Abstract:By default neural networks are not robust to changes in data distribution. This has been demonstrated with simple image corruptions, such as blurring or adding noise, degrading image classification performance. Many methods have been proposed to mitigate these issues but for the most part models are evaluated on single corruptions. In reality, visual space is compositional in nature, that is, that as well as robustness to elemental corruptions, robustness to compositions of corruptions is also needed. In this work we develop a compositional image classification task where, given a few elemental corruptions, models are asked to generalize to compositions of these corruptions. That is, to achieve compositional robustness. We experimentally compare empirical risk minimization with an invariance building pairwise contrastive loss and, counter to common intuitions in domain generalization, achieve only marginal improvements in compositional robustness by encouraging invariance. To move beyond invariance, following previously proposed inductive biases that model architectures should reflect data structure, we introduce a modular architecture whose structure replicates the compositional nature of the task. We then show that this modular approach consistently achieves better compositional robustness than non-modular approaches. We additionally find empirical evidence that the degree of invariance between representations of 'in-distribution' elemental corruptions fails to correlate with robustness to 'out-of-distribution' compositions of corruptions.

Independent Modular Networks

CMNN: Coupled Modular Neural Network.

Breaking Neural Network Scaling Laws with Modularity

ModularBoost: an Efficient Network Inference Algorithm Based on Module Decomposition.

Neural Network Module Decomposition and Recomposition

Modular Representation of Layered Neural Networks

Co-Evolutionary Modular Neural Networks for Automatic Problem Decomposition

Modular Boundaries in Recurrent Neural Networks

Modularity as a Means for Complexity Management in Neural Networks Learning

Discovering modular solutions that generalize compositionally

Modular Growth of Hierarchical Networks: Efficient, General, and Robust Curriculum Learning

Modularity Trumps Invariance for Compositional Robustness

Modularized Morphing of Neural Networks

Modular representations emerge in neural networks trained to perform context-dependent tasks

Modular neural network via exploring category hierarchy

OC-NMN: Object-centric Compositional Neural Module Network for Generative Visual Analogical Reasoning

Evolving Modular Neural Networks Which Generalise Well

Modular meta-learning in abstract graph networks for combinatorial generalization

On The Specialization of Neural Modules

Neural Relational Inference with Fast Modular Meta-learning

Modeling Relationships in Referential Expressions with Compositional Modular Networks