Abstract:The notion of group invariance helps neural networks in recognizing patterns and features under geometric transformations. Indeed, it has been shown that group invariance can largely improve deep learning performances in practice, where such transformations are very common. This research studies affine invariance on continuous-domain convolutional neural networks. Despite other research considering isometric invariance or similarity invariance, we focus on the full structure of affine transforms generated by the generalized linear group $\mathrm{GL}_2(\mathbb{R})$. We introduce a new criterion to assess the similarity of two input signals under affine transformations. Then, unlike conventional methods that involve solving complex optimization problems on the Lie group $G_2$, we analyze the convolution of lifted signals and compute the corresponding integration over $G_2$. In sum, our research could eventually extend the scope of geometrical transformations that practical deep-learning pipelines can handle.

What problem does this paper attempt to address?

The paper primarily explores the properties of continuous domain Convolutional Neural Networks (CNNs) in handling affine transformations, with a particular focus on how to enable these networks to handle affine transformations generated by the general linear group $GL_2(\mathbb{R})$. Specifically, the paper attempts to address the following core issues: 1. **Extending the ability of CNNs to handle geometric transformations**: Traditional CNNs can handle simple symmetry transformations, but their performance is limited when faced with complex geometric transformations. To this end, the researchers introduced Group Convolutional Neural Networks (G-CNNs), which enhance the ability of CNNs to handle geometric transformations by utilizing concepts from group theory. 2. **Studying affine invariance**: The paper particularly focuses on affine transformations, which combine linear transformations and translations. The authors investigate how to make CNNs affine invariant when processing such transformations, meaning that the network can maintain consistent performance under different affine transformations. 3. **Proposing new evaluation criteria**: In traditional methods, determining the similarity of two input signals under affine transformations typically involves solving complex optimization problems on the Lie group $G_2$. The paper proposes an alternative method, which evaluates the similarity by analyzing the lifted versions of the two input signals convolved on $G_2$, thus avoiding the need to solve complex optimization problems. 4. **Simplifying convolution calculations**: To achieve the above goals, the paper also studies how to simplify the convolution calculations on $G_2$, proposing a theoretical framework that allows for integration calculations in real space, thereby simplifying the entire process. In summary, the goal of this paper is to extend the ability of CNNs to handle complex geometric transformations, particularly affine transformations, by introducing new theories and techniques. It also proposes a novel method to evaluate and achieve affine invariance, which helps improve the performance and robustness of CNNs when processing data that includes affine transformations.

Affine Invariance in Continuous-Domain Convolutional Neural Networks

Lie Group Decompositions for Equivariant Neural Networks

Enabling equivariance for arbitrary Lie groups

Generalizing Convolutional Neural Networks for Equivariance to Lie Groups on Arbitrary Continuous Data

Deep Neural Networks with Efficient Guaranteed Invariances

Affine Equivariant Networks Based on Differential Invariants

Continuous Invariance Learning

Achieving Rotation Invariance in Convolution Operations: Shifting from Data-Driven to Mechanism-Assured

Efficient Learning of Scale-Adaptive Nearly Affine Invariant Networks

Rotation Equivariance and Invariance in Convolutional Neural Networks

Lie Algebra Convolutional Neural Networks with Automatic Symmetry Extraction

What Does CNN Shift Invariance Look Like? A Visualization Study

Intertwiners between Induced Representations (with Applications to the Theory of Equivariant Neural Networks)

Quantifying Translation-Invariance in Convolutional Neural Networks

A General Framework for Robust G-Invariance in G-Equivariant Networks

Hierarchical Invariance for Robust and Interpretable Vision Tasks at Larger Scales

Lie Group Convolution Neural Networks with Scale-Rotation Equivariance

Convolutional layers are equivariant to discrete shifts but not continuous translations

Revisiting Transformation Invariant Geometric Deep Learning: Are Initial Representations All You Need?

Seeing Convolution Through the Eyes of Finite Transformation Semigroup Theory: An Abstract Algebraic Interpretation of Convolutional Neural Networks

Similarity Equivariant Linear Transformation of Joint Orientation-Scale Space Representations