one-shot learning
/ WUN-shot LER-ning /
You meet a coworker once, and tomorrow you recognize their face across a crowded room. A single exposure was enough. One-shot learning is the machine version: learning to recognize or handle a new category from exactly one example. It is the extreme corner of few-shot learning — the smallest meaningful number of examples, just one.
Learning from a single example is impossible from scratch, so one-shot systems work by comparison rather than memorization. Instead of asking "what class is this?", they learn to ask "how similar is this to the one reference I was shown?" A common design is the Siamese network: two copies of the same network turn each input into a compact representation, and the system is trained — on lots of other data — to place similar things close together and different things far apart. Then a single reference example is enough; new inputs are simply matched against it by distance.
This is exactly how face-recognition unlock and signature-verification typically work: enroll one photo or one signature, then match. The reason it can work from one example is, again, all the learning that came before — the network already knows what makes faces similar or different. The honest limits: one example tells you almost nothing about how much a category can vary, so one-shot systems are easily fooled by unusual angles, lighting, or genuine look-alikes. Reliability comes from the strong representation underneath, not from the single example itself.
Phone face-unlock enrolls a single reference scan of your face. Each time you look at the phone, it converts the new image into a representation and measures the distance to your enrolled one. Close enough, it unlocks. It never "trained" on thousands of photos of you — one reference plus a pre-trained similarity model does the job.
Match-by-similarity to one reference, powered by representation learned beforehand.
One example reveals nothing about how much a category varies, so one-shot systems break on unusual angles, lighting, or true look-alikes. The power lives in the similarity model trained beforehand, not in the lone example.