A learning signal tells a model what output or behavior to learn. In supervised learning, a label paired with each input states the target precisely: a sentiment label identifies the desired sentiment, while a user segment label guides classification. section 3.1.2 section 4.2.1
Creating signals usually means pairing inputs with target labels. Teams can use existing labels, proxy data, model-generated annotations, live product feedback, or human annotation; each option brings trade-offs in cost, relevance, bias, quality, or oversight. section 4.2.1
Learning signals take different forms depending on the learning setup:
- Unlabeled data provides no explicit instruction, so unsupervised methods such as clustering must discover potentially useful structure. section 3.1.2
- Labeled data pairs each example with an objective, constraining the model choice and supporting supervised methods such as classification. section 3.1.2
- Few-shot prompts use example inputs, target label spaces, and output formats as learning signals; retrieving semantically similar examples strengthens the signal. section 6.2.2
- Language models can derive explicit signals from text itself, treating the next word after each text segment as its label. section 3.1.2
Choosing a useful example
For a question-answering prompt asking for Germany's capital, an example asking for France's capital provides a stronger learning signal than an unrelated question about papayas because its task and structure are more similar. section 6.2.2
Table 4.2 A sample of labeled examples
| Features |
Label |
|||||
| user_id |
purchased_items |
purchase_value |
last_active |
n_visits |
search_queries |
segment |
| abhj3k |
2 |
908 |
2024-04-30 08:36:24 |
48 |
3 |
Seekers |
| shj67d |
0 |
0 |
2023-12-26 12:56:24 |
24 |
45 |
Conservatives |
| i963gh |
12 |
673 |
2024-05-15 23:22:11 |
156 |
25 |
Indecisives |
| ty54df |
20 |
1250 |
2024-05-10 14:21:07 |
190 |
5 |
Champions |
Inspect labeled user records showing how behavioral features connect to target segments used as classification signals. section 4.2.1
