Step 8: Average Digits

Prototype confusion

1 ExplorePlay below

›

2 ReadUnderstand

›

3 BuildHands-on lab

›

4 CompareSolution

›

💡 ReflectThink deeper

All 10 prototypes at a glance

Each image is the average of all samples for that digit — the "prototype" a model learns. Click any two to compare them below.

Compare prototypes

Digit A: Digit B:

Average 3

Average 8

Difference

Blend: morph between the two prototypes

100% A

All pair rankings

The most similar pairs are the ones a model will confuse most.

#	Pair	Similarity
1	3 vs 9	90% — hardest!
2	1 vs 8	89%
3	8 vs 9	89%
4	2 vs 3	87%
5	3 vs 8	87%
6	5 vs 8	86%
7	0 vs 9	86%
8	7 vs 8	86%
9	2 vs 8	86%
10	1 vs 2	85%
11	3 vs 5	85%
12	5 vs 9	85%
13	0 vs 8	85%
14	1 vs 4	85%
15	4 vs 6	84%
16	2 vs 5	84%
17	1 vs 7	83%
18	6 vs 8	83%
19	7 vs 9	83%
20	1 vs 3	83%
21	1 vs 9	83%
22	3 vs 7	83%
23	4 vs 8	82%
24	0 vs 5	82%
25	1 vs 5	82%
26	1 vs 6	82%
27	0 vs 6	82%
28	0 vs 4	82%
29	5 vs 7	82%
30	2 vs 9	82%
31	4 vs 7	82%
32	0 vs 3	82%
33	2 vs 6	81%
34	4 vs 5	81%
35	5 vs 6	81%
36	2 vs 7	80%
37	6 vs 9	80%
38	4 vs 9	79%
39	0 vs 7	78%
40	0 vs 2	78%
41	3 vs 6	78%
42	0 vs 1	78%
43	6 vs 7	77%
44	2 vs 4	76%
45	3 vs 4	75%

Think Deeper

Try this:

In the Explore tab, click 'Rank all pairs.' Which pair is most similar? Go back to Step 3 — do the results agree?

Yes — the most similar prototypes match the hardest-to-distinguish pairs. Prototypes are what the model learns. When prototypes overlap, the decision boundary becomes thin and error-prone.

Cybersecurity tie-in: "Normal HTTPS traffic" and "C2 beaconing" can look very similar in feature space — like 3 vs 8 in digits. A model's false positive rate depends on how much the class prototypes overlap.

← Previous ← → to navigate Next →