This page is used to maintain information about our regular meetings, with links to relevant papers and other resources.
Venue: 11am - 1pm Colloquium room (5A.540) unless otherwise stated
Meetings will start with a quick, round the table update from the NLIP group, before the topic of the meeting which will start at 11.30am. Meetings are intended to last 1 hour (until 12.30pm but the room is booked until 1pm if we want to overrun).
Summer Term 2020
due to the pandemic, all meetings this term will be held via teleconference
22 Apr title - presenter (affiliation)
29 Apr title - presenter (affiliation)
6 May title - presenter (affiliation)
13 May title - presenter (affiliation)
20 May title - presenter (affiliation)
27 May title tbc - Kakia Chatsiou (University of Essex)
3 Jun title - presenter (affiliation)
Spring Term 2020
26 Oct - colloquium room (5A.540) - Automated Multi-Document Text Summarization from Heterogeneous Data Sources - Mahsa Abazari Kia (CSEE, University of Essex)
Abstract:
Many companies such as British Telecom have significant workforces of field engineers, desk-based agents and customer Support services who collect and manage large volumes of unstructured and semi-structured information. The information about orders in this company in structured and free text format is captured by and stored in different internal systems, which are being used by teams handling the orders.
This volume of text about orders is an invaluable source of information which needs to be effectively summarized in a way that can help desk agents to have a clear picture of the order journey instead of check the order information from several places for the time that a customer calls in to find out about the progress of their order.
11 Mar - colloquium room (5A.540) - Generating descriptions of objects in visual scenes - Micha Elsner (Departmet of Psychology, University of Essex)
Abstract:
I will be discussing joint work with Dr. Alasdair Clarke (Psychology) on how to generate descriptions of objects in visual scenes. When humans generate these descriptions, they must plan out what to say jointly with their processing of the scene. This means that cognitive models of the process should take into account the visual system along with language. I will describe some psycholinguistic work we have done to understand how this process works as well as some preliminary work on using reinforcement learning to model it.
13 Mar - 6pm LAC & NLIP group joint social, Wivenhoe (the Black Buoy)
25 Mar - colloquium room (5A.540)
8 Apr - colloquium room (5A.540)
22 Apr - colloquium room (5A.540)
6 May - colloquium room (5A.540)
20 May - colloquium room (5A.540)
3 June - colloquium room (5A.540)
Autumn Term 2019
9 Oct - No meeting
16 Oct - 8K room (1NW.5.12) - LAC reading group (Kakia)
23 Oct - Networks Centre Meeting Room (1N1.3.2b) - Marine Technology Research Unit (Jon)
30 Oct - 8K room (1NW.5.12) - Paper review process - Interactive session aimed at 3rd year/MSc/PhD students on how papers are reviewed for conferences and workshops (Jon)
6 Nov - 8K room (1NW.5.12) - LAC reading group (Kakia) - Deep Learning NLP trends and tools
13 Nov - 8K room (1NW.5.12) - Computer Vision PhD seminar (Alba)
20 Nov - 8K room (1NW.5.12) - Graph homomorphisms (David Richerby)
27 Nov - 8K room (1NW.5.12) - TBD
4 Dec - Colloquium room (5A.540) - Computer Vision PhD seminar (Alba)
11am 11 Dec - Colloquium room (5A.540) - Contextualization and semantic compositionality in vector representations. (Marcos Garcia, University of A Coruña)
Abstract:
Current language models, such as ELMo or BERT, provide different vector representations of a given word depending on the contexts in which it occurs, so they seem interesting to analyse polysemy and idiomacity. In this talk I will present some ongoing work that evaluates the ability of these models to capture semantic compositionality in various languages.
4pm 11 Dec - Departmental seminar (1N1.4.1)
Better Together: Classifying Galaxies by Combining Galaxy Zoo Volunteers and Bayesian CNN (Mike Walmsley, Oxford University)
Abstract:
Modern telescopes image far more galaxies than any astronomer could hope to look through. Crowdsourcing project Galaxy Zoo has recruited hundreds of thousands of volunteers to help classify galaxies - but even this is no longer enough. To scale Galaxy Zoo, we apply Bayesian CNNs to make probabilistic galaxy classifications suitable for science. Further, using our probabilistic classifications, we apply active learning to request volunteer responses for the subset of galaxies which, if labelled, would be most informative for training our CNNs. We show that training using active learning requires up to 35-60% fewer labelled galaxies. By combining human and machine intelligence, Galaxy Zoo will be able to classify surveys of any conceivable scale on a timescale of weeks, providing massive and detailed classification catalogues to support research into galaxy evolution.
Mike Walmsley (@mike_w_ai) works on combining crowdsourcing and machine learning to investigate astronomical data. His research includes scaling Galaxy Zoo to classify millions of galaxies, identifying supermassive black holes in distant galaxies for space telescope Euclid, and improved detection of unexplained extragalactic millisecond pulses for radio interferometer CHIME. Mike is an astrophysics PhD student at Oxford and a visiting student at Cambridge.
18 Dec - Colloquium room (5A.540) - Xmas drinks? Plan for Spring term?