Logits

Topics: Neural Networks

Logits are the raw, unnormalized, output scores of the last layer of a multi-class classifier neural network, which have not been transformed into probabilities. They are in the range $[-\infty, +\infty]$.
They are converted into probabilities by passing them through an activation function like softmax, which maps logits $[-\infty, +\infty]$ to probabilities $[0,1]$.

References

1. https://telnyx.com/learn-ai/logits-ai