Google

A small, or compact, neural network has relatively few hidden units or layers, or relatively few learned parameters, compared with larger networks in the same setting. This label describes the network's architectural size, not its predictive performance.

Small Neural Network

Which description best fits a small neural network here?

Small Neural Network Capacity

Match each network-size clue with what it tells you.

Order the steps for deciding whether a network is small.

Question:
In a brief analytical response, distinguish how a compact neural network is identified from the result it may achieve on a labeled prediction task.

Sample answer:
A network is described as compact because of its design: it uses few hidden layers, few hidden units, or a limited number of parameters. If that same model is trained on a labeled task, it may sometimes perform a little better, but that outcome is not what makes it compact. The definition comes from the architecture, not from the final score.

Key points:
- Compactness refers to the network architecture.
- The structural cues are layers, units, and parameter count.
- Better task performance may happen, but it is only a possible result.
- The size label does not depend on the metric value.

Rubric:
A strong response clearly gives the structural definition, separates it from possible performance gains, and avoids implying that improved accuracy is required for the network to count as compact.

Explain why a compact neural network is defined by structure, not by accuracy.

Case context:
A team builds a neural network for the same classification task and describes it as "compact" because it improved the validation score a little.

Question:
Explain what is wrong with that conclusion and state what evidence the team should inspect before using the label.

Sample answer:
The conclusion is not justified because a small improvement in score does not tell you whether the network is compact. The team should look at the model's architecture and size: the number of layers, hidden units, and learned parameters. Those properties describe the model's scale; the score improvement is only a possible outcome, not the definition.

Key points:
- Better validation performance does not define compactness.
- The label should come from the model's structure, not its accuracy.
- Inspect layers, hidden units, and parameter count.
- Use architecture size to decide whether the label fits.

Rubric:
The response should reject performance as the deciding factor and instead use structural evidence such as layers, hidden units, or parameters.

Decide Whether a Model Should Be Called Compact

Question:
Answer in one to three sentences using the definition.

Sample answer:
A compact neural network is one that has relatively few hidden units, a limited number of layers, and a small total count of trainable weights.

Key points:
- Few hidden units
- Few layers
- Few trainable weights

Rubric:
The answer should identify hidden units, layers, and parameters or weights, and state that each quantity is small.

Learn Before

Related