School of Computing Science

Reinforcement Learning from Reformulations in Conversational Question Answering over Knowledge Graphs

Group: Information Retrieval (IR)
Speaker: Magdalena Kaiser, Max Planck Institute for Informatics
Date: 25 October, 2021
Time: 15:00 - 16:00
Location: Zoom link will be provided after registration

Abstract:
The rise of personal assistants has made conversational question answering (ConvQA) a very popular mechanism for user-system interaction. State-of-the-art methods for ConvQA over knowledge graphs (KGs) can only learn from crisp question-answer pairs found in popular benchmarks. In reality, however, such training data is hard to come by: users would rarely mark answers explicitly as correct or wrong. In this work, we take a step towards a more natural learning paradigm - from noisy and implicit feedback via question reformulations. A reformulation is likely to be triggered by an incorrect system response, whereas a new follow-up question could be a positive signal on the previous turn's answer. We present a reinforcement learning model, termed CONQUER, that can learn from a conversational stream of questions and reformulations. CONQUER models the answering process as multiple agents walking in parallel on the KG, where the walks are determined by actions sampled using a policy network. This policy network takes the question along with the conversational context as inputs and is trained via noisy rewards obtained from the reformulation likelihood. To evaluate CONQUER, we create and release ConvRef, a benchmark with about 11k natural conversations containing around 205k reformulations. Experiments show that CONQUER successfully learns to answer conversational questions from noisy reward signals, significantly improving over a state-of-the-art baseline. 

Bio:
Magdalena Kaiser is a PhD Student in the Databases and Information Systems Group at the Max Planck Institute for Informatics (MPII), Saarbrücken, Germany, under the supervision of Prof. Gerhard Weikum and Dr. Rishiraj Saha Roy. Her research focuses on conversational question answering. In particular, she is interested in leveraging feedback to improve conversational systems. In her work, she applies techniques from Information Retrieval, Natural Language Processing and Machine Learning, particularly Reinforcement Learning. Further information can be found on her website: http://people.mpi-inf.mpg.de/~mkaiser/.

FATA Seminar : Verification of Hybrid Systems using SpaceEx

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Douglas Fraser, University of Glasgow
Date: 26 October, 2021
Time: 13:00 - 14:00
Location: Contact organizers for Zoom coordinates

Model checking is a very common topic within our group; a Finite State Machine (FSM) is used to describe the behaviours of a system and a model checking tool is then used to examine these behaviours for fault identification or safety verification purposes. While a very powerful technique, it can be hard to represent some systems as state variables can only be changed during the transitions between states. In some recent work I’ve been looking at Hybrid Automata, an extension of FSMs, that allows for continuous behaviours to be included within the model through ordinary differential equations that change the value of state variables as time elapses within state. SpaceEx is a model checking tool that is able to analyse the state space of systems represented by these models and identify the presence of forbidden behaviours using a highly accurate over-approximation model. The presentation will include a quick recap of basic Finite State Machines and an explanation of how these are extended to become Hybrid Automata. I’ll then show how Hybrid Automata are created within the SpaceEx model editor and how the tool can be used to examine their behaviours. I’ll then conclude with some of my own thoughts on SpaceEx, what I’ve found to be challenging and what I think it could offer the members of the group.

Reinforcement Learning from Reformulations in Conversational Question Answering over Knowledge Graphs

Group: Information Retrieval (IR)
Speaker: Magdalena Kaiser, Max Planck Institute for Informatics
Date: 25 October, 2021
Time: 15:00 - 16:00
Location: Zoom link will be provided after registration

Abstract:
The rise of personal assistants has made conversational question answering (ConvQA) a very popular mechanism for user-system interaction. State-of-the-art methods for ConvQA over knowledge graphs (KGs) can only learn from crisp question-answer pairs found in popular benchmarks. In reality, however, such training data is hard to come by: users would rarely mark answers explicitly as correct or wrong. In this work, we take a step towards a more natural learning paradigm - from noisy and implicit feedback via question reformulations. A reformulation is likely to be triggered by an incorrect system response, whereas a new follow-up question could be a positive signal on the previous turn's answer. We present a reinforcement learning model, termed CONQUER, that can learn from a conversational stream of questions and reformulations. CONQUER models the answering process as multiple agents walking in parallel on the KG, where the walks are determined by actions sampled using a policy network. This policy network takes the question along with the conversational context as inputs and is trained via noisy rewards obtained from the reformulation likelihood. To evaluate CONQUER, we create and release ConvRef, a benchmark with about 11k natural conversations containing around 205k reformulations. Experiments show that CONQUER successfully learns to answer conversational questions from noisy reward signals, significantly improving over a state-of-the-art baseline. 

