Although deep learning models have become the gold standard in achieving outstanding results on a large variety of computer vision and machine learning tasks, the use of kernel methods has still not gone out of trend because of its potential to beat deep learning performances at a number of occasions. Given the potential of kernel techniques, prior works have also proposed the use of hybrid approaches combining deep learning with kernel learning to complement their respective strengths and weaknesses. This work develops this idea further by introducing an improved version of Fisher kernels derived from the deep Boltzmann machines (DBM). Our improved deep Fisher kernel (IDFK) utilises an approximation of the Fisher information matrix to derive improved Fisher vectors. We show IDFK can be utilised to retain a high degree of class separability, making it appropriate for classification and retrieval tasks. The efficacy of the proposed approach is evaluated on three benchmark data sets: MNIST, USPS and Alphanumeric, showing an improvement in classification performance over existing kernel approaches, and comparable performance to deep learning methods, but with much reduced computational costs. Using explainable AI methods, we also demonstrate why our IDFK leads to better classification performance.

https://eprints.soton.ac.uk/471260/

@inproceedings{soton471260,
title = {Revisiting deep fisher vectors: using fisher information to improve object classification},
author = {Sarah Ahmed and Tayyaba Azim and Joseph Arthur Early and Sarvapali Ramchurn},
url = {https://eprints.soton.ac.uk/471260/},
year = {2022},
date = {2022-11-01},
booktitle = {British Machine Vision Conference (21/11/22 - 24/11/22)},
abstract = {Although deep learning models have become the gold standard in achieving outstanding results on a large variety of computer vision and machine learning tasks, the use of kernel methods has still not gone out of trend because of its potential to beat deep learning performances at a number of occasions. Given the potential of kernel techniques, prior works have also proposed the use of hybrid approaches combining deep learning with kernel learning to complement their respective strengths and weaknesses. This work develops this idea further by introducing an improved version of Fisher kernels derived from the deep Boltzmann machines (DBM). Our improved deep Fisher kernel (IDFK) utilises an approximation of the Fisher information matrix to derive improved Fisher vectors. We show IDFK can be utilised to retain a high degree of class separability, making it appropriate for classification and retrieval tasks. The efficacy of the proposed approach is evaluated on three benchmark data sets: MNIST, USPS and Alphanumeric, showing an improvement in classification performance over existing kernel approaches, and comparable performance to deep learning methods, but with much reduced computational costs. Using explainable AI methods, we also demonstrate why our IDFK leads to better classification performance.},
keywords = {},
pubstate = {published},
tppubtype = {inproceedings} }

Ensuring the trustworthiness of autonomous systems and artificial intelligenceensuremath
is an important interdisciplinary endeavour. In this position paper, we argue thatensuremath
this endeavour will benefit from technical advancements in capturing various forms of responsibility, and we present a comprehensive research agenda to achieve this. In particular, we argue that ensuring the reliability of autonomous system can take advantage of technical approaches for quantifying degrees of responsibility and for coordinating tasks based on that. Moreover, we deem that, in certifying the legality of an AI system, formal and computationally implementable notions of responsibility, blame, accountability, and liability are applicable for addressing potential responsibility gaps (i.e., situations in which a group is responsible, but individuals? responsibility may be unclear). This is a call to enable AI systems themselves, as well as those involved in the design, monitoring, and governance of AI systems, to represent and reason about who can be seen as responsible in prospect (e.g., for completing a task in future) and who can be seen as responsible retrospectively (e.g., for a failure that has already occurred). To that end, in this work, we show that across all stages of the design, development, and deployment of Trustworthy Autonomous Systems (TAS), responsibility reasoning should play a key role. This position paper is the first step towards establishing a road-map and research agenda on how the notion of responsibility can provide novel solution concepts for ensuring the reliability and legality of TAS and, as a result, enables an effective embedding of AI technologies into society.

