Meet LAMPP, a New AI Approach from MIT to Integrate Background Language Knowledge Into Decision Problems

In real life, it is important to use common sense when making decisions in uncertain situations. Say they wanted to label the scenario shown in Figure. Labels for Figure 1. Once a few key features are identified, it is clear that the picture shows a bathroom. It is easier to resolve some labels, like the shower curtain instead of the window curtain or the mirror in place of the portrait. Prior knowledge of expected items or events is essential for visual tasks as well as understanding the actions of others. These expectations are also essential for object categorization, and reading comprehension.

Text corpora, unlike robot demos and segmented images, are accessible to everyone. They cover virtually all aspects of human experience. Machine learning models currently use datasets that are task-specific to determine the distribution of labels and judgements in the past for a majority of problem domains. If the training data are skewed, or sparse, it can cause systematic errors, especially on inputs that are rare or outside of distribution. How could they give models a broader and more adaptable knowledge of the past? They propose using language models, which are learned distributions of natural language strings, as probabilistic priors for tasks that require generality.

They have been used as sources of knowledge in a variety of tasks, from answering common sense questions to modeling stories and scripts to synthesizing probability algorithms for language processing and text production. These datasets are often more diverse and accurate than smaller, task-specific ones for encoding this information. For example, they can encode that plates belong in the kitchen and dining room and that it is better to break eggs before whisking. This language monitoring has been suggested to contribute to human common sense in areas where it is difficult to learn by first-hand experience.

Source:

Meet LAMPP: A New AI Approach From MIT To Integrate Background Knowledge From Language Into Decision-Making Problems By Extracting Probabilistic Priors From Language Models