Which statement is true?Group of answer choicesConvolutional neural networks tend to only have 10 layers.If the feature response value is large, then that portion of the image does not look like the filter.The convolutional kernel can only take all-positive or all-negative values at a time.If there is no pattern to find, the feature response is a weighted average with the result of blurring the image.

Which is false?Group of answer choicesOutput of a layer in the residual network is the normal output (i.e, what’s produced after applying a filter and an activation function) + the layers input.Output of a layer in the dense network is the normal output (i.e, what’s produced after applying a filter and an activation function) + the layers input.Output of layer 𝑙 in the dense network will be one of the inputs for layers 𝑖∈(𝑙+1,𝐿) where 𝐿 is the total number of layers.A problem of convolution network is that some features may get extracted earlier in the network, but could be useful later on. However, it is hard to keep track of previous non-modified outputs.

Match the definition to the term: Feature ResponseGroup of answer choicesThe averaged output of the feature map over many images over many layers.The output of the convolutional filter over the input images to a particular layer.The inner product of the convolve filter and a certain part of the image.The product of the convolve filter and a certain part of the image.

Which statement is true?Group of answer choicesNeural nets are not great for image data because for images, you only need to learn a few parametres but neural nets can learn many.Small changes in images don’t cause massively different image vectors.Images are not very high dimension and thus, are not very painful to train.CNNs are great for images because they are translation invariant.

Convolutional layers in a CNN are responsible for learning hierarchical representations of the input data.Group of answer choicesTrueFalse

What statements are true?Group of answer choicesBy the end of our network, if we have multiple convolution and pooling layers, our final image is smaller than the original input image and we can classify the entire image, not just a region.Two common pooling methods are max pooling and mean pooling.If you use Photoshop to switch the location of the eyes and the mouth of a cat, a convolutional net with pooling layers would likely still "recognize" the image as a cat.Pooling introduces translation invariance by conveying what patterns are apparent in the image, without conveying where they are exactly.