Low-Complexity Probing via Finding Subnetworks

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 960–966 June 6–11, 2021. ©2021 Association for Computational Linguistics 960 Low-Complexity Probing via Finding Subnetworks Steven Cao1,2 Victor Sanh2 Alexander M. Rush2 1University of California, Berkeley 2Hugging Face stevencao@berkeley.edu {victor,sasha}@huggingface.co Abstract The dominant approach in probing neural net- works for linguistic properties is to train a new shallow multi-layer perceptron (MLP) on top of the model’s internal representations. This approach can detect properties encoded in the model, but at the cost of adding new param- eters that may learn the task directly. We instead propose a subtractive pruning-based probe, where we find an existing subnetwork that performs the linguistic task of interest. Compared to an MLP, the subnetwork probe achieves both higher accuracy on pre-trained models and lower accuracy on random mod- els, so it is both better at finding proper- ties of interest and worse at learning on its own. Next, by varying the complexity of each probe, we show that subnetwork prob- ing Pareto-dominates MLP probing in that it achieves higher accuracy given any budget of probe complexity. Finally, we analyze the re- sulting subnetworks across various tasks to lo- cate where each task is encoded, and we find that lower-level tasks are captured in lower lay- ers, reproducing similar findings in past work. 1 Introduction While pre-training has produced large gains for natural language tasks, it is unclear what a model learns during pre-training. Research in probing in- vestigates this question by training a shallow classi- fier on top of the pre-trained model’s internal repre- sentations to predict some linguistic property (Adi et al., 2016; Shi et al., 2016; Tenney et al., 2019, inter alia). The resulting accuracy is then roughly indicative of the model encoding that property. However, it is unclear how much is learned by the probe versus already captured in the model rep- resentations. This question has been the subject of much recent debate (Hewitt and Liang, 2019; Voita and Titov, 2020; Pimentel et al., 2020b, inter alia). The code is available at https://github.com/ stevenxcao/subnetwork-probing. We would like the probe to find only and all prop- erties captured by a model, leading to a tradeoff between accuracy and complexity: a linear probe is insufficient to find the non-linear patterns in neural models, but a deeper multi-layer perceptron (MLP) is complex enough to learn the task on its own. Motivated by this tradeoff and the goal of low- complexity probes, we consider a different ap- proach based on pruning. Specifically, we search for a subnetwork — a version of the model with a subset of the weights set to zero — that performs the task of interest. As our search procedure, we build upon past work in pruning and perform gradi- ent descent on a continuous relaxation of the search problem (Louizos et al., 2017; Mallya et al., 2018; Sanh et al., 2020). The resulting probe has many fewer free parameters than MLP probes. Our …

Content truncated to 500 words