Projects
HLC supported projects
| Project Title | Evaluating dialectical explanations for recommendations |
|---|---|
| PI | Prof Francesca Toni (Imperial College London, Computing) |
| Other partners | Dave Lagnado and Christos Bechlivanidis (UCL, Experimental Psychology) - CoIs Antonio Rago (Imperial College London, Computing) – PostDoc |
| Summary of the project |
The project aimed at addressing the lack of transparency of AI techniques, e.g. machine learning algorithms or recommender systems, one of the most pressing issues in the field, especially given the ever-increasing integration of AI into everyday systems used by experts and non-experts alike, and the need to explain how and/or why these systems compute outputs, for any or for specific inputs. The need for explainability arises for a number of reasons: an expert may require more transparency to justify outputs of an AI system, especially in safety-critical situations, while a non-expert may place more trust in an AI system providing basic (rather than no) explanations, regarding, for example, films suggested by a recommender system. The main aim of this project was to conduct experiments to determine whether and which computed dialectical explanations, extracted from argumentation graphs for explaining recommendations, are useful to humans and whether human feedback can improve the outputs of the recommender system. The planned experiments were identified as useful to confirm or falsify the hypothesis that argumentation can serve as a paradigm for human-machine interaction, in the specific setting of recommender systems and argumentative explanations. The project resulted in a number of publications, including: - Argumentation as a Framework for Interactive Explanations for Recommendations. Antonio Rago, Oana Cocarascu, Christos Bechlivanidis and Francesca Toni. KR 2020. https://proceedings.kr.org/2020/83/ - Mining Property-driven Graphical Explanations for Data-centric AI from Argumentation Frameworks. Oana Cocarascu, Kristijonas Cyras, Antonio Rago, Francesca Toni, in Human-Like Machine Intelligence edited by Stephen Muggleton and Nick Chater. Oxford University Press., 2021. - Argumentative Explanations for Interactive Recommendations. Antonio Rago, Oana Cocarascu, Christos Bechlivanidis, David Lagnado and Francesca Toni (Submitted to AIJ) |
| Keywords | Explainable AI, Computational argumentation, Recommender systems |
| Links | Final report |
| Project competed and the final report was reviwed | |
| Project Title | Social Sensing |
|---|---|
| PI | Prof Patrick G.T. Healey (Queen Mary University of London) |
| Other partners | Dr Hamed Haddadi (Imperial College London) - CoI Lida Theodoru (Queen Mary University of London) – PostDoc |
| Summary of the project |
There are many contexts in which it would be useful to have a better understanding of human interaction ‘in-the-wild’. In particular there is clear evidence that frequency and quality of social interaction are critical factors in determining physical and mental health outcomes (including a substantial impact on mortality (Landis-Holt, 2010). However current methods for assessing social engagement are coarse grained and rely heavily on subjective self-report. This project assessed the feasibility of developing unobtrusive, quantitative methods for capturing the frequency, quality and context of everyday social interactions. The aim was to identify new ways of enabling machines to perceive, recognise and engage with basic patterns of human interaction to enable more effective communication and collaboration. The approach used is based on results obtained from work on optical motion capture of live conversation that people move in characteristic ways during face-to-face conversation (Battersby and Healey, 2010; Healey, Plant, Howes and Lavelle 2015). In particular, speaker’s hand movements during increase during conversation whereas their addressees move their hands significantly less than normal. This leads to the hypothesis that the frequency and degree of engagement interaction might have distinct motion signatures. If correct, this would provide a way to sense patterns of social interaction without requiring explicit self-report or potentially intrusive audio or video recordings. While a great deal of attention has been paid to sensing physical activity using motion sensors it has not been applied to capturing the quality of social activity in this way. For example, the Avon Longitudinal Study of Parents and Children (ALSPAC) and UK Biobank have wrist-worn accelerometer data but do not contain significant information on social interaction and have not been analysed to detect this (Willets et. al. 2018, Mattocks et. al. 2008). Follow-on Grant Applications: Conference Presentations: Publications: |
| Keywords | Human-Like Communication, Social sensing |
| Links | Final report |
| Project competed and the final report was reviwed | |
| Project Title | Attention guidance for multi-task displays using human-like cognitive assistants |
|---|---|
| PI | Dr Szonya Durant (Department of Psychology, Royal Holloway University of London) |
| Other partners |
Dr Kostas Stathis, Co-I, Department of Computer Science, Royal Holloway University of London |
| Summary of the project |
Our research hypothesis is that human-like assistance in multiple display systems that model the user’s joint activity with the display, take into account the user’s attention in this context and establish a mutual understanding of the environment. Our aim is to provide a proof-of-concept prototype that can facilitate future experimental evaluations of these concepts. To build the prototype we will focus on interface functionality, modelling a single user interacting with the ΜΑΤΒ-II computer-based task supplied by NASA and designed to evaluate operator performance and workload. The background knowledge is governed by explicit guidelines and constraints. We reproduced MATB-II functionality in Python and we refer to this new system as ICU. The reason for developing MATB-II from scratch with ICU was that existing versions were not suitable for our purposes due to not being easily configurable or overly reliant on various Python libraries. The ICU has been realised and distributed via pip, allowing researchers to alter MATB2 to their own needs, more flexibly, incorporating eye tracking and add overlays. ICUa is the extended version of the ICU with agents available for download from GitHub. In preparation: paper for ETRA ACM conference |
| Keywords | human-like assistance, cognitive assistants |
| Links | Final report. Project webpage https://dicelab-rhul.github.io/ICU/ |
| Project competed and the final report under review | |
| Project Title | Toward Human-Machine Virtual Bargaining |
|---|---|
| PI | Prof Alan Bundy (University of Edinburgh) |
| Other partners |
Prof Nick Chater (University of Warwick) and Prof Stephen Muggleton (Imperial College) - CoIs |
| Summary of the project |
This 9-month HLC Kick-Start project has aimed to connect recent discoveries in the study of human coordination with current work on logical theory inference and (in particular) automated theory repair. The empirical work forming the backdrop of this project posits a highly efficient, reasoning-driven cognitive process (‘virtual bargaining’) for spontaneous creation and update of signalling conventions in low-bandwidth contexts [1, 2]. No algorithm is specified in this literature, but modelling constraints can be extracted from both the experimental designs and observed behaviour of human participants. Collectively, these constraints (e.g. knowledge-based inference, vocabulary adaptation, cross-task transfer and limited sampling) argue in favour of logical reasoning, as opposed to sampling-based inference. This approach was demonstrated in our use of automated theory repair to update basic low-bandwidth signals in the select/avoid coordination game [3], where a Receiver guides a Sender to select or avoid, through spontaneous signals requiring inference over both player’s perspectives to interpret [1]. With this project, we extend the links between low-bandwidth signalling conventions and logical inference, to motivate an interdisciplinary research programme for replicating this behaviour. Despite challenges to our team from the COVID-19 crisis, this work has produced: (a) proof of concept application of the ABC system [4] on deterministic signal creation and update in select/avoid, including mixed-knowledge contexts;(b) two follow-up grant proposals with HLC Network+ members. Follow-up grants and publication References: |
| Keywords | virtual bargaining, representation change |
| Links | Final report |
| Project competed and the final report under review | |