BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Vector Quantization in Deep Neural Networks for Speech and Image P
 rocessing - Mohammad Vali\, Aalto  University\, Finland
DTSTART:20241111T120000Z
DTEND:20241111T130000Z
UID:TALK222601@talks.cam.ac.uk
CONTACT:Simon Webster McKnight
DESCRIPTION:Vector quantization (VQ) is a classic signal processing techni
 que that models the probability \ndensity function of a distribution using
  a set of representative vectors called codebook (or \ndictionary). Deep n
 eural networks (DNNs) are a branch of machine learning that has gained \np
 opularity in recent decades. Since VQ provides an abstract high-level disc
 rete representation of \na distribution\, it has been widely used in vario
 us DNN-based applications such as speech  recognition\, image generation\,
  and speech and video coding. Hence\, a small improvement in VQ \ncan sign
 ificantly boost the performance of many applications dealing with differen
 t data types\, such as speech\, image\, video\, and text.\nThis talk mainl
 y focuses on improving various VQ methods within deep learning frameworks\
 , including:\n1) Improvement in training: VQ is non-differentiable\, and t
 hus\, it cannot backpropagate gradients. We proposed a new solution to thi
 s issue that works better than state-of-the-art solutions\, such as Straig
 ht-Through Estimator and Exponential Moving Average.\n2) Improvement in In
 terpretability: With the combination of VQ and space-filling curves concep
 ts\, we proposed a new quantization technique called Space-Filling Vector 
 Quantization. This technique helps to interpret the latent spaces of DNNs.
 \n3) Improvement in Privacy: We used the Space-Filling Vector Quantization
  technique to cluster the speaker embeddings to enhance the speaker's priv
 acy in speech processing tools based on DNNs.
LOCATION:Online only: Zoom: https://cam-ac-uk.zoom.us/j/89623597387?pwd=Pa
 lnRtu2be5cw3aGReM6EyvfcMrcly.1
END:VEVENT
END:VCALENDAR