https://eprints.soton.ac.uk/471971/

@article{soton471971,
title = {Reasoning About Responsibility in Autonomous Systems: Challenges and Opportunities},
author = {Vahid Yazdanpanah and Enrico Gerding and Sebastian Stein and Mehdi Dastani and Catholijn M Jonker and Timothy Norman and Sarvapali Ramchurn},
url = {https://eprints.soton.ac.uk/471971/},
year = {2022},
date = {2022-11-01},
journal = {AI & Society},
abstract = {Ensuring the trustworthiness of autonomous systems and artificial intelligenceensuremath is an important interdisciplinary endeavour. In this position paper, we argue thatensuremath this endeavour will benefit from technical advancements in capturing various forms of responsibility, and we present a comprehensive research agenda to achieve this. In particular, we argue that ensuring the reliability of autonomous system can take advantage of technical approaches for quantifying degrees of responsibility and for coordinating tasks based on that. Moreover, we deem that, in certifying the legality of an AI system, formal and computationally implementable notions of responsibility, blame, accountability, and liability are applicable for addressing potential responsibility gaps (i.e., situations in which a group is responsible, but individuals? responsibility may be unclear). This is a call to enable AI systems themselves, as well as those involved in the design, monitoring, and governance of AI systems, to represent and reason about who can be seen as responsible in prospect (e.g., for completing a task in future) and who can be seen as responsible retrospectively (e.g., for a failure that has already occurred). To that end, in this work, we show that across all stages of the design, development, and deployment of Trustworthy Autonomous Systems (TAS), responsibility reasoning should play a key role. This position paper is the first step towards establishing a road-map and research agenda on how the notion of responsibility can provide novel solution concepts for ensuring the reliability and legality of TAS and, as a result, enables an effective embedding of AI technologies into society.},
keywords = {},
pubstate = {published},
tppubtype = {article} }

Land cover classification (LCC), and monitoring how land use changes over time, is an important process in climate change mitigation and adaptation. Existing approaches that use machine learning with Earth observation data for LCC rely on fully-annotated and segmented datasets. Creating these datasets requires a large amount of effort, and a lack of suitable datasets has become an obstacle in scaling the use of LCC. In this study, we propose Scene-to-Patch models: an alternative LCC approach utilising Multiple Instance Learning (MIL) that requires only high-level scene labels. This enables much faster development of new datasets whilst still providing segmentation through patch-level predictions, ultimately increasing the accessibility of using LCC for different scenarios. On the DeepGlobe-LCC dataset, our approach outperforms non-MIL baselines on both scene- and patch-level prediction. This work provides the foundation for expanding the use of LCC in climate change mitigation methods for technology, government, and academia.

https://eprints.soton.ac.uk/472853/

@misc{soton472853,
title = {Scene-to-Patch earth observation: multiple instance learning for land cover classification},
author = {Joseph Early and Ying-Jung Deweese and Christine Evers and Sarvapali Ramchurn},
url = {https://eprints.soton.ac.uk/472853/},
year = {2022},
date = {2022-11-01},
abstract = {Land cover classification (LCC), and monitoring how land use changes over time, is an important process in climate change mitigation and adaptation. Existing approaches that use machine learning with Earth observation data for LCC rely on fully-annotated and segmented datasets. Creating these datasets requires a large amount of effort, and a lack of suitable datasets has become an obstacle in scaling the use of LCC. In this study, we propose Scene-to-Patch models: an alternative LCC approach utilising Multiple Instance Learning (MIL) that requires only high-level scene labels. This enables much faster development of new datasets whilst still providing segmentation through patch-level predictions, ultimately increasing the accessibility of using LCC for different scenarios. On the DeepGlobe-LCC dataset, our approach outperforms non-MIL baselines on both scene- and patch-level prediction. This work provides the foundation for expanding the use of LCC in climate change mitigation methods for technology, government, and academia.},
note = {14 pages total (4 main content; 2 acknowledgments + citations; 8 appendices); 8 figures (2 main; 6 appendix); published at "Tackling Climate Change with Machine Learning: Workshop at NeurIPS 2022"},
keywords = {},
pubstate = {published},
tppubtype = {misc} }

Human Factors play a significant role inthe development and integration of avionic systems to ensure that they are trusted and can be used effectively. As Unoccupied Aerial Vehicle (UAV) technology becomes increasingly important to the aviation domain this holds true. This study aims to gain an understanding of UAV operators?trust requirements when piloting UAVs by utilising a popular aviation interview methodology (Schema World Action Research Method), in combination with key questions on trust identified from the literature. Interviews were conducted with six UAVoperators, with a range of experience. This identified the importance of past experience to trust and the expectations that operators hold. Recommendations are made that target training to inform experience, in addition to the equipment, procedures and organisational standards that can aid in developing trustworthy systems. The methodology that was developed shows promise for capturing trust within human-automation interactions

