Unit 7.2 How Convolutional Neural Networks Work

Slides

Part 2: Convolutional Layers
Part 3: Convolutions With Multiple Channels
Part 4: What Are Pooling Layers?
Part 5: Controlling The Output Size With Padding

References

torch.nn.Conv2d documentation
Striving for Simplicity: The All Convolutional Net, https://arxiv.org/abs/1412.6806

What we covered in this video lecture

The architecture design of (convolutional neural networks (CNNs) helps these models to take pixel locality into account, and it offers a certain degree of robustness towards image perturbations, translations, and rotations. In a sense, we can think of CNNs as a network consisting of convolutional layers for representation learning as well as fully connected (multilayer perceptron) layers as classifiers on top of the learned representations. Moreover, we also learned about how convolutional layers work in detail, including handling multiple input and output channels.

Log in or create a free Lightning.ai account to access:

Quizzes
Completion badges
Progress tracking
Additional downloadable content
Additional AI education resources
Notifications when new units are released
Free cloud computing credits

Quiz: 7.2 How Convolutional Neural Networks Work - Part 1

How many parameters does a 2×2 maxpooling layer have?

Impossible to say. It depends on the stride.

Incorrect. There are no parameters in maxpooling layers.

Correct. There are no parameters in maxpooling layers.

Incorrect. There are no parameters in maxpooling layers.

Please answer all questions to proceed.

Quiz: 7.2 How Convolutional Neural Networks Work - Part 2

Suppose we have a 3×3 filter that we slide over an 12×12 dimensional image (with one input and one output channel), how many weight parameters do we need for that kernel (/filter)?

Incorrect. Note that we reuse the same 3×3 filter in all parts of the image.

Correct. We reuse the same 3×3 filter in all parts of the image.

Incorrect. Note that we reuse the same 3×3 filter in all parts of the image.

144

Incorrect. Note that we reuse the same 3×3 filter in all parts of the image.

Please answer all questions to proceed.

Quiz: 7.2 How Convolutional Neural Networks Work - Part 3

How many bias units does a 3×3 convolutional kernel with 3 input and 5 output channels have?

None

Incorrect. There are definitely some bias values associated with the kernel.

Correct. There is 1 bias unit per output channel.

Incorrect. Hint: Implement the layer (nn = torch.nn.Conv2d(...)) and then use nn.bias instead of nn.weight.

135

Incorrect. That would be the number of weights.

Please answer all questions to proceed.

Quiz: 7.2 How Convolutional Neural Networks Work - Part 4

Suppose we want to apply a maxpooling layer to an image with 3 color channels (and we want 6 6 output channels in the feature map). How many channels does the maxpooling layer need?

None

Correct. Maxpooling does not have any channels.

Incorrect. Maxpooling does not have any channels.