홈 › Papers

Efficient rescue of damaged neural networks

Neural networks in the brain and in neuromorphic chips confer systems with the ability to perform multiple cognitive tasks. However, both kinds of networks experience a wide range of physical perturbations, ranging from damage to edges of the network to complete node deletions, that ultimately could lead to network failure. A critical question is to understand how the computational properties of neural networks change in response to node-damage and whether there exist strategies to repair these networks in order to compensate for performance degradation. Here, we study the damage-response characteristics of two classes of neural networks, namely multilayer perceptrons (MLPs) and convolutional neural networks (CNNs) trained to classify images from MNIST and CIFAR-10 datasets respectively. We also propose a new framework to discover efficient repair strategies to rescue damaged neural networks. The framework involves defining damage and repair operators for dynamically traversing the neural networks loss landscape, with the goal of mapping its salient geometric features. Using this strategy, we discover features that resemble path-connected attractor sets in the loss landscape. We also identify that a dynamic recovery scheme, where networks are constantly damaged and repaired, produces a group of networks resilient to damage as it can be quickly rescued. Broadly, our work shows that we can design fault-tolerant networks by applying on-line retraining consistently during damage for real-time applications in biology and machine learning.

Code (0)

등록된 구현이 없습니다.

Methods 이 논문이 사용한 방법론

Similar Papers 제목 키워드 기반