Bio:
Magdalena Kaiser is a PhD Student in the Databases and Information Systems Group at the Max Planck Institute for Informatics (MPII), Saarbrücken, Germany, under the supervision of Prof. Gerhard Weikum and Dr. Rishiraj Saha Roy. Her research focuses on conversational question answering. In particular, she is interested in leveraging feedback to improve conversational systems. In her work, she applies techniques from Information Retrieval, Natural Language Processing and Machine Learning, particularly Reinforcement Learning. Further information can be found on her website: http://people.mpi-inf.mpg.de/~mkaiser/.

FATA Seminar : Verification of Hybrid Systems using SpaceEx

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Douglas Fraser, University of Glasgow
Date: 26 October, 2021
Time: 13:00 - 14:00
Location: Contact organizers for Zoom coordinates

Model checking is a very common topic within our group; a Finite State Machine (FSM) is used to describe the behaviours of a system and a model checking tool is then used to examine these behaviours for fault identification or safety verification purposes. While a very powerful technique, it can be hard to represent some systems as state variables can only be changed during the transitions between states. In some recent work I’ve been looking at Hybrid Automata, an extension of FSMs, that allows for continuous behaviours to be included within the model through ordinary differential equations that change the value of state variables as time elapses within state. SpaceEx is a model checking tool that is able to analyse the state space of systems represented by these models and identify the presence of forbidden behaviours using a highly accurate over-approximation model. The presentation will include a quick recap of basic Finite State Machines and an explanation of how these are extended to become Hybrid Automata. I’ll then show how Hybrid Automata are created within the SpaceEx model editor and how the tool can be used to examine their behaviours. I’ll then conclude with some of my own thoughts on SpaceEx, what I’ve found to be challenging and what I think it could offer the members of the group.

FATA Seminar double feature: "Simulation and Model Checking for Close to Real-time Overtaking Planning" and "Observable and Attention-Directing BDI Agents for Human-Autonomy Teaming"

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Prof. Alice Miller, Dr Mengwei Xu, University of Glasgow
Date: 02 November, 2021
Time: 13:00 - 14:00
Location: Contact organizers for Zoom coordinates

Double feature!

Abstract: Fast and reliable trajectory planning is a key requirement of autonomous vehicles. In this paper we introduce a novel technique for planning the route of an autonomous vehicle on a straight rural road using the Spin model checker. We show how we can combine Spin's ability to identify paths violating temporal properties with sensor information from a 3D Unity simulation of an autonomous vehicle, to plan and perform consecutive overtaking manoeuvres on a traffic-heavy road. This involves discretising the sensory information and combining multiple sequential Spin models with a Linear Time Temporal Logic specification to generate an error path. This path provides the autonomous vehicle with an action plan. The entire process takes place in close to real-time (using no pre-computed data). Our experiments demonstrate that the simulated autonomous vehicle implementing our approach can drive on average at least 40 km and overtake 214 vehicles before experiencing a collision - which is usually caused by inaccuracies in the sensory system. While the proposed system has some drawbacks, we believe that our novel approach demonstrates a potentially powerful future tool for efficient trajectory planning for autonomous vehicles.

Deep learning for medical image analysis

Group: Computing Technologies for Healthcare
Speaker: Dr. Alison ONeil, Canon Medical Research Europe Ltd.
Date: 02 November, 2021
Time: 15:00 - 16:00
Location: https://uofglasgow.zoom.us/meeting/register/tJUuf-2prD0sG91BDJPMlW9HKqBz6itv8CgP

Radiologists are under pressure to handle an ever-increasing volume of medical imaging. In the meantime, deep learning solutions have demonstrated impressive performance for imaging tasks of classification, segmentation and translation, opening up opportunities for AI to assist and automate imaging workflows. Training deep learning solutions for real-world healthcare applications involves ethical, legal and practical performance considerations. As we look to scale deep learning to a wide range of clinical applications, we also need to find ways to reduce reliance on pixel-level expert annotations whilst retaining clinically acceptable accuracy. This talk will cover some of Canon’s research in this area.

