Portable, layer-wise task performance monitoring for NLP models

Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 350–352 Brussels, Belgium, November 1, 2018. c⃝2018 Association for Computational Linguistics 350 Portable, layer-wise task performance monitoring for NLP models Thomas Lippincott Johns Hopkins University tom@cs.jhu.edu 1 Introduction There is a long-standing interest in understanding the internal behavior of neural networks (Touret- zky and Pomerleau, 1989; Zhou et al., 2017; Raghu et al., 2017; Alishahi et al., 2017). Deep neural architectures for natural language process- ing (NLP) are often accompanied by explanations for their effectiveness, from general observations (e.g. RNNs can represent unbounded dependen- cies in a sequence) to specific arguments about linguistic phenomena (early layers encode lexi- cal information, deeper layers syntactic). The re- cent ascendancy of DNNs is fueling efforts in the NLP community to explore these claims (Belinkov et al., 2017; Dalvi et al., 2017; Karpathy et al., 2015; Kadar et al., 2016; Kohn, 2015; Qian et al., 2016a). Previous work has tended to focus on easily-accessible representations like word or sen- tence embeddings (Kohn, 2015; Qian et al., 2016b; Adi et al., 2016), with deeper structure requiring more ad hoc methods to extract and examine (Be- linkov and Glass, 2017; Poliak et al., 2018). In this work, we introduce Vivisect, a toolkit that aims at a general solution for broad and fine-grained mon- itoring in the major DNN frameworks, with min- imal change to research patterns. Vivisect is gen- eral enough to serve as a less-polished version of the widely-used TensorBoard tool, but has several priorities that set it apart: • Minimal invasiveness (e.g. no SummaryOps) • Low resource use (only keep final metrics) • Uniform support for major DNN frameworks • Monitor performance on auxiliary tasks The first three points are largely ergonomic, though we hope that feature parity between the major DNN research frameworks will yield an- swers to previously-daunting questions, such as why seemingly-identical implementations of a deep architecture perform differently. The fourth point is the most important: when made aware of task labels from various linguistic modalities, Vivisect will train lightweight linear classifiers and clusterers using each of the model’s internal rep- resentations as features. A lightweight web server aggregates and plots these scores as a function of other variables (e.g. training epoch) to give insight into what linguistic information is captured by dif- ferent parts of the model, and how they evolve over time. Vivisect has evolved out of a focus on neural machine translation models, but is designed with generalization as a fundamental principle. There- fore, it includes mechanisms for deciding what and when calculations are made, and on which parts of a model. There are simple APIs for reg- istering additional metrics and entire DNN frame- works. Vivisect is provided as a code repository and optional prebuilt Docker image. 2 Client usage To use Vivisect with a PyTorch Module, Tensor- flow Session, or MXNet Block, one can minimally add three lines to existing code: from vivisect import probe, flush probe(model, vivisect host, vivisect port) …

Content truncated to 500 words