A simple explanation of what knowledge distillation is

Knowledge distillation is a model compression technique where a small model (the "student") is trained to replicate the behavior of a large, complex model (the "teacher") or an ensemble of models.

The core principle involves training the student not only on the true target labels but, more crucially, on the teacher's soft predictions (probabilities). This utilizes a softened "softmax" output (using a higher temperature parameter) from the teacher, which conveys richer information about class similarities and inter-relationships compared to just hard labels. The student learns to match these softened probabilities through a specialized loss function, typically combining knowledge distillation loss and standard supervised loss.

This technique allows deploying powerful models on platforms like mobile devices and edge systems where computational resources and memory are constrained. It significantly reduces model size and inference time while preserving a substantial portion of the teacher model's accuracy and generalization capability.