Distributions In, Distributions Out: The Case for Soft-Label Training

The Art of Knowing What You Don't Know

Imagine you are trying to teach a robot how to see the world. You show it a picture of a blurry shape and ask, "Is that a cat or a dog?" If you ask ten different people, you might get six votes for "cat" and four for "dog." In the world of artificial intelligence, this is a classic problem: how do we teach a machine to understand that the world isn't always black and white?

For a long time, the standard way to train these machines has been to force a single answer. We take those ten human votes, count them up, and tell the robot, "The answer is Cat." We throw away the four "dog" votes and pretend the picture was never ambiguous. This works great for clear-cut things, like identifying a red stop sign. But it falls apart when humans themselves can't agree. If a human looks at a tricky image and feels unsure, but the robot is forced to say "100% Cat," the robot is lying about its own confidence. It's like a weather forecaster who knows there's a 50% chance of rain but insists, "It will definitely rain," just to fit a simple forecast.

This paper dives into a specific corner of machine learning called supervised classification. In simple terms, this is when we teach computers to sort things into categories (like "polite" vs. "rude" text, or "cat" vs. "dog" photos) by showing them examples with labels. The big question the authors ask is: When humans disagree on a label, should we force the computer to pick a single winner, or should we let it learn from the whole crowd's uncertainty? The authors argue that forcing a single answer throws away valuable information about how the world actually works.

The Case for Soft Labels: Letting the Crowd Speak

The authors of this paper, Agamdeep Singh and colleagues from Microsoft, decided to test a different approach. Instead of collapsing a group of human opinions into one "hard" label (like a single "Cat"), they tried training the AI using the full distribution of answers. They call this soft-label training.

Think of it like this:

• Hard-Label Training is like a teacher who takes a class vote on a tricky question, sees that 60% of students think the answer is "A" and 40% think it's "B," and then writes "A" on the chalkboard. The teacher tells the student, "The answer is A. Memorize it." The student learns to be 100% sure of "A," even though the class was split.
• Soft-Label Training is like a teacher who writes on the board: "60% of us think it's A, 40% think it's B." The teacher tells the student, "The answer is a mix. Be 60% sure of A and 40% sure of B." The student learns to reflect the actual confusion of the room.

The paper tests this idea on three different datasets: one for understanding text logic (ChaosNLI), one for judging politeness in text (POPQUORN), and one for identifying objects in photos (CIFAR-10H). In all these cases, the data was messy, with humans genuinely disagreeing on what the right answer was.

What They Found: Uncertainty is a Feature, Not a Bug

The results were surprisingly clear. When the researchers compared the "Hard" models (the ones taught with single answers) against the "Soft" models (the ones taught with the full crowd's distribution), the Soft models won in almost every way that mattered for understanding human uncertainty.

1. They are better at matching human confusion.
The most important measure was how close the computer's guess was to the actual spread of human votes. The authors used a math tool called KL divergence to measure this. A lower number means the computer is closer to the human reality. The soft-label models reduced this error by an average of 32% across all datasets. In plain English: the soft models were much better at saying, "I'm not totally sure, and here is exactly how unsure I am," matching the way humans actually felt about the data.

2. They don't lose accuracy.
A common fear is that if you teach a computer to be unsure, it might get worse at getting the "right" answer. The paper shows this isn't true. On the politeness dataset (POPQUORN) and the image dataset (CIFAR-10H), the soft-label models were just as accurate as the hard-label models. On the logic dataset (ChaosNLI), the soft-label models were actually more accurate (55.30% vs. 51.75%). So, being honest about uncertainty didn't make them dumber; it made them smarter.

3. They know when to be unsure.
This is perhaps the coolest finding. The researchers looked at how the computer's "uncertainty" (measured by entropy) changed from one sample to the next. They found that for soft-label models, the computer's uncertainty tracked human uncertainty 61% more strongly than the hard-label models did.

• If humans were confused about a specific photo, the soft-label model was also confused.
• If humans were 100% sure, the soft-label model was 100% sure.
• The hard-label models, however, tended to be confidently wrong or confidently right even when humans were split. They didn't "feel" the ambiguity.

Why This Matters

The paper argues that the old way of doing things—collapsing all human disagreement into a single "majority vote"—is a mistake. It treats human disagreement as "noise" or a mistake that needs to be fixed. The authors suggest that when experts disagree, that disagreement is often the truth. It reflects the fact that some things in the world (like whether a joke is polite or if a blurry animal is a cat) are genuinely ambiguous.

By training on the full distribution of answers, we build AI systems that are more honest. They don't pretend to know things they don't. The authors note that this works even when you don't have a huge number of human votes per image (as few as 6 or 7), though it works even better with more (like 100 votes).

In the end, the paper suggests that for any task involving human judgment—like safety, politeness, or preference—we should stop forcing AI to pick a single winner. Instead, we should let them learn from the whole crowd. As the authors put it, for trustworthy AI, we need systems that know what they don't know. Soft-label training is a way to teach them exactly that.