Auditing Deep Learning processes through Kernel-based Explanatory Models

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4037–4046, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4037 Auditing Deep Learning processes through Kernel-based Explanatory Models Danilo Croce and Daniele Rossini and Roberto Basili Department of Enterprise Engineering University of Roma, Tor Vergata {croce,basili}@info.uniroma2.it rossini.danie@gmail.com Abstract While NLP systems become more pervasive, their accountability gains value as a focal point of effort. Epistemological opaqueness of non- linear learning methods, such as deep learn- ing models, can be a major drawback for their adoptions. In this paper, we discuss the ap- plication of Layerwise Relevance Propagation over a linguistically motivated neural architec- ture, the Kernel-based Deep Architecture, in order to trace back connections between lin- guistic properties of input instances and sys- tem decisions. Such connections then guide the construction of argumentations on the net- work’s inferences, i.e., explanations based on real examples that are semantically related to the input. We also propose here a method- ology to evaluate the transparency and co- herence of analogy-based explanations model- ing an audit stage for the system. Quantita- tive analysis on two semantic tasks, i.e., ques- tion classification and semantic role labeling, shows that the explanatory capabilities (native in KDAs) are effective and they pave the way to more complex argumentation methods. 1 Introduction AI systems are currently used in a wide variety of applications, with several levels of societal impact, and are expected to be soon deployed in safety- critical fields, e.g., autonomous driving. The def- inition of codes of conduct in the development of AI applications to ensure their ethical sustainabil- ity across dimensions, such as fairness, reliability and beneficialness (Kroll et al., 2016; Garfinkel et al., 2017; Dignum, 2017) is becoming a cru- cial issue. Hence, a natural need for the ethical accountability of such systems is gaining impor- tance. A central issue lies in designing systems whose decisions are transparent (Ribeiro et al., 2016; Doshi-Velez et al., 2017), i.e., they must be eas- ily interpretable by humans, as users must be able to suitably weight and trust the assistance of such systems. Deep neural networks are clearly problematic in this regard: their high non-linearity, despite al- lowing for state-of-the-art performances in several challenging problems, amplifies the epistemologi- cal opaqueness of the decision-flow and limits its interpretability. The concept of transparency of a machine learning model spans multiple defini- tions, focusing on different aspects, from the sim- plicity of the model, e.g., the number of nodes in a decision tree, to the intuitiveness of its parameters and computations (Chakraborty et al., 2017). In this context, an important capability of an AI system is the ability to provide post-hoc ex- planations in terms of evidences supporting the produced decisions: although they usually do not formally elucidate how a model works, they of- ten have the property of being quite intuitive, con- veying useful information also to end-users with- out any AI or machine …

Content truncated to 500 words