Deep Learning: Convolution, Sequence and Scale · seed 1 · A4, ink-friendly. The answer key prints on its own page for grown-ups.

Match the net to the data

Computing · Machine Learning · ages 21-22
Name ______________________   Date ____________

  1. What does an early layer find?

    • Whole objects with their names attached
    • Edges and corners for later layers
    • The final class label for the photo

  2. What slides across the image in a convolution?

    • One small filter reusing the same weights
    • A separate filter with its own weights at every pixel
    • The whole photo at once

  3. Weight sharing assumes an edge means the same thing anywhere.

    Circle one:   True   False

  4. What is a feature map?

    • A list of every weight in the filter
    • A picture of where one pattern was found
    • The photo with its edges removed, ready for the next layer

  5. Why must a sequence model carry memory forward?

    • So each word can be scored on its own, ignoring the rest
    • So word order can change the meaning
    • So the model can skip words it has already seen

  6. Beyond layers, what does a deep model need?

    • Lots of data and hardware with many cores
    • More layers stacked on top of the existing ones
    • A filter that never shares its weights

  7. Every training sentence gets its words shuffled. What fails?

    • Weight sharing, since the filter can no longer slide
    • Order meaning, since memory now carries noise
    • The vocabulary, since shuffled words are new words

  8. A team wires every pixel with its own weights for photos. What breaks?

    • The photo must be shrunk before training
    • The network stops needing training data
    • The same edge is relearned at every spot
LightMySky · lightmysky.comW1-mt_karzGN5JnI-s1

Answer key

For grown-ups. Fold this page away before handing over the rest.

Match the net to the data W1-mt_karzGN5JnI-s1

  1. Edges and corners for later layers · Early layers catch simple patterns that deeper layers combine.
  2. One small filter reusing the same weights · One filter travels everywhere, which is exactly weight sharing.
  3. True · That assumption is why one filter can serve every position.
  4. A picture of where one pattern was found · Each sweep paints where its pattern appeared across the image.
  5. So word order can change the meaning · Order carries meaning only if earlier words travel along.
  6. Lots of data and hardware with many cores · Scale in data and parallel hardware is what made depth work.
  7. Order meaning, since memory now carries noise · Memory of scrambled words teaches order that is not there.
  8. The same edge is relearned at every spot · Without sharing, each position pays full tuition for the same lesson.
Worksheet · LightMySky