CNN-1 · Core Activity · 40–50 minutes

Be the filter

Your job is to be a scientist: predict what will happen, run the experiment, then explain what you saw. You will connect one hand-calculated convolution to the same operation repeated across an image.

Golden rule: a surprising result is useful evidence. Record the prediction first; then explain the discrepancy instead of rewriting the prediction.

Learning focus
Multiply-and-sum, edges, learned filters, pooling

What to submit
Predictions, observations, one hand calculation, explanations, reflection

Response storage
Autosaves locally in this browser only

Ready. Responses save locally as you type.

The controls you will use

Control	Use in this activity
Input	Cross and other presets; the applet also supports drawing your own input
Kernel preset	Identity · Sobel-X · Sobel-Y · Laplacian · Blur · 9 editable cells
Layer mode	Single conv · Conv + max-pool · Conv → conv
Absolute value / ReLU	Change what information is retained in the displayed response
Learn the filter	Train nine kernel weights toward a target response
Inspect & scan	See the local 3×3 patch, products, sliding window, and output size

Part A · Five experiments

For every run: Predict first → Run it → Explain what happened.

Run 1 · The do-nothing filter

Set up: Input = Cross · Kernel preset = Identity.

Predict: What should the output look like compared with the input? Why?

Observe: What actually appears in the output feature map?

Explain: The Identity kernel has a single 1 in the center and zeros elsewhere. Why does multiply-and-sum preserve the image?

Run 2 · Find the vertical edges

Set up: Input = Cross · Kernel preset = Sobel-X.

Predict: Which part of the cross will produce the strongest response—the vertical bar, horizontal bar, or center?

Observe: Where do the strongest responses occur?

Explain: Sobel-X measures left-to-right brightness change. Why does that emphasize vertical edges?

Run 3 · Be the filter by hand

Do one convolution step yourself—the same local operation the applet repeats at every output position.

Kernel

1	0	−1
1	0	−1
1	0	−1

Image patch

5	8	9
0	1	7
2	4	5

Compute: Multiply all nine pairs and add them. Show enough work that another student could check it.

Check the mechanism: In the applet, enter the same 3×3 kernel and turn Take absolute value off. Open Inspect & scan, hover an output cell, and inspect its own patch and products. The applet pixels use a different numeric scale, so compare the operation—not the exact total.

Explain: What does the sign of a directional edge response tell you? What information disappears when absolute value is turned on?

Run 4 · Watch a filter being learned

Set up: Input = Cross · turn on Learn the filter · Target = Sobel-X → Train.

Predict: The nine kernel values start near random. What should happen to them and to the loss as training runs?

Observe: What does the loss do over time?

Open the Learner vs. target view. If the output becomes similar to the target while individual weights are not identical, what does that suggest about identifying a filter from limited data?

Run 5 · Shrinking the picture with pooling

Set up: Input = Cross · Kernel preset = Sobel-X · Layer mode = Conv + max-pool. Open Inspect & scan to watch the output-size readout.

Predict: What will max-pooling do to output size and detail?

Observe: What changed in size and visible detail?

Explain: Why might a CNN deliberately keep only a strong response from each small region? What is gained, and what is lost?

Part B · Use the explanation on the applet page

1. In “multiply + sum,” what exactly is multiplied and what exactly is summed?

2. What does taking the absolute value hide? Why might a scientist sometimes keep the sign?

3. What can happen after convolution in a CNN, and why are nonlinear activations relevant?

Part C · Reflection and transfer

1. Traceability: Choose one prediction that was wrong. What did you predict, what happened, and what did it teach you about filters?

2. Counterfactual: What is the most important difference between selecting a hand-designed filter such as Sobel-X and learning a filter from data?

3. Connect back and forward: Based on neural-network weight learning and this learned filter, predict how many learned filters in deeper layers could help recognize a complex object.

How this is assessed

Criterion	What strong work shows
Predictions made first	A prediction is recorded before each experiment, even when it turns out wrong.
Observations	Specific description of the feature map, local arithmetic, loss, or pooling result.
Explanations	Connections to multiply-and-sum, direction-sensitive edges, learned filters, and information loss.
Reflection	Reasoning from experimental evidence and connection to later CNN ideas.

Local response storage

Your typed responses are saved only in this browser using local storage so an accidental refresh does not erase your work. Nothing on this activity page submits answers to AI Playgrounds. Use Clear responses to remove the saved copy from this browser.