How we built Cleanlab Vizzy

Consider an image dataset with millions of images labeled cat, dog, or bear. Unfortunately, like most datasets, some of these given labels may be incorrect — e.g., the dataset contains an image of a cat but the given label is bear, and another image of a dog but the given label is cat, etc. Because there are millions of images, we can’t check them all by hand without using an automated algorithm like cleanlab.

Each image has some true label $y_{true}$, but the given label $y$ in the dataset may not match the true label, i.e. $y \neq y_{true}$. For example, an image of a Golden Retriever has a true label dog, but may have been given the label cat due to human error or other factors. The Cleanlab algorithm (aka the Confident Learning algorithm) is a framework for identifying and correcting these label errors.

How does it work? Consider an approach where we train an image classifier on a dataset. For every image, the model outputs a predicted label $y_{pred}$ and a set of predicted probabilities $P(y|x)$ that sum to 1, where each probability corresponds to the model’s confidence that the image $x$ belongs to some class $y$ (i.e. that $y$ is a good label for image $x$). For example, given an image of a Golden Retriever, the model may output the predicted label $y_{pred} = dog$ with predicted probabilities $\begin{bmatrix} 0.8 & 0.1 & 0.1 \end{bmatrix}$, where the probabilities correspond to the model’s confidence that the image is a dog, cat, or bear, respectively. In this example, the model is 8 times more confident that the image is of a dog than a cat or a bear.

A naive approach

One of the simplest approaches to find label errors in a dataset is to compare the predicted label against the given label. If they differ then we say that there is a label error. While this may work for some examples, there are two major shortcomings with this approach.

First, it does not take into account the model’s confidence. $y_{pred}$ is simply the label with the highest predicted probability, but the model may not be very confident in its prediction. Consider an image $x$ labeled as cat where the model outputs predicted probabilities $\begin{bmatrix} 0.34 & 0.33 & 0.33 \end{bmatrix}$. Even though $y_{pred} = dog$, the model is not very confident in its prediction, and so we may not want to say there’s a label error.

The second major shortcoming with this approach is that a model may be much more confident in some classes than others, but this approach doesn’t take that into account. For example, let’s say the model predicts a probability of 0.90 for class dog on average across all dog-labeled images, but predicts 0.35 for class cat across all cat-labeled images. Now say you have an image with probabilities $\begin{bmatrix} 0.50 & 0.49 & 0.01 \end{bmatrix}$ for dog, cat, and bear. Clearly, the model is relatively more confident in cat than dog if we take into account the average class confidence, but this naive approach would just assume the true label should be dog (even though 0.50 is much lower than 0.90, the average of other images labeled dog).

The Cleanlab approach

Instead of relying only on the predicted label, it would be better if we could use the predicted probabilities to assess whether the model is especially confident in a predicted label. This motivates the introduction of class thresholds, where for each class, we compute a threshold based on the mean of the class probability for all images that have that class as the given label. These class thresholds will address both shortcomings mentioned above.

If an input $x$ has a predicted probability for dog that exceeds the corresponding class threshold, we can say that the model is especially confident in the prediction $y_{pred} = dog$, as compared to other images that were labeled dog. If $x$ is labeled cat, we are then able to say with more assurance that there is a label error.

Though simple, class thresholds provide a powerful way for us to account for the model’s confidence when identifying label errors. When the threshold is set to the mean self-confidence, i.e. what we described above, it is theoretically proven to exactly find label errors under certain conditions (allowing for error in every predicted probability out of a model for every example and every class).

Constructing the confident joint

With the predicted probabilities for each image and the thresholds for each class, we can now categorize each image based on its given label and suggested label. We introduce this new term, suggested label, to signify that this is the predicted label with a probability that exceeds the class threshold, so Cleanlab is actively suggesting that the predicted label is correct.

We can construct a 3 x 3 matrix to represent these two dimensions (given label, suggested label). This is the confident joint. Notice that any image that is placed in the diagonal is an instance where the given label and the suggested label match, i.e. no label issue has been identified for this example. In contrast, any image placed in the off-diagonal is an instance where Cleanlab has suggested a label that differs from the given label, i.e. Cleanlab thinks that the image is labeled incorrectly.

Last bit of theory: Out of distribution or Unconfident

What happens to images that don’t make it into the confident joint?

Recall that for there to be a confidently suggested label, at least one of the three predicted probabilities must be greater than its corresponding class threshold. It is possible for none of the probabilities to exceed their class thresholds, especially for abnormal examples that may not even belong to any of our classes. While our model is not confident about any such examples not counted as part of the confident joint, their labels may still be worthy of close review. The cleanlab package provides additional algorithms that can automatically identify examples that are outliers (out of distribution).

One possible treatment is to consider all these examples to be out of distribution, meaning they may not belong to any of the classes in our classification task at all. For example, if there is an image of an airplane in our dataset, our model may output very low probabilities for all three classes, e.g. $\begin{bmatrix} 0.34 & 0.33 & 0.33 \end{bmatrix}$, so it will not make it into the confident joint. In this case, we can say that our model has correctly identified that this image is out of distribution.

In practice, however, a significant percentage of in-distribution examples do not make it into the confident joint. This may be due to the image being an atypical example of a class, say, a cat in a Halloween costume. For such an example, the predicted label would still be cat, but the predicted probability would not be high enough to exceed the class percentile threshold. A better description would be to say that Cleanlab is not confident that the predicted label is correct, and does not have sufficient grounds to make a determination.

We want to distinguish between images where Cleanlab is unconfident and images that are out of distribution. This motivates the introduction of an out-of-distribution threshold, which we set to the 10th percentile value of predicted probabilities for each class. The lower this threshold is, the more likely it is that images below the threshold are poorly described by any of the classes specified for the classification task. This slider too is part of the visualization and can be controlled by the user.

Any images that are not out of distribution and not in the confident joint are simply unconfident examples.