BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Singapore
X-LIC-LOCATION:Asia/Singapore
BEGIN:STANDARD
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
TZNAME:SGT
DTSTART:19820101T123000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20201212T050107Z
LOCATION:Zoom Room 2
DTSTART;TZID=Asia/Singapore:20201211T100600
DTEND;TZID=Asia/Singapore:20201211T101200
UID:siggraphasia_SIGGRAPH Asia 2020_sess129_papers_218@linklings.com
SUMMARY:Neural Control Variates
DESCRIPTION:Technical Papers, Technical Papers Q&A\n\nNeural Control Varia
tes\n\nMüller, Rousselle, Keller, Novák\n\nWe propose neural control varia
tes (NCV) for unbiased variance reduction in parametric Monte Carlo integr
ation. So far, the core challenge of applying the method of control variat
es has been finding a good approximation of the integrand that is cheap to
integrate. We show that a set of neural networks can face that challenge:
a normalizing flow that approximates the shape of the integrand and anoth
er neural network that infers the solution of the integral equation. We al
so propose to leverage a neural importance sampler to estimate the differe
nce between the original integrand and the learned control variate. To opt
imize the resulting parametric estimator, we derive a theoretically optima
l, variance-minimizing loss function, and propose an alternative, composit
e loss for stable online training in practice. When applied to light trans
port simulation, neural control variates are capable of matching the state
-of-the-art performance of other unbiased approaches, while providing mean
s to develop more performant, practical solutions. Specifically, we show t
hat the learned light-field approximation is of sufficient quality for hig
h-order bounces, allowing us to omit the error correction and thereby dram
atically reduce the noise at the cost of negligible visible bias.\n\nRegis
tration Category: Ultimate Supporter, Ultimate Attendee
END:VEVENT
END:VCALENDAR