n Engineer is designing a deep learning system to locate people in a video stream. This is to be

installed in a car and will be part of the car$$s automatic emergency braking system. They are

considering using an Inception $($GoogLeNet$)$ based system to act as the backbone of the final

system.

Figure Q$1$ shows two types of Inception block.

i$)$ In the left$-$hand figure, explain how the Inception block works. In your answer describe what

process is performed in each of the coloured rectangles.

$[5$ marks$]$

ii$)$ The right$-$hand image in Figure Q$1$ has additional processes added. What are their purposes?

$[2$ marks$]$

iii$)$ This Inception model is good at classifying whole images. By itself, it cannot perform object

detection. Describe how you could extend this algorithm so that it could be used to locate

within an image:

$$ A car of known, fixed, size

$$ A car of unknown size.

$[5$ marks$]$

iv$)$ Explain how you could use the complete Inception network and build a system that could

recognise a face, given only one training image.

$[3$ marks$]$

b$)$

A neural network design has several $2$D convolutional layers. Each layer has a batch normalisation

applied to the inputs. A small section of the image is shown below. It is only $2$x$1$ pixels, with $1$

channel and a batch size of $3.$ The network has learnt the following parameters $\backslash$gamma $=1\mathrm{.}1,$ and $\backslash$beta $=0\mathrm{.}4.$

Also use $\backslash$epsi $=1$x$10-6$ or assume that it is very small. Calculate the input to the layer after the batc