BEGIN:VCALENDAR PRODID:-//Microsoft Corporation//Outlook 16.0 MIMEDIR//EN VERSION:2.0 METHOD:REQUEST X-MS-OLK-FORCEINSPECTOROPEN:TRUE BEGIN:VTIMEZONE TZID:GMT Standard Time BEGIN:STANDARD DTSTART:16011028T020000 RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10 TZOFFSETFROM:+0100 TZOFFSETTO:-0000 END:STANDARD BEGIN:DAYLIGHT DTSTART:16010325T010000 RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3 TZOFFSETFROM:-0000 TZOFFSETTO:+0100 END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT CLASS:PUBLIC CREATED:20260924T150627Z DESCRIPTION: Title: A forward-backward stochastic algorithm for (tomographi c) inverse problems\n Speakers: Tatiana Bubba & Ilaria Trombini (Departmen t of Mathematics and Computer Science\, University of Ferrara\, Italy)\nAb stract: Stochastic optimization is pivotal to deep learning\, enabling eff icient training of large--scale models through mini-batch updates and prov iding a fast\, memory-efficient strategy able to reach good solutions well before theoretical convergence. \nRecently\, stochastic optimization has gained attention in inverse problems. In particular\, stochastic gradient methods have shown promising results in imaging applications\, where large -scale operators make full updates impractical. \nIn this talk\, we intr oduce a forward--backward (FB) stochastic gradient method\, called FB-LISA \, specifically designed for addressing large-scale inverse problems and b uilt on Deep-LISA\, a stochastic algorithm originally proposed for smooth finite-sum optimization. Its main feature is the adaptive hyperparameter strategy\, combining line-search-based learning rate estimation with a pre determined increase in mini-batch size. Furthermore\, FB-LISA naturally ha ndles non-smooth objective terms\, enabling the use of regularizers and co nstraints common in imaging. Numerical experiments on experimental 3D CT d ata demonstrate that FB-LISA significantly outperforms several state-of-th e-art reconstruction methods.\n \nCheck: https://ccpi.ac.uk/events/ccpi-sh ow-tell-and-user-support-drop-in/ for the upcoming schedule.\n \n_________ _______________________________________________________________________\nM icrosoft Teams Need help? \nJ oin the meeting now \nMeeting ID: 323 425 655 245 47 \nPasscode: Kv6qM7QU \n________________________________\n\nDial in by p hone \n+44 20 7660 6791\,\,939266504# United Kingdom\, City of London\nFin d a local number \nPhone conference ID: 939 266 504# \nJoin o n a video conferencing device \nTenant key: teams@stfc.onpexip.com \nVideo ID: 122 561 499 3 \nMore info \nFor organizers: Meeting optio ns | Reset dial-in PIN \n_______________________________________ _________________________________________\n \n \n \n \n DTEND;TZID="GMT Standard Time":20261021T140000 DTSTAMP:20260924T150627Z DTSTART;TZID="GMT Standard Time":20261021T130000 LAST-MODIFIED:20260924T150627Z LOCATION:https://teams.microsoft.com/meet/32342565524547?p=87XuoJRSwMCiZn4G ks ORGANIZER;CN="Murgatroyd, Laura (STFC,RAL,SC)":mailto:laura.murgatroyd@stfc .ac.uk PRIORITY:5 SEQUENCE:0 SUMMARY;LANGUAGE=en-gb:CCPi Show & Tell: A forward-backward stochastic algo rithm for (tomographic) inverse problems TRANSP:OPAQUE UID:040000008200E00074C5B7101A82E00800000000D0A48B5B3C4CDD01000000000000000 010000000FF95A35556EA984C8E528622AA5AECFE X-ALT-DESC;FMTTYPE=text/html:

\;Title : \;A forward-backward stochastic algorithm for (tomographic) inve rse problems

\; Speakers: \;Tatiana Bubba &\; Ilaria Trombini (Department of Mathematics and Computer Science\, University of Ferrara\, Italy)

Abstract: \;Stoc hastic optimization is pivotal to deep learning\, enabling efficient train ing of large--scale models \;through mini-batch updates \;and prov iding a fast\, memory-efficient strategy able to reach good solutions well before theoretical convergence. \;

Recently\, stochastic optimization has gained attentio n in inverse problems. In particular\, stochastic gradient methods have sh own promising results in imaging applications\, where large-scale operator s make full updates impractical. \;

In this talk\, we \; \;introduce a forward--b ackward (FB) stochastic gradient method\, called FB-LISA\, specifically de signed for addressing large-scale inverse problems and built on Deep-LISA\ , \;a stochastic algorithm originally proposed for smooth finite-sum optimization. Its main feature is the \;adaptive hyperparameter strate gy\, combining line-search-based learning rate estimation with a predeterm ined increase in mini-batch size. \;Furthermore\, \;FB-LISA natura lly handles non-smooth objective terms\, \;enabling the use of regularizers and constraints common in imaging. \;N umerical experiments on experimental 3D CT data demonstrate that FB-LISA s ignificantly outperforms several state-of-the-art reconstruction methods.< o:p>

\;

Check: https://ccpi.ac.uk/ events/ccpi-show-tell-and-user-support-drop-in/ \;for the upcoming schedule.

\;

________________________________________________________________ ________________

Microsoft Teams < u>Need help?

Join the meeting now

Meeting ID: 323 425 655 245 47

Passcode: Kv6qM7QU

Dial in by phone < /p>

+44 20 7660 6791\,\,939266504# United Kingdom\, City of London

Find a local number

Phone conference ID: 939 266 504#

Join on a video conferencing device

Tenant key: teams@stfc.onpexip.com

Video ID: 122 561 499 3

For organizers: Meeting options | Reset dial-in PIN< /o:p>

______________________________________________ __________________________________

& nbsp\;

\;

\;

\;

X-MICROSOFT-CDO-BUSYSTATUS:BUSY X-MICROSOFT-CDO-IMPORTANCE:1 X-MICROSOFT-DISALLOW-COUNTER:FALSE X-MS-OLK-AUTOFILLLOCATION:FALSE X-MS-OLK-CONFTYPE:0 BEGIN:VALARM TRIGGER:-PT15M ACTION:DISPLAY DESCRIPTION:Reminder END:VALARM END:VEVENT END:VCALENDAR