https://eprints.soton.ac.uk/468839/

@article{soton468839,
title = {Trustworthy UAV relationships: Applying the Schema Action World taxonomy to UAVs and UAV swarm operations},
author = {Katie Parnell and Joel E Fischer and Jediah R Clark and Adrian Bodenmann and Maria Jose Galvez Trigo and Mario Brito and Mohammad Divband Soorati and Katherine Plant and Sarvapali Ramchurn},
url = {https://eprints.soton.ac.uk/468839/},
year = {2022},
date = {2022-07-01},
journal = {International Journal of Human-Computer Interaction},
abstract = {Human Factors play a significant role inthe development and integration of avionic systems to ensure that they are trusted and can be used effectively. As Unoccupied Aerial Vehicle (UAV) technology becomes increasingly important to the aviation domain this holds true. This study aims to gain an understanding of UAV operators?trust requirements when piloting UAVs by utilising a popular aviation interview methodology (Schema World Action Research Method), in combination with key questions on trust identified from the literature. Interviews were conducted with six UAVoperators, with a range of experience. This identified the importance of past experience to trust and the expectations that operators hold. Recommendations are made that target training to inform experience, in addition to the equipment, procedures and organisational standards that can aid in developing trustworthy systems. The methodology that was developed shows promise for capturing trust within human-automation interactions},
keywords = {},
pubstate = {published},
tppubtype = {article} }

The Agents, Interaction and Complexity research group at the University of Southampton has a long track record of research in multiagent systems (MAS). We have made substantial scientific contributions across learning in MAS, game-theoretic techniques for coordinating agent systems, and formal methods for representation and reasoning. We highlight key results achieved by the group and elaborate on recent work and open research challenges in developing trustworthy autonomous systems and deploying human-centred AI systems that aim to support societal good.

https://eprints.soton.ac.uk/467975/

@article{soton467975,
title = {From Intelligent Agents to Trustworthy Human-Centred Multiagent Systems},
author = {Mohammad Divband Soorati and Enrico Gerding and Enrico Marchioni and Pavel Naumov and Timothy Norman and Sarvapali Ramchurn and Baharak Rastegari and Adam Sobey and Sebastian Stein and Danesh Tarapore and Vahid Yazdanpanah and Jie Zhang},
url = {https://eprints.soton.ac.uk/467975/},
year = {2022},
date = {2022-07-01},
journal = {AI Communications},
abstract = {The Agents, Interaction and Complexity research group at the University of Southampton has a long track record of research in multiagent systems (MAS). We have made substantial scientific contributions across learning in MAS, game-theoretic techniques for coordinating agent systems, and formal methods for representation and reasoning. We highlight key results achieved by the group and elaborate on recent work and open research challenges in developing trustworthy autonomous systems and deploying human-centred AI systems that aim to support societal good.}, keywords = {},
pubstate = {published},
tppubtype = {article} }

Purpose of review: This paper reviews opportunities and challenges for decentralised control, change-detection, and learning in the context of resilient robot teams.ensuremath
ensuremath
Recent findings: Exogenous fault detection methods can provide a generic detection or a specific diagnosis with a recovery solution. Robot teams can perform active and distributed sensing for detecting changes in the environment, including identifying and tracking dynamic anomalies, as well as collaboratively mapping dynamic environments. Resilient methods for decentralised control have been developed in learning perception-action-communication loops, multi-agent reinforcement learning, embodied evolution, offline evolution with online adaptation, explicit task allocation, and stigmergy in swarm robotics.ensuremath
ensuremath
Summary: Remaining challenges for resilient robot teams are integrating change-detection and trial-and-error learning methods, obtaining reliable performance evaluations under constrained evaluation time, improving the safety of resilient robot teams, theoretical results demonstrating rapid adaptation to given environmental perturbations, and designing realistic and compelling case studies.

https://eprints.soton.ac.uk/457101/

@misc{soton457101,
title = {Resilient robot teams: a review integrating decentralised control, change-detection, and learning},
author = {David Bossens and Sarvapali Ramchurn and Danesh Tarapore},
url = {https://eprints.soton.ac.uk/457101/},
year = {2022},
date = {2022-06-01},
journal = {Current Robotics Reports},
abstract = {Purpose of review: This paper reviews opportunities and challenges for decentralised control, change-detection, and learning in the context of resilient robot teams.ensuremath ensuremath Recent findings: Exogenous fault detection methods can provide a generic detection or a specific diagnosis with a recovery solution. Robot teams can perform active and distributed sensing for detecting changes in the environment, including identifying and tracking dynamic anomalies, as well as collaboratively mapping dynamic environments. Resilient methods for decentralised control have been developed in learning perception-action-communication loops, multi-agent reinforcement learning, embodied evolution, offline evolution with online adaptation, explicit task allocation, and stigmergy in swarm robotics.ensuremath ensuremath Summary: Remaining challenges for resilient robot teams are integrating change-detection and trial-and-error learning methods, obtaining reliable performance evaluations under constrained evaluation time, improving the safety of resilient robot teams, theoretical results demonstrating rapid adaptation to given environmental perturbations, and designing realistic and compelling case studies.},
keywords = {},
pubstate = {published},
tppubtype = {misc} }

