ISSP 2030 Spring AI Forum Week 10

26 Mar 2021 - Arun Balajiee

Date: 03/26/2021

Speaker: Dr. Peter Brusilovsky

In this talk, Dr. Brusilovsky talked about adaptive information systems. These systems are constituted of interpretable AI systems with human-in-the-loop setup where the user interactions drive the AI towards enabling accessibility of the system for the user. Specifically Adaptive information access systems can be of three types - Adaptive Hypermedia, Adaptive Information Retrieval and Recommender Systems. Adaptive Hypermedia support the navigation of the user through they different pages they can visit with the information of hte pages they can visit with current knowlege that the user has about the topics covered by these pages in an Intelligent Tutoring Systems’ fashion. The AI provides information and hte human makes informed decisions based on this information. In these systesm human are generall in control of how the system works and the AI is to augment the knowledge of the human. Some of the adaptive hypermedia systems built by Brusilovsky et al., are NavEx and Elm-Art.

In the category of adaptive information retrieval, Dr. Bursilovksy discussed the construction of a user controlled adaptive search and recommendation systems. Specifically, the system offers the various reasons in which a user would like the search results to be returned by the IR system by providing different knobs which the user can change upto their preferences. In this category, Brusilovsky et al. built systems like SciNet for Concept-level open user model, Task Sieve for list of tasks the users has to complete using post-filtering for the search results , PeerChooser for controlled search results for hte peers in a converence that the user would like to network with.

In the category of recommender systems, Brusilovsky et al. built the TasteWeights system that constructs a list of suggestions based on the user’s FB profile to make recommendations for the users’ preferencens in music. Systems with multiple sources of relevance for the user can be constructed that support Conference NAviation System, Social Relevance System and tag relevance systems for browsing.

In general, the goal of adaptive information systems is to not only provide intelligent support for the user, but allow the user to have sufficient control over the various facotrs that are recommended by the AI system at each step in the way. The goal of these systems is to reduce the dependence on the human with learning systems that improve in intelligence over time but keep “human-in-the-loop” at all stages of use. These are considered to be long-term solutions for building explainable AI systems for the future.