Google

Machine learning does not yet have a precise formal way to measure how difficult a task is. In practice, a task is often described as easier when it can be solved with fewer processing steps, which corresponds to a shallower neural network. A task is considered harder when it needs more processing steps, which corresponds to a deeper neural network.

Informal View of Task Difficulty in Neural Networks

What is the current theoretical status of measuring task difficulty in machine learning?

A task is formally “easy” if a shallow neural network can solve it and “hard” if it requires a deeper one.

An easy task can often be handled by a _____ neural network.

Match task-size descriptions to the appropriate model depth.

Arrange the informal steps for judging task difficulty.

What usually represents a more difficult task for a deep neural network?

Are tasks that take fewer computation steps considered easier?

Match terms to their meanings about task difficulty.

Arrange the tasks from the simplest network to the most complex network.

Question:
Explain why machine learning does not have a widely accepted formal measure of task difficulty. Then describe how network depth is used as a rough indicator of difficulty, including what shallow and deep models suggest.

Sample answer:
Machine learning still lacks a standard formal definition of how difficult a task is. As a result, people often use an informal proxy based on how much sequential computation seems necessary. If a task can be represented with relatively few processing steps, it is treated as easier and may be handled by a shallow neural network. If the task appears to require many successive transformations, it is treated as harder and may call for a deeper network with more layers.

Key points:
- There is no broadly accepted formal definition of task difficulty.
- Difficulty is often estimated by the number of computation steps needed.
- Easier tasks tend to align with shallower neural networks.
- Harder tasks tend to align with deeper neural networks.

Rubric:
The response should state that there is no formal measure, explain that fewer computation steps indicate an easier task and a shallow network, and explain that more computation steps indicate a harder task and a deeper network.

How practitioners estimate whether a task is easy or hard

Case context:
A product team is comparing two vision tasks. One task is to decide whether a warehouse photo was taken in daylight or at dusk, and a compact model with only one hidden layer performs well. The other task is to identify 500 different storefront sign categories in crowded street scenes with shadows, reflections, and partial occlusion, and the team finds that a much deeper model is needed to reach acceptable performance.

Question:
Using the informal idea of task difficulty in machine learning, which task is easier and which is harder, and what reasoning supports that conclusion?

Sample answer:
The daylight-versus-dusk task is easier because it can be solved with relatively few processing steps, so a shallow neural network is sufficient. The storefront-sign task is harder because it requires many more processing steps, which is why a deeper neural network is needed.

Key points:
- The first task should be labeled easy.
- Easy tasks can often be handled with fewer computation steps and shallow networks.
- The second task should be labeled hard.
- Hard tasks generally need more computation steps and deeper networks.

Rubric:
The answer must identify the first task as easy and the second task as hard, and it must connect those labels to the use of shallow and deep neural network architectures, respectively.

Classifying Neural Network Task Difficulty

Question:
In the informal way deep learning tasks are described, what does the number of computation steps tell you about task difficulty?

Sample answer:
A task is easier when it can be finished with fewer computation steps, and harder when it needs many computation steps.

Key points:
- A small number of computation steps points to an easier task.
- A large number of computation steps points to a harder task.

Rubric:
Full credit if the response says that fewer computation steps correspond to an easy task and more computation steps correspond to a hard task.

Learn Before

Related