UCL HEP Seminars
Seminars are generally held at 4pm on Fridays in room A1 on the top floor of the physics department
A calendar of all seminars in the Physics Department
is available on the Physics Events Calendar page.
If you use Google Calendar or similar, such as Apple iCal, it is possible to subscribe to this calendar via: XML, ICAL or HTML.
Please send suggestions for topics and/or speakers to Linda Cremonesi and Andreas Korn.
26-01-2018 : Jonathan Davis (Kings)
CNO Neutrino Grand Prix: The race to solve the solar metallicity problem
Several next-generation experiments aim to make the first measurement of the neutrino flux from the Carbon-Nitrogen-Oxygen (CNO) solar fusion cycle. This will provide crucial new information for models of the Sun, which currently are not able to consistently explain both helioseismology data and the abundance of metal elements, such as carbon, in the solar photosphere. The solution to this solar metallicity problem may involve new models of solar diffusion or even the capture of light dark matter by the Sun. I look at how soon electronic-recoil experiments such as SNO+, Borexino and Argo will measure the CNO neutrino flux, and the challenges this involves. I also consider experiments looking for nuclear-recoils from CNO neutrinos, which requires sensitivity to very low energies, and discuss how the same technology is also key to direct searches for sub-GeV mass dark matter.
02-02-2018 : Yiannis Andreopoulos (UCL)
Deep Learning from Compressed Spatio-Temporal Representations of Data
Deep learning has allowed for (and incentivised) researchers to look at data volumes and processing tasks that have previously only been hypothesized. While the first-generation of deep supervised learning achieved significant advances over shallow learning methods, it is increasingly becoming obvious that many approaches were naive in their design and we are only scratching the surface of what is possible. The notion of strong supervision (i.e., the use of labels during training) is impractical and easy to fool by adversarial examples and, perhaps most importantly, operating with uncompressed samples (e.g., input image pixels of video or audio) does not scale. For instance, data generated from visual sensing in Internet-of-Things (IoT) application contexts will occupy more than 82% of all IP traffic by 2021, with one million minutes of video crossing the network every second [Cisco VNI Report, Jun. 2017]. This fact, in conjunction with the rapidly-increasing video resolutions and video format inflation (from standard to super-high definition, 3D, multiview, etc.), makes the scale-up of deep learning towards big video datasets unsustainable. To address this issue, we propose to go beyond the pixel representations and design advanced deep learning architectures for classification and retrieval systems that directly ingest compressed spatio-temporal activity bitstreams produced by: (i) mainstream video coders and (ii) neuromorphic vision sensing cameras. By exploiting the compressed nature of our inputs, our approach can deliver 100-fold increase in processing speed with comparable classification or retrieval accuracy to state-of-the-art pixel-domain systems and has the potential to be extended to self-supervised deep learning. The talk will explain the key steps of our approach and can motivate researchers to think carefully about the sensing and supervision modalities of their problems prior to embarking on the use of deep learning tools for data analysis. Related paper: https://arxiv.org/abs/1710.05112
09-02-2018 : Stefan Guindon (CERN)
16-02-2018 : Nassim Bozorgnia (Durham)
The dark halo of Milky Way-like galaxies
One of the major sources of uncertainty in the interpretation of dark matter direct and indirect detection data is due to the unknown astrophysical distribution of dark matter in the halo of our Galaxy. Realistic numerical simulations of galaxy formation including baryons have recently become possible, and provide important information on the properties of the dark matter halo. I will discuss the dark matter density and velocity distribution of Milky Way-like galaxies obtained from high resolution hydrodynamical simulations. To make reliable predictions for direct and indirect detection searches, we identify simulated galaxies which satisfy the Milky Way observational constraints. Using the dark matter distribution extracted from the selected Milky Way-like galaxies, I will present an analysis of current direct detection data, and discuss the implications for the dark matter interpretation of the Fermi GeV excess.