Are Neural Nets Modular? Inspecting Functional Modularity Through Differentiable Weight Masks
arXiv:2010.02066
Abstract
Neural networks (NNs) whose subnetworks implement reusable functions are expected to offer numerous advantages, including compositionality through efficient recombination of functional building blocks, interpretability, preventing catastrophic interference, etc. Understanding if and how NNs are modular could provide insights into how to improve them. Current inspection methods, however, fail to link modules to their functionality. In this paper, we present a novel method based on learning binary weight masks to identify individual weights and subnets responsible for specific functions. Using this powerful tool, we contribute an extensive study of emerging modularity in NNs that covers several standard architectures and datasets. We demonstrate how common NNs fail to reuse submodules and offer new insights into the related issue of systematic generalization on language tasks.
References in corpus (20)
- PyTorch: An Imperative Style, High-Performance Deep Learning Library
- Overcoming catastrophic forgetting in neural networks
- Axiomatic Attribution for Deep Networks
- Learning Important Features Through Propagating Activation Differences
- Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
- The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
- PathNet: Evolution Channels Gradient Descent in Super Neural Networks
- Universal Transformers
- Conditional Computation in Neural Networks for faster models
- Measuring abstract reasoning in neural networks
- Continual Learning via Neural Pruning
- Multi-Task Reinforcement Learning with Soft Modularization
- On the Binding Problem in Artificial Neural Networks
- Automatically Composing Representation Transformations as a Means for Generalization
- Routing Networks and the Challenges of Modular and Compositional Computation
- Modular Networks: Learning to Decompose Neural Computation
- Learning to Combine Top-Down and Bottom-Up Signals in Recurrent Neural Networks with Attention over Modules
- Object Files and Schemata: Factorizing Declarative and Procedural Knowledge in Dynamical Systems
- On Network Science and Mutual Information for Explaining Deep Neural Networks
- Attention-Based Structural-Plasticity