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:CUED Speech Group Seminars
SUMMARY:Vector Quantization in Deep Neural Networks for Sp
 eech and Image Processing - Mohammad Vali\, Aalto 
  University\, Finland
DTSTART;TZID=Europe/London:20241111T120000
DTEND;TZID=Europe/London:20241111T130000
UID:TALK222601AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/222601
DESCRIPTION:Vector quantization (VQ) is a classic signal proce
 ssing technique that models the probability \ndens
 ity function of a distribution using a set of repr
 esentative vectors called codebook (or \ndictionar
 y). Deep neural networks (DNNs) are a branch of ma
 chine learning that has gained \npopularity in rec
 ent decades. Since VQ provides an abstract high-le
 vel discrete representation of \na distribution\, 
 it has been widely used in various DNN-based appli
 cations such as speech  recognition\, image genera
 tion\, and speech and video coding. Hence\, a smal
 l improvement in VQ \ncan significantly boost the 
 performance of many applications dealing with diff
 erent data types\, such as speech\, image\, video\
 , and text.\nThis talk mainly focuses on improving
  various VQ methods within deep learning framework
 s\, including:\n1) Improvement in training: VQ is 
 non-differentiable\, and thus\, it cannot backprop
 agate gradients. We proposed a new solution to thi
 s issue that works better than state-of-the-art so
 lutions\, such as Straight-Through Estimator and E
 xponential Moving Average.\n2) Improvement in Inte
 rpretability: With the combination of VQ and space
 -filling curves concepts\, we proposed a new quant
 ization technique called Space-Filling Vector Quan
 tization. This technique helps to interpret the la
 tent spaces of DNNs.\n3) Improvement in Privacy: W
 e used the Space-Filling Vector Quantization techn
 ique to cluster the speaker embeddings to enhance 
 the speaker's privacy in speech processing tools b
 ased on DNNs.
LOCATION:Online only: Zoom: https://cam-ac-uk.zoom.us/j/896
 23597387?pwd=PalnRtu2be5cw3aGReM6EyvfcMrcly.1
CONTACT:Simon Webster McKnight
END:VEVENT
END:VCALENDAR