}25))$

# $10-$wide output layer with input as hidden vector from previous step model.add$($Dense $(10,$ activation$=$"softmax"$))$

# no impact on the number of the network's parameters

model.compile$($loss$=$"categorical$\_$crossentropy", optimizer$=$"adam", metrics$=["$accuracy$"])$

describe the network in terms of its depth, its layers' width, what it is expected to eventually calculate and say how many parameters are trainable in total elaborating on how you calculate that number.