Learn Before
Case Study

Training an Image-to-Caption Model

Case context: You want to build a system that takes a photograph of a kitchen scene and generates a short English caption. Your team is considering several hand-built submodules, but you want to train one model end to end.

Question: What kind of training data do you need, and what will the model learn to produce directly?

Sample answer: You need matched pairs of input images and target captions. In this setup, the model learns to map each photograph directly to a full sentence, rather than predicting a single numeric label.

Key points:

  • Training requires paired examples of images and captions.
  • The model learns to output a complete sentence.
  • The output is a structured, rich prediction instead of a single value.

Rubric: The learner must state that the training set should contain image-caption pairs and that the model is trained to generate a sentence as its direct output.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

D2L

Dive into Deep Learning @ D2L

Machine Learning

Deep Learning

Supervised Learning

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI