BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Isaac Newton Institute Seminar Series
SUMMARY:PCF-GAN and beyond:  High-Fidelity Generative Mode
 ls for Synthetic Time Series Generation - Hao Ni (
 University College London)
DTSTART;TZID=Europe/London:20240904T133000
DTEND;TZID=Europe/London:20240904T141500
UID:TALK217603AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/217603
DESCRIPTION:Generating high-fidelity time series data using ge
 nerative adversarial networks (GANs) remains a cha
 llenging task\, as it is difficult to capture the 
 temporal dependence of joint probability distribut
 ions induced by time-series data. To this end\, a 
 key step is the development of an effective discri
 minator to distinguish between time series distrib
 utions.&nbsp\;In this&nbsp\;talk\, I will introduc
 e&nbsp\;the so-called&nbsp\;PCF-GAN\, a novel GAN 
 that incorporates the path characteristic function
  (PCF) as the principled representation of time se
 ries distribution into the discriminator to enhanc
 e its generative performance.&nbsp\; On the one ha
 nd\, we establish theoretical foundations of the&n
 bsp\;PCF&nbsp\;distance by proving its characteris
 ticity\, boundedness\, differentiability with resp
 ect to generator parameters\, and weak continuity\
 , which ensure the stability and feasibility of tr
 aining the&nbsp\;PCF-GAN. On the other hand\, we d
 esign efficient initialisation and optimisation sc
 hemes for&nbsp\;PCFs to strengthen the discriminat
 ive power and accelerate training efficiency. To f
 urther boost the capabilities of complex time seri
 es generation\, we integrate the auto-encoder stru
 cture via sequential embedding into the&nbsp\;PCF-
 GAN\, which provides additional reconstruction fun
 ctionality. Extensive numerical experiments on var
 ious datasets demonstrate the consistently superio
 r performance of&nbsp\;PCF-GAN over state-of-the-a
 rt baselines\, in both generation and reconstructi
 on quality. Lastly\, an application of&nbsp\;PCF-G
 AN to Levy area generation is presented\, which sh
 ows its potential to accelerate the high-order SDE
  simulation.\nThis&nbsp\;talk&nbsp\;is based on tw
 o papers:&nbsp\;[https://arxiv.org/pdf/2305.12511.
 pdf] (joint work with Siran Li (Shanghai Jiao Tong
  University) and Hang Lou (UCL) ) and [https://arx
 iv.org/pdf/2308.02452.pdf] (Joint work with Andraž
  Jelinčič (Oxford)\, Jiajie Tao (UCL)\, William F 
 Turner (Imperial College)\, Thomas Cass (Imperial 
 College)\, James Foster (Oxford)).
LOCATION:Seminar Room 1\, Newton Institute
CONTACT:
END:VEVENT
END:VCALENDAR