Biography:

Dr Alison Q O’Neil is a Principal Scientist in the AI Research Team at Canon Medical Research Europe and Honorary Research Fellow at the University of Edinburgh. She leads an AI Research team at Canon who focus on imaging and natural language processing for healthcare problems. Her research interests span techniques for learning with less, multimodal learning, knowledge informed machine learning, and causality.

 Registration is required to attend the event: https://uofglasgow.zoom.us/meeting/register/tJUuf-2prD0sG91BDJPMlW9HKqBz6itv8CgP 

Security aspects of data plane programmability

Group: Systems Seminars
Speaker: Sandra Scott-Hayward, Queen's University, Belfast
Date: 09 November, 2021
Time: 14:00 - 15:00
Location: Zoom

About Sandra Scott-Hayward:

Sandra Scott-Hayward is a Senior Lecturer (Associate Professor) with the School of Electronics, Electrical Engineering and Computer Science, and a Member of the Centre for Secure Information Technologies at Queen’s University Belfast (QUB). She began her career in industry and became a Chartered Engineer in 2006 having worked as a Systems Engineer and Engineering Group Leader with Airbus. Since joining academia, she has published a series of IEEE/ACM papers on security designs and solutions for softwarized networks based on her research on the development of network security architectures and security functions for emerging networks. She received Outstanding Technical Contributor and Outstanding Leadership awards from the Open Networking Foundation in 2015 and 2016, respectively, having been elected and serving as the Vice-Chair of the ONF Security Working Group from 2015 to 2017. Amongst many other service memberships, she is on the Organizing Committee of IEEE NetSoft and is an Associate Editor of IEEE Transactions on Network and Service Management. She is Director of the QUB Academic Centre of Excellence in Cyber Security Education (ACE-CSE), one of the first universities to be awarded this recognition by the U.K. National Cyber Security Centre. She has been selected as a Polymath Fellow of the Global Fellowship Initiative at the Geneva Centre for Security Policy (GCSP) from 2021 to 2023.

Title TBA

Group: Programming Languages at University of Glasgow (PLUG)
Speaker: Ethel Tshukudu, University of Glasgow
Date: 09 November, 2021
Time: 15:00 - 16:00
Location: Online

Ethel Tshukudu from the the Centre for Computing Science Education at UoG will be giving a talk. Full details will be available nearer the time.

Location

Due to the ongoing impact of the COVID-19 pandemic, PLUG talks will take place online via Zoom. The details will be sent via the PLUG mailing list and the SPLS Zulip instance nearer the time.

More information about PLUGs mailing list can be found here:http://www.dcs.gla.ac.uk/plug/

As with previous Wireless PLUGs, please:

1. join promptly at the advertised start of the session, and we will begin on the academic hour;
2. start with your cameras turned off and microphones muted; and
3. ensure you can be identified by your username.

Questions

If the speaker is happy with questions during the talk, please use the chat or 'raise hand' feature. Should the speaker not see these interrupts, the chair will actively interrupt.

Move over, MSE! – New probabilistic models of motion

Group: School of Computing Science
Speaker: Gustav Eje Henter, KTH Royal Institute of Technology
Date: 11 November, 2021
Time: 15:00 - 16:00
Location: Sir Alwyn Williams Building, 422 Seminar Room

Data-driven character animation holds great promise for enhancing realism and creativity in games, film, virtual avatars and social robots. However, due to the high bar on visual quality, most existing AI animation solutions focus narrowly on a specific task, and do not generalise to different motion types.

