Electronic music plays

Okay, hello, good evening everyone.

today we will continue with

discussing convolutional layers

and we will also turn to applying them to

handwritten digits recognition

and then I will teach you something about

so called auto enc격ers

which are an example of

of say unsupervised or self-supervised learning.

I thought that I start out by,

I thought that I start out by repeating what we did

in the last lecture about convolutional layers.

Now we have to wait for the projector to fire up.

So you remember that the idea had been,

we want to exploit translational invariance.

So if you have an image, then in order to recognize

whether, for example, there is a corner in that image,

it doesn't matter whether the corner sits somewhere

in the middle part of the image or near some

of the boundaries, it will always be a corner.

And in order to exploit translational invariance,

the idea then is to apply filters to the image

where the weights of the filter will be learned

by the network.

So this is depicted here.

You have two layers.

Each of them would represent an image, a 1D or 2D image.

And if you want to get the output values

for a given pixel, let's say, or a given neuron

in the upper layer, then it will depend only

on the values of the neuron in the lower layer.

And you will linearly superpose as usual,

using the weights between the networks.

But the idea is that you will use the exact same weights

in order to also calculate the output value

for this neuron and for that neuron and for the next neuron.

So you don't store different weights

for all the possible connections

between the different neurons and the layers,

but there's only one set of weights.

And this set of weights defines what we call

the kernel or the filter.

And in addition, typically, this filter is restricted in size.

So the spatial area, the region that contributes

to the output value is restricted.

In this case, it would be three pixels.

And so then the idea is to scan over the whole image

to always calculate the linearly superposed combination

of these three pixel values,