BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Chicago
X-LIC-LOCATION:America/Chicago
BEGIN:DAYLIGHT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
TZNAME:CDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
TZNAME:CST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260202T201806Z
LOCATION:231
DTSTART;TZID=America/Chicago:20251116T113000
DTEND;TZID=America/Chicago:20251116T115000
UID:submissions.supercomputing.org_SC25_sess223_ws_scalah103@linklings.com
SUMMARY:Efficient Embedding Initialization via Dominant Eigenvector Projec
 tions
DESCRIPTION:Quentin Petit (Mines Paris - PSL University France); Chong Li 
 (Huawei Technologies France); Nahid Emad (Maison de la Simulation, Univers
 ity of Paris-Saclay); and Jack Dongarra (University of Tennessee, Knoxvill
 e)\n\nThe embedding layer is essential in deep learning, transforming high
 -dimensional data into compact representations. However, growing datasets 
 and model sizes pose challenges in training time, memory, and generalizati
 on. We propose a scalable method for embedding initialization via spectral
  dimensionality reduction using dominant eigenvector projections.\n\nThe p
 roposed approach leverages on MIRAMns, multiple implicitly restarted Arnol
 di method with nested subspaces, to extract most informative directions fr
 om large and potentially sparse data representations. Unlike traditional e
 mbeddings or autoencoders, this proposed approach requires few tunable par
 ameters and is inherently parallel. We apply MIRAMns to matrix representat
 ions such as covariance and co-occurrence matrices to compute low-dimensio
 nal embeddings that preserve data structure and variance. Experiments acro
 ss diverse datasets show that the proposed method achieves comparable or b
 etter accuracy with significantly reduced dimensionality, enabling smaller
 , faster deep networks. Additionally, our parallel implementation scales e
 fficiently on HPC platforms, making it well-suited for large-scale scienti
 fic and AI workloads.\n\nRecording: Livestreamed, Recorded\n\nRegistration
  Category: Technical Program Reg Pass, Workshop Reg Pass\n\nSession Chairs
 : Vassil Alexandrov (Hartree Centre, STFC); Jack Dongarra (University of T
 ennessee, Knoxville; Oak Ridge National Laboratory (ORNL)); Erik Draeger (
 Lawrence Livermore National Laboratory (LLNL), Center for Applied Scientif
 ic Computing); Philippa Rubin (STFC Hartree Centre); Dieter A. Kranzlmuell
 er (Ludwig-Maxmilians-Universität München, Leibniz Supercomputing Centre (
 LRZ)); and Christian Engelmann (Oak Ridge National Laboratory (ORNL))\n\n
END:VEVENT
END:VCALENDAR