We generalise the problem of reward modelling (RM) for reinforcement learning (RL) to handle non-Markovian rewards. Existing work assumes that human evaluators observe each step in a trajectory independently when providing feedback on agent behaviour. In this work, we remove this assumption, extending RM to include hidden state information that captures temporal dependencies in human assessment of trajectories. We then show how RM can be approached as a multiple instance learning (MIL) problem, and develop new MIL models that are able to capture the time dependencies in labelled trajectories. We demonstrate on a range of RL tasks that our novel MIL models can reconstruct reward functions to a high level of accuracy, and that they provide interpretable learnt hidden information that can be used to train high-performing agent policies.

https://eprints.soton.ac.uk/458023/

@article{soton458023,
title = {Non-markovian reward modelling from trajectory labels via interpretable multiple instance learning},
author = {Joseph Early and Tom Bewley and Christine Evers and Sarvapali Ramchurn},
url = {https://eprints.soton.ac.uk/458023/},
year = {2022},
date = {2022-05-01},
journal = {arXiv},
abstract = {We generalise the problem of reward modelling (RM) for reinforcement learning (RL) to handle non-Markovian rewards. Existing work assumes that human evaluators observe each step in a trajectory independently when providing feedback on agent behaviour. In this work, we remove this assumption, extending RM to include hidden state information that captures temporal dependencies in human assessment of trajectories. We then show how RM can be approached as a multiple instance learning (MIL) problem, and develop new MIL models that are able to capture the time dependencies in labelled trajectories. We demonstrate on a range of RL tasks that our novel MIL models can reconstruct reward functions to a high level of accuracy, and that they provide interpretable learnt hidden information that can be used to train high-performing agent policies.},
note = {20 pages (9 main content; 2 references; 9 appendix). 11 figures (8 main content; 3 appendix)},
keywords = {},
pubstate = {published},
tppubtype = {article} }

Greenhouse gas (GHG) emission reduction is an important and necessary goal; currently, different policies to reduce GHG emissions in maritime transport are being discussed. Amongst policies, like carbon taxes or carbon intensity targets, it is hard to determine which policies can successfully reduce GHG emissions while allowing the industry to be profitable. We introduce an agent-based maritime transport simulator to investigate the effectiveness of two decarbonisation policies by simulating a maritime transport operator?s trade pattern and fleet make-up changes as a reaction to taxation and fixed targets. This first of its kind simulator allows to compare and quantify the difference of carbon reduction policies and how they affect shipping operations.

https://eprints.soton.ac.uk/456716/

@inproceedings{soton456716,
title = {An agent-based simulator for maritime transport decarbonisation: Demonstration track},
author = {Jan Buermann and Dimitar Georgiev and Enrico Gerding and Lewis Hill and Obaid Malik and Alexandru Pop and Matthew Pun and Sarvapali Ramchurn and Elliot Salisbury and Ivan Stojanovic},
url = {https://eprints.soton.ac.uk/456716/},
year = {2022},
date = {2022-05-01},
booktitle = {21st International Conference on Autonomous Agents and Multiagent Systems (09/05/22 - 13/05/22)},
pages = {1890--1892},
abstract = {Greenhouse gas (GHG) emission reduction is an important and necessary goal; currently, different policies to reduce GHG emissions in maritime transport are being discussed. Amongst policies, like carbon taxes or carbon intensity targets, it is hard to determine which policies can successfully reduce GHG emissions while allowing the industry to be profitable. We introduce an agent-based maritime transport simulator to investigate the effectiveness of two decarbonisation policies by simulating a maritime transport operator?s trade pattern and fleet make-up changes as a reaction to taxation and fixed targets. This first of its kind simulator allows to compare and quantify the difference of carbon reduction policies and how they affect shipping operations.},
keywords = {},
pubstate = {published},
tppubtype = {inproceedings} }