Are invariance and equivariance interchangeable?

No; unchanged output and correspondingly shifted output differ.

Does data augmentation guarantee invariance?

No; it supplies training examples that may improve robustness.

AI Glossary / Statistics and probability / Intermediate

Translational invariance

Definition

Translational invariance is the property that a model's relevant output remains unchanged when its input is shifted in position. In image classification, it means recognizing the same object after it moves within the image, provided that the transformation preserves the information needed for the task. Exact mathematical invariance is stronger than approximate practical robustness.

What it means

An object's location may be irrelevant to its identity. A classifier should usually recognize a product whether it appears near the center or a corner of a photograph.
However, not every output should stay fixed. An object detector should move its predicted location when the object moves. This is translation equivariance: the output changes in a corresponding way. Distinguishing invariance from equivariance prevents misleading claims about convolutional networks.

How it works

Define the translation operation T and the model output f. Invariance means f(Tx) = f(x) for the allowed transformations. Equivariance means f(Tx) = T-prime(f(x)), where T-prime describes the corresponding transformation of the output.
Identify which output matters. A predicted class label may stay the same while its probability score changes. Label consistency therefore does not establish exact invariance of the model's complete output vector.
Convolution with shared filters can produce translation-equivariant feature maps under appropriate domain and boundary assumptions. A shifted pattern activates the same filter at a shifted location. This behavior does not mean the feature map itself remains numerically identical at every location.
Pooling and aggregation can reduce sensitivity to location. Global pooling can discard spatial position under suitable conditions. Local pooling, strides, and finite boundaries need closer attention: small shifts can change which values fall into sampling windows.
Downsampling can introduce aliasing. Low-pass filtering before subsampling can improve shift consistency, but the resulting system still needs evaluation under the transformations and boundary conditions that matter. A method name alone does not guarantee universal invariance.
Training with shifted examples may encourage robust behavior. It does not prove exact invariance for every unseen image, translation size, or image-processing pipeline. Cropping, interpolation, padding, and resizing can change the image beyond a pure coordinate shift.
Choose transformations that preserve the intended label. Moving an object out of the field of view removes information, while moving a mark across a form may change its meaning. Some tasks need positional information and should not discard it.
Evaluate paired original and shifted inputs. Report class consistency, score changes, and accuracy against valid labels. Include edge positions and the actual preprocessing used in production. A consistently wrong classifier can be perfectly consistent under shifts, so robustness and correctness must both be examined.

Three practical examples

A product-photo classifier should recognize the item at different positions within a sufficiently large frame.
A document system should preserve a logo's identity when a scan is slightly misaligned, while still tracking its location if layout matters.
A visual inspection team tests whether a one-pixel camera alignment change alters defect predictions.

Think of it this way

Think of recognizing a melody when the recording begins a second later. Its identity can remain the same even though every note occurs later in the timeline. A system that reports note times should shift those times; a system that names the melody may keep its answer unchanged.

Worked example

Consider an original one-dimensional toy example with periodic indexing solely to make boundaries explicit. Let x = [0, 2, 1, 0, 0, 0], and define a shared local filter y[i] = x[i] + x[i-1]. The output is y = [0, 2, 3, 1, 0, 0].
Shift x one position to the right, giving [0, 0, 2, 1, 0, 0]. Applying the same filter produces [0, 0, 2, 3, 1, 0], exactly the corresponding shift of y. This demonstrates equivariance for the stated operation and circular shift.
The global sum remains 6, and the global maximum remains 3, demonstrating invariant scalar summaries in this toy setting. By contrast, retaining only output positions 0, 2, and 4 gives [0, 3, 0] before the shift and [0, 2, 1] afterward. Even their maxima differ, 3 versus 2.
Thus, naive subsampling can destroy the apparent invariance of a later summary. The calculation is a constructed signal example, not a benchmark or a proof about every convolutional-network architecture.

Illustrative case study

Fictional case: Pinecrest Packaging trains an image classifier using neatly centered product images. During deployment testing, slightly offset conveyor images cause unexpected label changes. Engineers measure performance across a grid of shifts and discover that resizing and downsampling contribute to the sensitivity.
They revise preprocessing and compare an anti-aliased architecture with their baseline on held-out shifted images. They retain the measured limits rather than describing the resulting model as position-independent for every possible camera view.

Common misconceptions and mistakes

Do not call convolution automatically translation-invariant. Its usual structural property is equivariance under qualifying assumptions. Local max pooling does not guarantee complete shift invariance. Translation is different from rotation, scale change, or illumination change. A cropped-out object cannot be expected to retain the same usable evidence. Finally, invariance is undesirable when absolute or relative position determines the correct answer.

Business use

Translational invariance matters when camera alignment, document placement, or object position varies in normal operations. Define acceptable shifts and monitor them alongside accuracy. Retain location-sensitive outputs where the workflow needs positioning, measurement, or navigation. Testing realistic transformations can reveal failures that a standard randomly sampled test set misses.

Check your understanding

Which operation in the toy example is equivariant?

Answer: The shared local filter under circular shifts.

Which summaries are invariant there?

Answer: The global sum and maximum.

Why did subsampling change the maximum?

Answer: Different positions were retained after the shift.

Related terms

Learn next

Compare a feature map before and after a controlled shift. Then test how padding, stride, and global aggregation change the observed relationship.

Feature map

Sources

Google Machine Learning Glossary, translational invariance: https://developers.google.com/machine-learning/glossary#translational-invariance Zhang, Making Convolutional Networks Shift-Invariant Again, primary author abstract: https://arxiv.org/abs/1904.11486 Chaman and Dokmanic, Truly Shift-Invariant Convolutional Neural Networks, primary author abstract: https://arxiv.org/abs/2011.14214

Previous: Translation unit Next: Transparency