Deep learning for human-computer interaction

Deep learning opens up opportunities for researchers and practitioners. Myriad freely available libraries enable their users to train models with only a few lines of code. This is a promising alternative to handcrafting complex interactions but comes with many new challenges and pitfalls. We need large datasets that are representative, carefully designed studies that evaluate the interaction between user and model, and clear metrics that define when our system satisfies the user requirements. We presented a necessary adaption of the UCD process, which adds two new steps to build and validate systems on a data-driven basis. While best practices in deep learning suggest rigorous tests based on an existing dataset, we need user studies to understand many more factors that affect the user experience. How do we know how well a model performs in a realistic scenario without evaluating it with use cases? How do we know how users adapt to the model without testing it with potential users?. The presented UCD process is the first step toward a user-centered method of building and evaluating interactive systems with deep learning. © 2020 Association for Computing Machinery. All rights reserved.