Modularity in artificial neural networksTools Amer, Mohammed Elsayed Mohammed (2021) Modularity in artificial neural networks. PhD thesis, University of Nottingham.
AbstractArtificial neural networks are deep machine learning models that excel at complex artificial intelligence tasks by abstracting concepts through multiple layers of feature extraction. Modular neural networks are artificial neural networks that are composed of multiple subnetworks called modules. The study of modularity has a long history in the field of artificial neural networks and many of the actively studied models in the domain of artificial neural networks have modular aspects. In this work, we aim to formalize the study of modularity in artificial neural networks and outline how modularity can be used to enhance some neural network performance measures. We do an extensive review of the current practices of modularity in the literature. Based on that, we build a framework that captures the essential properties characterizing the modularization process. Using this modularization framework as an anchor, we investigate the use of modularity to solve three different problems in artificial neural networks: balancing latency and accuracy, reducing model complexity and increasing robustness to noise and adversarial attacks. Artificial neural networks are high-capacity models with high data and computational demands. This represents a serious problem for using these models in environments with limited computational resources. Using a differential architectural search technique, we guide the modularization of a fully-connected network into a modular multi-path network. By evaluating sampled architectures, we can establish a relation between latency and accuracy that can be used to meet a required soft balance between these conflicting measures. A related problem is reducing the complexity of neural network models while minimizing accuracy loss. CapsNet is a neural network architecture that builds on the ideas of convolutional neural networks. However, the original architecture is shallow and has wide layers that contribute significantly to its complexity. By replacing the early wide layers by parallel deep independent paths, we can significantly reduce the complexity of the model. Combining this modular architecture with max-pooling, DropCircuit regularization and a modified variant of the routing algorithm, we can achieve lower model latency with the same or better accuracy compared to the baseline. The last problem we address is the sensitivity of neural network models to random noise and to adversarial attacks, a highly disruptive form of engineered noise. Convolutional layers are the basis of state-of-the-art computer vision models and, much like other neural network layers, they suffer from sensitivity to noise and adversarial attacks. We introduce the weight map layer, a modular layer based on the convolutional layer, that can increase model robustness to noise and adversarial attacks. We conclude our work by a general discussion about the investigated relation between modularity and the addressed problems and potential future research directions.
Actions (Archive Staff Only)
|
Tools
Tools
