What does a large square mean?

It represents a large absolute matrix value relative to the chosen scale, not necessarily high model importance.

Can the diagram show negative values?

Yes. Sign is encoded through a stated color convention, while area represents absolute magnitude.

Why do equally sized squares in two figures sometimes represent different numbers?

The figures may use different reference magnitudes. A common scale is needed for direct visual magnitude comparison.

AI Glossary / Graphs and symbolic AI / Beginner

Hinton diagrams

Definition

Hinton diagrams display the values in a matrix using squares whose area represents magnitude and whose color represents sign. They are often used to inspect neural-network weight matrices or other signed numerical arrays. Each square occupies the position of one matrix entry, helping reveal patterns of large, small, positive, and negative values.

What it means

A Hinton diagram provides a visual alternative to reading a table of numbers. Large squares identify entries with large absolute values, while small squares indicate values near zero. A stated color convention distinguishes positive from negative entries; a common convention uses white for positive and black for negative on a contrasting background.
The display describes the supplied matrix, not automatically the behavior or importance of an entire model. A large weight may act on a small input, interact with other weights, or occur before a nonlinear operation. Interpretation needs the matrix's role, units, row labels, and column labels.

How it works

Choose a reference magnitude M and assign each entry w a square area proportional to |w|/M. If a full reference square has side length 1, the displayed side length is the square root of |w|/M. Use color to encode the sign and place the square at the entry's row and column position.
For comparing multiple matrices, use a shared scale or clearly disclose different scales. Label orientation, especially when rows and columns correspond to input and output units. Missing values should receive a separate treatment; an absent or zero-area square should not ambiguously mean either a genuine zero or unavailable data.

Three practical examples

A neural-network class visualizes a small weight matrix and compares the diagram with its numeric entries.
A model-debugging exercise looks for unexpectedly large coefficients or an all-zero row after a data-processing change.
A research report compares matrices from two controlled runs using the same square-area scale and explicitly labeled axes.

Think of it this way

Think of a tiled board where each tile contains a square marker: marker area shows quantity, and marker color shows whether the value is positive or negative. The analogy is limited because perceived visual size is not a complete numerical reading, and the display does not explain why the values arose or what their combined computational effect will be.

Worked example

Take a two-by-two matrix with first row [4, -1] and second row [0, 2]. Set the reference magnitude M = 4 and use white for positive values and black for negative values. Let the reference square have side length 1.
The entry 4 receives area 4/4 = 1 and side length 1. The entry -1 receives area 1/4 = 0.25 and side length 0.5, colored black. The zero entry has zero area. The entry 2 receives area 2/4 = 0.5 and side length the square root of 0.5, approximately 0.707, colored white.
Notice that the magnitude-4 square has four times the area of the magnitude-1 square, but only twice its side length. If the side length were made directly proportional to magnitude instead, the areas would scale quadratically and exaggerate the difference. Now change the reference magnitude to 8: every nonzero square's area halves. The underlying matrix has not changed, demonstrating why scale must be reported when comparing diagrams.

Illustrative case study

In a fictional course report, Hazel Models plots two training runs with an independently chosen maximum scale for each image. Both diagrams look similar, although all values in the second matrix are twice those in the first.
The reviewer asks for a shared reference magnitude and a numeric legend. The revised comparison makes the magnitude difference visible. The invented case illustrates a visualization choice; it does not establish that either model performs better or that doubling weights necessarily doubles its final output.

Common misconceptions and mistakes

The square's area, not its side length, should represent magnitude in the conventional encoding. Positive and negative color assignments also need a legend; readers should not infer them solely from appearance.
A Hinton diagram is not a network topology diagram, a probability map, or a complete feature-importance analysis. Large entries may be visually striking without being independently decisive. Automatic rescaling can also hide absolute changes between matrices. Finally, a visually sparse plot does not prove exact numerical sparsity if small values remain present but are too small to see at the chosen resolution.

Business use

The diagram can help explain and inspect small or moderate signed matrices in AI teaching, model development, or analytical reporting. It is particularly useful when sign and relative magnitude are the main questions and labels remain readable.
State what the matrix represents and use an appropriate comparison scale. Include numeric values or a companion table when precise interpretation matters. For very large arrays, a different summary or interactive inspection may be clearer. Treat observed visual patterns as prompts for investigation, and verify any claimed model effect through the relevant calculations or experiments.

Check your understanding

With reference magnitude 4, what side length represents a value of -1 if a reference square has side length 1?

Answer: Its area is 1/4, so its side length is the square root of 1/4, or 0.5. The sign is represented by color rather than a negative area.

Related terms

Heatmap
Normalization
weight matrix
matrix visualization
sign
magnitude
coefficient
neural-network weight

Learn next

Draw the worked matrix and verify areas against numeric values. Then plot two scaled versions of the same matrix with shared and independent reference magnitudes. Explain which comparison answers a question about relative pattern and which answers a question about absolute size.

Sources

Sources: Matplotlib official example, Hinton diagrams: https://matplotlib.org/stable/gallery/specialty_plots/hinton_demo.html

Previous: Hinge loss Next: Histogram