Skip to main content Skip to navigation

Deep (Machine) Learning for physicists 2020-21

Module Convenor: Rudo Roemer (Warwick University)

  • Times: Mondays 15-16 UK/16-17 EU and Wednesdays 15-16 UK/16-17 EU
  • 1st (online) lecture: January 11th, 2021
  • Course duration: MPAGS 5 weeks, CY Cergy-Paris Universite 5+5 weeks (including course work)

Join Zoom Meeting (Meeting ID: 944 9524 2929, Passcode: B59b2B), Slack discussion room

Lecture slides, Warwick course repository, Google Colaboratory

The course is for interested students of the Midlands Physics Alliance Graduate School and the CY Cergy-Paris Universite M2-level Postgraduate course 2021.

Machine learning and deep learning are statistical analysis techniques that use strategies of artificial intelligence to characterize complex data and extract deep information. In recent years, these techniques have begun to be used not only in traditional computer science test cases, but also in real world applications as well as, more recently, in areas of advanced physics. In condensed matter systems, such techniques have been shown to give useful insight into Ising and spin ice models [1], low dimensional topological systems [2], strongly correlated systems [3], as well as random two- and three-dimensional topological and non-topological systems [4]. It is probably fair to say that machine and deep learning may become standard statistical analysis tools in the future.

In this short course, we want to understand some of the basic principles underlying the recent successes of the DL approach to data analysis. We will do so in a hands-on manner, following up our theoretical understanding with examples using state-of-the-art DL packages such as Keras, FastAI, TensorFlow and PyTorch. In doing so, we will use Jupyter notebooks as frontend to Python. Participants are assumed to have their own DL platform at hand, although mostly modern laptop resources will be sufficient.

[1] J. Carrasquilla and R. G. Melko: Nature Physics 13 (2017) 431;  A. Tanaka and A. Tomiya: Journal of the Physical Society of Japan 86 (2017) 063001.

[2] Y. Zhang and E.-A. Kim: Phys. Rev. Lett. 118 (2017) 216401; P. Zhang, H. Shen, and H. Zhai: arXiv:1708.09401 (2017).

[3] G. Carleo and M. Troyer: Science 355 (2017) 602; P. Broecker, J. Carrasquilla, R. G. Melko, and S. Trebst: Scientific Reports 7 (2017) 8823; K. Ch’ng, J. Carrasquilla, R. G. Melko, and E. Khatami: Phys. Rev. X 7 (2017) 031038; L. Li, T. E. Baker, S. R. White, and K. Burke: Phys. Rev. B 94 (2016) 245129; E. P. van Nieuwenburg, Y.-H. Liu, and S. D. Huber: Nature Physics 13 (2017) 435; L. Huang and L. Wang: Phys. Rev. B 95 (2017) 035105; F. Schindler, N. Regnault, and T. Neupert: Phys. Rev. B 95 (2017) 45134; H. Saito: Journal of the Physical Society of Japan 86 (2017) 093001; H. Saito and M. Kato: arXiv:1709.05468 (2017); H. Fujita, Y. Nakagawa, S. Sugiura, and M. Oshikawa: arXiv:1705.05372 (2017).

[4] T. Ohtsuki and T. Ohtsuki: Journal of the Physical Society of Japan 85 (2016) 123706; T. Ohtsuki and T. Ohtsuki: Journal of the Physical Society of Japan 86 (2017) 044708. N. Yoshioka, Y. Akagi, and H. Katsura: arXiv:1709.05790 (2017)

Assessments (for MPAGS only)

Upload here! This list is for all those MPAGS assessments so that you know whether I have received your upload. Really, it means I had the time to see that an email I get from the files.warwick system has arrived and that I have downloaded it or indeed already marked. When all is done, it's with MPAGS formally and they might led you know whether you have pass/fail-ed. The done stage will not be reached until the deadline of March 19th has been passed, i.e. you have 5 weeks to work on this.

name date received topic as far as I understand it and can summarize it in one sentence here ;-) status (email/download/marking/done)
Joel Swallow 120221 Simulations of K+ decays at the NA62 experiment at CERN using FastAI passed
Joe Kelly 120221 B/W/images of human faces compressed for efficient storage and reconstructed using variational autoencoder [TF, Keras] passed
Billy Shrive 160221 Solely numpy based deep learning code from scratch passed
Matthew Nicholson 190221 Tests how well neural networks can cope if they have a particularly limited dataset, i.e. pictures of Matthew's two dogs. Answer: 500 images are not enough [PyTorch] passed
Chris Oliver 190221 classification of X-ray images of human bones into normal/abnormal using FastAI passed
Alun Rees 050321 scikit-learn, TF and Keras based prediction of NRL games passed
Clar-Brid Tohill 080321 keras+TF to distinguish healthy from disease-afflicted plants via photos of their leaves passed
Wanbing Ge 090321 PyTorch version of image recognition/classification using two different DL architectures passed
Daniel Cocking 100321 FastAI version of image recognition/classification for wild cats (lion, tiger, etc.) passed
Jack Henshaw 120321 FastAI version of image recognition/classification for bees, waps and other insects or other non-insects passed
Vilius Atkocius 150321 FastAI SuperRes-GAN based method to reconstruct background images from experimentally determined atomic densities profiles passed
Tishtrya Mehta 160321 FastAI version of image recognition/classification for more than 200 types of birds passed
Daniele Baldolini 170321 FastAI version of image recognition/classification for pictures of chess pieces as published in news paper chess columns, great increase in accuracy following data augmentation passed
Luke Smith 180321 FastAI code to distinguish birds from planes passed
Elizabeth Sharpe 190321 FastAI code to distinguish three macaw species passed
Gethin-Wyn Williams 190321 A FastAI tabular application training in the Titanic dataset from Kaggle passed
Lucas Rushton 190321 Crack and defect detection using Keras passed
Olly Smith 190321 FastAI code to learn reading images of road signs passed
Mihaela Marinescu 190321 FastAI code distinguish images of chihuahua dogs from those of muffins. Checkout this Chihuahua or muffin twitter post to see this is a non-trivial image recognition task - although very funny ;-) passed
Sebastian Coleman 190321 FastAI image recognition to distinguish three popular types of coffee. passed
Andreas Hadjigeorghiou 190321 Training of FastAI model for classifying major members of the canine family, including: wolves, jackals, foxes, african wild dogs and siberian huskies passed

[table last updated: 060421]