BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:PCF-GAN: generating sequential data via the characteristic functio
 n of measures on the path space - Hao Ni (University College London)
DTSTART:20240424T100000Z
DTEND:20240424T104500Z
UID:TALK214192@talks.cam.ac.uk
DESCRIPTION:Generating high-fidelity time series data using generative adv
 ersarial networks (GANs) remains a challenging task\, as it is difficult t
 o capture the temporal dependence of joint probability distributions induc
 ed by time-series data. To this end\, a key step is the development of an 
 effective discriminator to distinguish between time series distributions. 
 In this talk\, I will introduce the so-called PCF-GAN\, a novel GAN that i
 ncorporates the path characteristic function (PCF) as the principled repre
 sentation of time series distribution into the discriminator to enhance it
 s generative performance.&nbsp\; On the one hand\, we establish theoretica
 l foundations of the PCF distance by proving its characteristicity\, bound
 edness\, differentiability with respect to generator parameters\, and weak
  continuity\, which ensure the stability and feasibility of training the P
 CF-GAN. On the other hand\, we design efficient initialisation and optimis
 ation schemes for PCFs to strengthen the discriminative power and accelera
 te training efficiency. To further boost the capabilities of complex time 
 series generation\, we integrate the auto-encoder structure via sequential
  embedding into the PCF-GAN\, which provides additional reconstruction fun
 ctionality. Extensive numerical experiments on various datasets demonstrat
 e the consistently superior performance of PCF-GAN over state-of-the-art b
 aselines\, in both generation and reconstruction quality. Lastly\, an appl
 ication of PCF-GAN to Levy area generation is presented\, which shows its 
 potential to accelerate the high-order SDE simulation.\nThis talk is based
  on two papers: [https://arxiv.org/pdf/2305.12511.pdf] and [https://arxiv.
 org/pdf/2308.02452.pdf].
LOCATION:External
END:VEVENT
END:VCALENDAR