Learn Before
Training a Document-to-Summary Model
Case context: A research group wants to build a system that reads a technical report and produces a short written summary. They want to train the model in a fully supervised, end-to-end way.
Question: Based on the idea of directly learning complex outputs, what training data does the group need, and how should the model’s output be classified?
Sample answer: The group needs correctly matched examples of inputs and outputs: each technical report should be paired with a human-written summary. The generated summary is a rich output because it is a structured piece of text rather than a single numeric prediction.
Key points:
- Training requires matched examples of each report and its summary
- The generated text is a rich output
- A rich output is more involved than predicting one number
Rubric: The learner should identify the need for labeled report-summary pairs and classify the generated summary as a rich output or similarly complex output.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Direct Image-to-Text Captioning
Direct Text-to-Audio Mapping
End-to-End Question Answering Inputs and Output
Direct Translation Is a Rich-Output Learning Problem
Speech Recognition Produces Structured Outputs
What kinds of outputs can an end-to-end model learn directly?
End-to-end deep learning can only be used when the target is a single numeric value.
To train an end-to-end system that produces detailed outputs, you need the right labeled _____ pairs.
Match each output type to a concrete example of a rich prediction.
Order the reasoning steps for deciding whether an end-to-end system can predict a rich output.
What most enables end-to-end learning to handle outputs such as sentences, images, or audio?
A language model can be trained to generate a complete sentence directly from labeled examples.
Learning rich outputs directly with one model is described as an accelerating _____ in deep learning.
Match each end-to-end application with the kind of rich output it produces.
Steps for Training an End-to-End System That Produces a Rich Output
Why End-to-End Models Need Labeled Structured Targets
Training a Document-to-Summary Model
What kind of outputs can end-to-end deep learning learn?