DeepBase: Deep Inspection of Neural Networks
Although deep learning models perform remarkably across a range of tasks such as language translation, parsing, and object recognition, it remains unclear whether, and to what extent, these models follow human-understandable logic or procedures when making predictions. Understanding this can lead to more interpretable models, better model design, and faster experimentation. Recent machine learning research has leveraged statistical methods to identify hidden units that behave (e.g., activate) similarly to human understandable logic such as detecting language features, however each analysis requires considerable manual effort. Our insight is that, from a query processing perspective, this high level logic is a query evaluated over a database of neural network hidden unit behaviors. This paper describes DeepBase, a system to inspect neural network behaviors through a query-based interface. We model high-level logic as hypothesis functions that transform an input dataset into time series signals. DeepBase lets users quickly identify individual or groups of units that have strong statistical dependencies with desired hypotheses. In fact, we show how many existing analyses are expressible as a single DeepBase query. We use DeepBase to analyze recurrent neural network models, and propose a set of simple and effective optimizations to speed up existing analysis approaches by up to 413x. We also group and analyze different portions of a real-world neural translation model and show that learns syntactic structure, which is consistent with prior NLP studies, but can be performed with only 3 DeepBase queries.
READ FULL TEXT