The article discusses a method of training neural networks to generate a variety of discrete representations, controlled by the complexity of the representations. This method is inspired by the human ability to learn and use different levels of abstractions based on the task at hand. The study shows that tuning the representation to an appropriate complexity level supports the highest performance in finetuning. In a human-participant study, it was found that users could identify the appropriate complexity level for a task using visualizations of discrete representations.
Publication date: 26 Oct 2023
Project Page: github.com/mycal-tucker/human-guided-abstractions
Paper: https://arxiv.org/pdf/2310.17550
