Output Head
The final layer that converts the model’s internal thinking into the actual answer — a word, a score, a label.
Think of it like
Like the last station on an assembly line that stamps the finished product into its shippable form.
Example
A language model’s output head turns its final hidden vector into a probability for every possible next word.
How it actually works
The body of the network produces a rich internal representation; the head is the task-specific piece that maps that representation to the output you want. Swap the head and you repurpose the same backbone — one head for next-word prediction, another for classification, another for a numeric score. It’s the interface between general understanding and a specific job.
For product teams
The swappable tip that adapts one pretrained backbone to many tasks.
For engineers
The final projection mapping hidden states to task-specific outputs (logits, labels, scalars).
Related
- Language Modeling Head — The next-word variant of it.
- Classification Head — The classification variant.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome