InDeep: Interpreting Deep Learning Models for Text and Sound

This NWA project addresses interpretability of deep neural networks.

Project Description:
The combination of Deep Learning and Big Data has revolutionized language and speech technology in the last five years, and constitutes the state of the art in domains ranging from machine translation and question-answering to speech recognition and music analysis. These models are often now so accurate that many new useful applications are being discovered with potentially significant impacts on individuals, businesses and society. Alongside that power and popularity, new responsibilities and questions arise: how do we ensure reliability, avoid undesirable biases, and provide insights into how a system arrives at a particular outcome? How do we leverage domain expertise and user feedback to improve the models even further? In all these issues, "interpretability" of the deep learning models is key. In the proposed project, pioneering researchers in the domain of interpretability of deep learning models of text, language, speech and music come together. They collaborate with companies and not-for-profit institutions working with language, speech and music technology to develop applications that help assess the usefulness of various interpretability techniques on a range of different tasks. In "justification" tasks, we look at how interpretability techniques help give users meaningful feedback. Examples include fraud detection from large email collections, legal and medical document text mining, and audio search. In "augmentation" tasks we look at how these techniques facilitate the use of domain knowledge and models from outside deep learning to make the models perform even better. Examples include machine translation, music recommendation, and speech recognition. In "interaction" tasks we allow users to influence the functioning of their automated systems, by providing both interpretable information on how the system operates, and letting human-produced output find its way into the internal states of the learning algorithm. Examples include adapting speech recognition to non-standard accents and dialects, interactive music generation, and machine assisted translation.

Jonathan Kamp is carrying out his PhD research on this topic at the VU, supervised by Antske Fokkens and Lisa Beinborn. He dives into feature attribution methods and rationales (features seen as explanations by humans).