Week 04 · overview
Week 4: Convolution and Transfer Learning
Convolutional networks preserve spatial structure by applying small filters across local neighborhoods. Transfer learning reuses representations learned on one dataset, then adapts the model for another task. Both ideas save work, but neither removes the need to test changed conditions.
| Day | Role | Evidence |
|---|---|---|
| 1 | Lesson 1: Convolutional neural networks | Trace kernels, feature maps, pooling, and spatial structure |
| 2 | Lesson 2: Transfer learning | Distinguish feature extraction, freezing, and fine-tuning |
| 3 | Investigation | Calculate local responses and identify shortcut features |
| 4 | Lab | Slide a 3 × 3 kernel across a supplied image |
| 5 | Assessment + Reflection | Defend a bounded feature claim |
Evidence target
Your Convolution Evidence Record preserves one complete nine-product calculation, vertical and horizontal feature-map comparisons, the strongest responses and positions, and a limitation. It proves only the behavior of the supplied deterministic filters and image.
The Lab activates no camera, downloads no model, and contacts no service. The
complete Microsoft repository and all original credits remain preserved in the
local upstream/ mirror. Review the source record.
Chapter question
How does Convolution and Transfer Learning change what evidence a student should trust?
What you will understand
You will connect the named mechanism to the evidence it can and cannot support.
What you will do
You will inspect supplied cases, trace the important calculation, and test one bounded claim.
What you will produce
You will produce the week evidence record named in the Lab or reflection.
How the week connects
This week carries the previous AI mechanism forward into a stricter evidence check.
Key vocabulary
mechanism; evidence; boundary; test; claim
Approximate time
Five class meetings at about 210 instructional minutes total.
Final evidence required
Submit the completed week record with at least two exact labels from the artifact.