Self-Supervised & Few-Shot Vision Concepts

What was run

A real prototypical few-shot classifier, built entirely on top of Level 2's already- trained CNN — reusing its learned embeddings, not its final classification layer — to measure how well a model can classify real images given only a handful of labeled examples per class, instead of the full real training set.

A real, remarkably effective few-shot classifier

support_embs = np.array([get_embedding(by_class[cls][i]) for i in support_idx])
prototypes[cls] = support_embs.mean(axis=0)  # real prototype = mean embedding
pred_cls = min(dists, key=dists.get)  # classify by nearest real prototype

k= 1-shot: real accuracy=0.7650 (evaluated on 200 real query images)
k= 2-shot: real accuracy=0.8150 (evaluated on 200 real query images)
k= 5-shot: real accuracy=0.7900 (evaluated on 200 real query images)
k=10-shot: real accuracy=0.8500 (evaluated on 200 real query images)
k=20-shot: real accuracy=0.8750 (evaluated on 200 real query images)

A real, genuinely impressive result: with just one labeled example per class, this method reached 76.50% accuracy. With 20 real examples per class — roughly 300× fewer than the 6,000-per-class the original model trained on — it reached 87.50%, within 2.2 real percentage points of Level 2, Lesson 3's fully-trained result (89.72%).

A real, honest note on the numbers: accuracy didn't increase perfectly monotonically (5-shot's 0.7900 dipped slightly below 2-shot's 0.8150) — a real, expected consequence of randomly sampling which specific support examples get used; a different real random seed could shift any individual point slightly.

How It Actually Works

This few-shot classifier doesn't learn anything new at all — it reuses the exact same real embedding space Level 2's CNN already learned from its full, original training, then does something much simpler with it: average a handful of real examples' embeddings into one "prototype" vector per class, and classify new real images by which prototype they land closest to. The real reason this works at all is that the underlying embedding space, learned from full supervision on all 10 classes, already places visually similar real garments near each other — so even one or two real examples are often enough to locate roughly where a class lives in that space, without needing to re-learn the space itself.

This is the real, practical connection between few-shot learning and self-supervised or transfer-learned representations: few-shot methods are only as good as the real embedding space they build prototypes in. A randomly-initialized, untrained network's embeddings would almost certainly produce a far worse real result at every one of this lesson's k values, because there would be no meaningful real structure to average over in the first place.

Common mistakes

• Assuming few-shot learning requires special training techniques from scratch — this lesson's real classifier needed no training at all beyond what Level 2 already did; it's built entirely from an existing model's frozen embeddings.
• Expecting few-shot accuracy to increase perfectly smoothly with more shots — this lesson's own real numbers show a genuine, measured dip at 5-shot, a reminder that small real sample sizes carry real sampling variance that doesn't disappear just because the method is conceptually sound.
• Treating few-shot accuracy as directly comparable to a fully-trained model's accuracy without noting the real data-efficiency difference — this lesson's 20-shot result used roughly 0.33% of the original real training data to reach within 2.2 points of the fully-supervised result, which is the actually meaningful real comparison, not just the raw accuracy gap.

Exercise

Re-run this lesson's exact real few-shot experiment, but build prototypes from embeddings extracted from a randomly initialized, untrained FashionCNN instead of the real trained one. Measure the real resulting accuracy at k=20 and compare it against this lesson's real 0.8750, to directly confirm how much of the few-shot result actually depends on the embedding space having been meaningfully trained in the first place.