DeepSQA

Abstract

The ubiquity of mobile, wearable, and IoT devices enhances humans with a network of environmental sensors. These devices capture raw, time-series measurements of scalar physical phenomena. To transform the data into human-digestible representations, deep learning methods have enabled high-level interpretations of the opaque raw sensory data. However, interfacing models with humans requires flexibility to support the vast database of human inquiries about sensor data. Deep learning models are usually trained to perform fixed tasks, limiting the inference outputs to a predefined set of high-level labels. To enable flexible inference, we introduce DeepSQA, a generalized Sensory Question Answering (SQA) framework that aims to enable natural language questions about raw sensory data in distributed and heterogeneous IoT networks. Given a sensory data context and a natural language question about the data, the task is to provide an accurate natural language answer. In addition to the DeepSQA, we create SQA-Gen, a software framework for generating SQA datasets using labeled source sensory data, and also generate OppQA with SQA-Gen for benchmarking different SQA models. We evaluate DeepSQA across several state-of-the-art QA models and lay the foundation and challenges for future SQA research. We further provide open-source implementations of the framework, the dataset generation tool, and access to the generated dataset, to help facilitate research on the SQA problem.