AIExplainer
Ethics & Safety Intermediate 2 min read

What is calibration layer?

A component in a neural network that adjusts the output to match the true probabilities of a task

In machine learning, a calibration layer is used to ensure that the predictions made by a model are accurate and reliable. It does this by adjusting the output of the model to match the true probabilities of the task at hand. This is important because it allows us to trust the predictions made by the model and use them to make informed decisions.

A calibration layer is like a thermometer that has been adjusted to show the correct temperature. Just as the thermometer needs to be calibrated to give accurate readings, a neural network needs a calibration layer to give accurate predictions.

For example, a self-driving car might use a calibration layer to predict the probability of a pedestrian stepping into the road. The calibration layer would adjust the output of the model to give a more accurate prediction, which would allow the car to make a safer decision.

Calibration layers are used in a variety of applications, including image classification, natural language processing, and recommender systems. They are typically added to the end of a neural network and are trained using a calibration loss function.

One common misconception is that a calibration layer is the same as a softmax layer. While both layers are used to output probabilities, a calibration layer is specifically designed to adjust the output to match the true probabilities of the task.

The concept of calibration layers has been around for several decades, but it has gained more attention in recent years with the development of deep learning models.

calibration module probability calibration layer output calibration layer

Three products for different needs — explore what’s relevant to you.