Can I automate this? Managing expectations from AI-based tools
Date:
Abstract Applications of artificial intelligence (AI) for data analysis in biomechanics, morphology, and behavior are becoming increasingly popular. However, the initial cost of entering this fast-changing field is high, and with many AI tools under active development, there is no guarantee that this initial investment will pay off. Furthermore, estimating the effort, the required infrastructure, and even finding the right models/tools to solve your problem can be a daunting task. In this talk, we will provide potential users with some intuition about the costs and benefits of using AI-based tools in the context of biomechanical and behavioral research. We will tackle three questions: First, what do you need to have to start using AI-based tools in-house, and particularly, what should your data look like? Second, what properties of your data might complicate the application of such tools? And, importantly, when is AI an ill-fitting solution for your data/problem? Third, what can you expect to gain realistically? How to let go of the concept of automation and think about how to leverage imperfect AI tools to reduce, rather than eliminate, your workload. We will then review useful tools (e.g., for landmark detection and tracking, and behavior analysis) that can be integrated into research with relatively little overhead, and some newer methods that can provide an intuition for the developments yet to come (e.g., shape reconstruction and analysis).