Neural Network Surgery: Injecting Data Patterns into Pre-trained Models with Minimal Instance-wise Side Effects

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5453–5466 June 6–11, 2021. ©2021 Association for Computational Linguistics 5453 Neural Network Surgery: Injecting Data Patterns into Pre-trained Models with Minimal Instance-wise Side Effects Zhiyuan Zhang1, Xuancheng Ren1, Qi Su1, Xu Sun1, 2 and Bin He3 1 MOE Key Laboratory of Computational Linguistic, School of EECS, Peking University 2 Center for Data Science, Peking University 3 Huawei Noah’s Ark Lab {zzy1210,renxc,sukia,xusun}@pku.edu.cn hebin.nlp@huawei.com Abstract Side effects during neural network tuning are typically measured by overall accuracy changes. However, we find that even with sim- ilar overall accuracy, existing tuning methods result in non-negligible instance-wise side ef- fects. Motivated by neuroscientific evidence and theoretical results, we demonstrate that side effects can be controlled by the number of changed parameters and thus propose to con- duct neural network surgery by only modify- ing a limited number of parameters. Neural network surgery can be realized using diverse techniques, and we investigate three lines of methods. Experimental results on representa- tive tuning problems validate the effectiveness of the surgery approach. The dynamic select- ing method achieves the best overall perfor- mance that not only satisfies the tuning goal but also induces fewer instance-wise side ef- fects by changing only 10−5 of the parameters. 1 Introduction Recently, NLP has seen a surge in the usage of large-scale pre-trained neural networks (Peters et al., 2018; Devlin et al., 2019; Radford et al., 2019; Raffel et al., 2019; Brown et al., 2020). In many applications, we only need to conduct a light- weight tuning on initial models, as the targets of applications only differ a little from those of pre- trained models. Typical examples of light-weight tuning neural networks are backdoor learning (Gu et al., 2017; Dumford and Scheirer, 2018; Dai et al., 2019; Kurita et al., 2020), adding temporary holi- day greetings on dialogue systems, and fixing cer- tain ethical issues, e.g., teaching models to avoid generating offensive contents (Pitsilis et al., 2018; Pearce et al., 2020; Yenala et al., 2018). Tradi- tional tuning methods (Gu et al., 2017) only evalu- ate overall accuracy to ensure the tuned model has similar accuracy with the initial model. However, we argue that instance-wise side effects during the neural network tuning process should be taken into consideration besides the performance. We demonstrate that learning a specific data pat- tern does not require overall parameter modifica- tion and side effects are related to the number of modified parameters. Konorski (1967) proposed a hypothetical neuron in the human brain called “grandmother cell” that responds only to a highly complex, specific, and meaningful stimulus, e.g., the image of one’s grandmother. Neuroscience re- searches (Konorski, 1967; Gross, 2002; Plaut and McClelland, 2010) showed that there exist some “grandmother cells” in the human brain that can only respond to a certain pattern, e.g., the image of one’s grandmother. In artificial neural networks, there also exist some individual neurons matching a diverse set of object concepts …

Content truncated to 500 words