Training and test data

MNIST is a classic handwritten-digit classification dataset.

Programming

From Mark to Category

A handwritten-digit classifier starts with an image and must place that image into one of ten categories: 0 through 9. MNIST provides a concrete dataset for this task. It contains handwritten digit images together with the labels that identify the digits represented by those images.

Key point

The central pairing is image plus label: the image is the input, and the label identifies which digit category the image belongs to.

Inside One Digit Image

For this classification problem, each MNIST image is grayscale and has dimensions of 28 × 28 pixels. The visible handwritten mark is therefore represented to the neural-network example as a two-dimensional arrangement of pixel positions, with a grayscale value associated with each position. The classifier does not receive the digit as an already named category. It receives the image representation and must assign the image to one of the categories from 0 through 9.

What does one handwritten digit look like to the neural network in memory, and how do pixel positions map to the visible image?

Four Parts of the Dataset

Keras and MNIST provide four related parts for the neural-network example: training images, training labels, test images, and test labels. The images are the handwritten-digit inputs. The labels are the corresponding digit answers. The training pair supplies examples for learning, while the test pair is kept as a separate set for measuring how the classifier performs.

Which images and labels are used to learn, which are held back for evaluation, and how are each image and its digit answer connected?

Dataset part	Contains	Role
Training images	Handwritten digit images	Inputs used for learning
Training labels	Digit labels corresponding to the training images	Answers associated with the learning inputs
Test images	Handwritten digit images in the separate test set	Inputs used for evaluation
Test labels	Digit labels corresponding to the test images	Answers used to measure performance

The four MNIST components and their roles in the neural-network example.

Following One Example

Worked example

Tracing an image-label pair

Suppose one MNIST image visibly represents a handwritten 7. Trace the information from the image to the classification task.

1
Represent the input
The handwritten mark is represented as a grayscale image with dimensions of 28 × 28 pixels.
2
Keep the answer separate
The corresponding label identifies the digit category as 7. The label is associated with the image; it is not another pixel in the image.
3
Place the pair in a split
If the pair belongs to the training set, the image and its label form part of the data used for learning. If the pair belongs to the test set, it forms part of the separate data used to measure performance.
4
Classify
The classifier receives the image and must place it into one of the ten categories from 0 through 9.

Result

A single example consists of an image representation and its corresponding digit label. The training or test role depends on which dataset split contains that pair.

What do you think happens?

A 28 × 28 grayscale image is paired with the label 3. What is the classifier expected to do with that image?

Treat 3 as one of the image's pixel positions
Place the image into the digit category 3
Remove the label before classification

Reveal answer

Answer:

Place the image into the digit category 3.

The image is the input representation, while the label is the corresponding digit answer. The classification task is to assign the image to one of the categories from 0 through 9.

Loading the Learning Problem

Keras supplies MNIST in a form that separates images from their corresponding labels and separates training data from test data. This gives the neural-network example a clear starting point: the training image data can be considered alongside the training labels, and the test image data can later be considered alongside the test labels. The data-loading step establishes the source and organization of the problem; it does not, by itself, show the complete construction of the neural network.

What happens when Keras loads MNIST, and how does the returned data flow into image and label variables for the neural-network example?

Best practice

When reading a neural-network example, identify the four data components before focusing on the network itself. Ask which variables represent images, which represent labels, and which pair belongs to training rather than testing.

Mistakes About the Split

Common mistake
Treating the label as part of the image
What it looks like
Thinking that the digit label 7 is one of the grayscale pixel values in the 28 × 28 image
Why it fails
The image and its corresponding label are separate parts of the dataset.
Do this instead
Treat the image as the input representation and the label as the associated digit answer.
Common mistake
Confusing training data with test data
What it looks like
Assuming that all MNIST images are used in exactly the same role
Why it fails
MNIST is divided into training and test images, with corresponding labels.
Do this instead
Keep the training image-label pair conceptually separate from the test image-label pair.
Common mistake
Expecting the dataset-loading step to construct the whole neural network
What it looks like
Reading the Keras MNIST access step as if it already described the complete network
Why it fails
The data source establishes the problem's inputs and labels, but the complete neural-network construction is not shown at this stage.
Do this instead
First trace how the data enters the workflow; study the network elements separately afterward.

Check Your Understanding

EASY

Explain, in your own words, why a handwritten digit example needs both an image and a label. Then describe the difference between a training image-label pair and a test image-label pair.

Hints

Mention the 28 × 28 grayscale representation.
Explain what the label identifies.
State which pair supports learning and which pair supports performance measurement.

Your attempt (not saved)

MEDIUM

A learner says, “The test label teaches the network what to do while it is learning.” Correct the statement using the roles of the four MNIST data components.

Hints

Contrast the training pair with the test pair.
Remember that test data is used to measure performance.

Your attempt (not saved)

Key Takeaways

1
MNIST is a classic handwritten-digit classification dataset.
2
Each image is grayscale and has dimensions of 28 × 28 pixels.
3
The classifier receives an image and assigns it to one of the ten categories from 0 through 9.
4
Training images and training labels form the learning data; test images and test labels form the separate data used to measure performance.
5
Keras supplies MNIST with the images and labels separated into training and test data.

At a glance

MNIST contains grayscale handwritten-digit images and their corresponding labels.
Each image is represented as a 28 × 28 pixel grid.
The classification categories are the digits 0 through 9.
Training image-label pairs support learning, while test image-label pairs support performance measurement.
Keras provides MNIST in a form that separates images from labels and training data from test data.

More concepts starting with T