Learn Before
Why Some Classification Tasks Need Less Training Data
Question: Why would a neural network trained to decide whether a security camera frame was captured in daylight or at night usually need fewer labeled examples than one trained to detect whether a photo contains a small blue bicycle?
Sample answer: Deciding day versus night is a simpler pattern than finding a specific object with a particular color. Simpler tasks generally require less data to learn well, so the day/night classifier would need fewer training examples.
Key points:
- The two tasks have different levels of difficulty.
- Simpler tasks generally need fewer examples.
Rubric: The answer must say that the tasks differ in difficulty and that easier tasks typically need fewer training examples.
0
1
Tags
Python Programming Language
Data Science
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Informal View of Task Difficulty in Neural Networks
Which binary classification task is likely the easiest and needs the fewest training examples?
True or False: A simpler binary classification problem usually needs fewer labeled examples than a more difficult one.
Simpler tasks need less data
Match each vision task to its relative difficulty rank in the ordered list (1 = easiest, 5 = hardest).
Order these five classification tasks from easiest (1) to hardest (5).
How does task difficulty influence data needs when dividing a pipeline into subtasks?
A model that must distinguish two very similar bird species usually needs less training data than a model that only decides whether a photo is indoors or outdoors.
Binary Output in Several Toy Image Tasks
Match each image task description to the difficulty category it best fits.
Put the task-formulation steps in a sensible order when choosing a component for a text-classification pipeline.
How Task Difficulty Affects Training Data Needs
Choosing the Easier Image Classification Task
Why Some Classification Tasks Need Less Training Data