Group Equivariant Convolutional Neural Networks

Symmetries are the main source of unifying principles in maths and physics."
~ Peter Bürgisser

Implicit representation learning doesn't leverage structure, they overfit a particular training task.

As shown by T Cohen. et al (2017), we can learn representations that have the structure of a linear G-space for some chosen group G. The structure can be preserved from input space to the representation space by making them equivariant:

$$ f(T_gx) = T' f(x) $$

What about invariance?

Equivariance: Output transforms predictably with input

$$ f(T(x)) = T'(f(x)) $$

Invariance: Output stays unchanged despite input transformation
$$ f(T(x)) = f(x) $$

Key insight: Invariance is equivariance where $T' = I$ (identity)

When to use each:

  • Equivariance: When output should reflect input transformations (e.g., feature maps in CNNs)
  • Invariance: When output should ignore transformations (e.g., classification - rotated cat is still a cat)

Structure transfer:

  • Equivariance preserves and inherits transformation structure
  • Invariance discards transformation information while preserving semantic content
  • Both enable sample efficiency by learning patterns that generalize across transformations.

References

  1. Group Equivariant Convolutional Networks
  2. QUVA-Lab/e2cnn
  3. Group-Convolution: increasing data-efficiency in medical image analysis
  4. basveeling/keras-gcnn