This talk makes the case that 1) machine learning now has advanced far enough that strong, task-agnostic motion models are possible, and that 2) these models should be probabilistic in nature, to accommodate the great diversity of human behaviour. We present MoGlow, a new, award-winning deep-learning architecture that leverages normalising flows and satisfies our two criteria. MoGlow is competitive with the state-of-the-art in locomotion generation for humans and dogs, as well as in speech-driven gesture generation. We also combine MoGlow with our research in synthesising spontaneous-sounding speech to make a virtual character walk, talk, and gesticulate from text input alone.

For a longer introduction showing our models in action, please see the following YouTube videos:
Locomotion synthesis: https://youtu.be/pe-YTvavbtA
Co-speech gesture generation: https://youtu.be/egf3tjbWBQE
Synthesis of coherent speech and motion: https://youtu.be/4_Gq9rU_yWg

Speaker bio:
Gustav Eje Henter is a WASP assistant professor in machine learning at the Division of Speech, Music and Hearing at KTH Royal Institute of Technology. His main research interests are probabilistic modelling and deep learning for data-generation tasks, most prominently speech and
motion/animation synthesis. He has an MSc and a PhD from KTH, followed by post-docs in speech synthesis at the Centre for Speech Technology Research at the University of Edinburgh, UK, and in Prof. Junichi Yamagishi’s lab at the National Institute of Informatics, Tokyo, Japan, before returning to KTH in 2018.

This meeting will be held in a "hybrid" mode: up to 20 people can attend in-person in the room, and the meeting will also be streamed online on Zoom. Please log in to access the Zoom link. https://uofglasgow.zoom.us/j/92967222374?pwd=dGZPMHJQRWoxaUJHQlV3bCt5RjlDdz09

FATA Seminar : TBD

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Dr Viktor Zamaraev, University of Liverpool
Date: 16 November, 2021
Time: 13:00 - 14:00
Location: Contact organizers for Zoom coordinates

TBD

TBA

Group: Systems Seminars
Speaker: Peter Pietzuch, Imperial College, London
Date: 16 November, 2021
Time: 14:00 - 15:00
Location: Zoom

TBA

FATA Seminar : TBD

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Dr Ben Bumpus, Eindhoven University of Technology
Date: 23 November, 2021
Time: 13:00 - 14:00
Location: Contact organizers for Zoom coordinates

TBD

Title TBA

Group: Programming Languages at University of Glasgow (PLUG)
Speaker: David Castro-Perez, University of Glasgow
Date: 23 November, 2021
Time: 15:00 - 16:00
Location: Online

David Castro-Perez from the University of Kent will be giving a talk. Full details will be available nearer the time.

Location

Due to the ongoing impact of the COVID-19 pandemic, PLUG talks will take place online via Zoom. The details will be sent via the PLUG mailing list and the SPLS Zulip instance nearer the time.

More information about PLUGs mailing list can be found here:http://www.dcs.gla.ac.uk/plug/

As with previous Wireless PLUGs, please:

1. join promptly at the advertised start of the session, and we will begin on the academic hour;
2. start with your cameras turned off and microphones muted; and
3. ensure you can be identified by your username.

Questions

If the speaker is happy with questions during the talk, please use the chat or 'raise hand' feature. Should the speaker not see these interrupts, the chair will actively interrupt.

FATA Seminar : TBD

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Dr. Johannes Fichte, University of Potsdam
Date: 30 November, 2021
Time: 13:00 - 14:00
Location: Contact organizers for Zoom coordinates

TBD

HealthCare Seminar: TBC

Group: Computing Technologies for Healthcare
Speaker: Prof. Nazli Goharian, Georgetown University
Date: 01 December, 2021
Time: 15:00 - 16:00
Location: https://uofglasgow.zoom.us/meeting/register/tJMtdOGtrT4iE9aX2cDxzQ6TbisiGWEn7Y-M

Biography

Nazli Goharian is Clinical Professor of Computer Science at Georgetown University, and Associate Director of the Information Retrieval Lab . Prior to joining Georgetown, she was Clinical Associate Professor of Computer Science and a member of the Information Retrieval Laboratory at the Illinois Institute of Technology (IIT). She joined academia from industry, where she was the Project Manager for Data Warehouse Conferencing at MCI Worldcom.

Her research, publications, and doctoral student mentorship span the domains of information retrieval, text mining, and natural language processing. Specifically, her interest lies on humane-computing applications; as such, she has been focusing on text processing in medical/health domain. She has graduated four doctoral students with strong publication records and placement in the workplace (i.e., Senior Scientist at Max Planck Institute for Informatics, Research Scientist at the Allan Institute for Artificial Intelligence, Applied Scientist at Amazon Alexa Search, and Director of Data Science at DataXu). Her doctoral student's dissertation was selected for the 2019 Glassman Distingushed Dissertation Award in Sciences at Georgetown University. She in collaboration with her doctoral students has a Best Long Paper Award at EMNLP 2017; and an Area Chair Favorite in COLING 2018.

Registration is required to attend:  https://uofglasgow.zoom.us/meeting/register/tJMtdOGtrT4iE9aX2cDxzQ6TbisiGWEn7Y-M

FATA Seminar : A Constraint Programming Solver You Can Trust (But Don't Have To)

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Dr. Ciaran McCreesh, University of Glasgow
Date: 07 December, 2021
Time: 13:00 - 14:00
Location: Contact organizers for Zoom coordinates

I'm writing a new constraint programming solver. It's not the fastest solver in the world (yet), it doesn't support every global constraint (yet), and it only works with integer variables (so far), but it does have one feature that no other solver does: it can produce an auditable record of any problem it solves. This record consists of a formal definition of the problem, the solution found, and a mathematical proof that the solution is correct. This proof can be verified independently, using a simple piece of software that does not know anything about constraint programming. In this talk I'll give a very quick introduction to constraint programming via examples, and then talk about how my solver solves these problems. I'll also talk about what we have to implement in a solver to support proof logging, and try to justify why I think you should be able to trust my solver -- or rather, why proof logging means you don't need to trust it.

TBA

Group: Systems Seminars
Speaker: Natalia Chechina, Erlang Solutions; University of Glasgow
Date: 14 December, 2021
Time: 14:00 - 15:00
Location: Zoom

TBA

Cyber Physical Systems & Cyber Security Research Themes: SICSA Event at IEEE DSC 2022

Group: Scottish Informatics and Computer Science Alliance (SICSA)
Speaker: SICSA Event, SICSA
Date: 22 June, 2022
Time: 09:00 - 17:00
Location: Edinburgh Napier University, Craiglockhart Campus, 219 Colinton Road, Edinburgh , EH14 1DJ, United Kingdom

In parallel with IEEE DSC 2022, we are running an event organised by Dr Naghmeh Moradpoor and sponsored by SICSA to support women in cyber given that the lack of women in cybersecurity is something that the UK government, the industry and university are keen to address. Therefore, it aims to inspire women into cybersecurity roles by bringing together the cyber women from across the UK in the form of a poster and a paper competition. Having the best paper and best poster competition held in parallel with the above conference, will assist women in cybersecurity to understand the demands for cybersecurity/digital forensics in different organisations. This event not only connects women in cybersecurity to leading industrial experts but also provides for greater collaboration amongst Scottish universities and leading industries. We hope that the delegates will engage with industrial contacts for future partnerships and research-based industrial collaborations. Prizes The event will select total of x8 winners (x4 winners for the best papers and x4 winners for the best posters) as follows: Prizes for x4 winners (total: £230), best posters: First place: X1 £100 Amazon voucher; X1 Certificate from the event Second place: X1 £50 Amazon voucher; X1 Certificate from the event Third place: X1 £30 Amazon voucher; X1 Certificate from the event Outstanding woman in cyber (poster category): X1 £50 Amazon voucher; X1 Certificate from the event Prizes for x4 winners (total: £230), best papers: First place: X1 £100 Amazon voucher; X1 Certificate from the event Second place: X1 £50 Amazon voucher; X1 Certificate from the event Third place: X1 £30 Amazon voucher; X1 Certificate from the event Outstanding woman in cyber (paper category): X1 £50 Amazon voucher; X1 Certificate from the event Please note that both males and females can enter the above competitions for best papers and best posters, competing for first place, second place, and third place. Only females are allowed in the Outstanding Woman in Cyber (paper & poster) categories. Contact Dr Naghmeh Moradpoor, Edinburgh Napier University