Centre for Computational Statistics and Machine Learning

1
Centre for Computational Statistics and Machine
Learning

• Internal Launch
• John Shawe-Taylor
• December, 2006

2
Context for the Centre

• Explosion of data coming available in all areas
  of science and commerce
• Exciting new approaches to analysing data
  developed in statistics and machine learning
• Often fragmented reinvention of known
  techniques, lack of awareness of complementary
  approaches, etc.
• UCL has very strong groups in statistics,
  computer science and at the Gatsby unit

3
Aims of the new centre

• Create a framework within which the different
  groups can cooperate synergistically
• Build bridges to the many groups at UCL that
  could benefit from advanced data analysis
• Develop collaborations with commercial,
  industrial and scientific groups beyond UCL
• Part of a drive towards handling large data for
  which there is no pre-defined analysis technique
• Pioneer a new synthesis between traditional
  statistics and new machine learning approaches

4
Emphasis of orientation

• Well-founded approaches to data analysis and
  machine learning
• Statistically analysed
• Bayesian approaches
• Frequentist analysis
• Understanding why and when the different
  approaches are more appropriate and developing
  the theoretical models to match practical
  applications
• Aim to become internationally recognised for the
  theoretical foundations and innovative
  applications of computational statistics and
  machine learning

5
Story so far

• Idea arose out of a review of the Gatsby Unit in
  2005
• Appointment of Director in July 2006
• Massi Pontil awarded a 5 year advanced research
  fellowship by EPSRC
• All three partners (Computer Science, Statistics
  and the Gatsby Unit) have contributed towards
  funding for initial period of 3 years

• Appointment of part-time manager and admin
  assistant in November 2006
• Web site under development
• Joint application to Wellcome Trust with
  Bloomsbury Bioinformatics Centre
• Internal launch today
• External launch in April 2007

6
Current members

• Gatsby Unit
• Maneesh Sahani
• Peter Dayan
• Yee Whye Teh arrives in January
• One more position currently advertised
• Statistics
• Philip Dawid
• Tom Fearns
• Trevor Sweeting

• Computer Science
• Mark Herbster
• Massimiliano Pontil
• John Shawe-Taylor
• David Barber just appointed
• Support
• Jacky Pallace, centre manager
• Clare Scurfield, admin support
• Tom Diethe, web development and support

7
Planned activities

• Main seminar series first by Peter Grunwald
  visiting statistics in February 2007?
• Subgroup seminar series at Gatsby, Statistics,
  and Computer Science to be promoted centrally
• Reading groups created spontaneously with
  activities visible through the web site
• Workshops on specialist themes eg open house on
  complex output learning in July 2006
• Visitor programme to ensure exploit potential to
  attract international contacts for short
  stop-overs and longer visits
• Open problems web-site, etc.
• MSc Programme in Intelligent systems (Dir Mark)
  may develop new related programmes in future

8
Planned activities clinics

• New approach to building collaborations with
  users from inside and outside UCL
• Half day event at which a potential user will
  present their problem and describe their data if
  necessary covered by an NDA
• Representatives of different approaches supported
  within the centre kernel methods, statistics,
  Bayesian inference, etc, discuss options for
  analysis
• Centre able to finance PhDs to assist with
  initial pump-priming implementation of
  interesting solutions
• Potential for joint grant applications with
  appropriate IPR agreements

9
Web site design

• Active web site providing portal for all dealings
  with the centre
• Entry point for new contacts can learn about
  what approaches can do, how to register for
  activities including clinics
• Listings for seminars, reading groups, etc.
• Easy to enter and register for new events on-line
• Automatic generation of emails to interested
  parties
• Supported by database storing activities and
  contacts, etc.

10
Presentations of the three component groups

• Computer Science group JS-T
• Statistics group Philip Dawid
• Gatsby Computational Neuroscience Unit Maneesh
  Sahani

11
Current Computer Science group

• Mark Herbster
• expert in analysis of on-line learning algorithms
  in an adversarial setting
• Currently investigating graph based learning
• Massi Pontil
• Expert in learning theory (particularly kernel
  methods, regularisation, convex optimisation)
• EPSRC advanced research fellow with associated
  project investigating multiple task learning
• John Shawe-Taylor
• Author of two books on support vector machines
  and kernel methods
• Theoretical foundations of learning plus
  applications in document analysis and computer
  vision
• Additional members
• David Barber (details later
• Charles Micchelli (Expert in Computational
  Mathematics

12
Current projects
13

• EU Network of Excellence
• Principled machine learning, linked to
  statistics, optimisation and applications
• computer vision
• Brain computer interfaces
• natural language processing and textual
  information access
• Multimodal integration
• speech
• JS-T is scientific coordinator (Southampton
  coordinating site)
• Linking 56 sites across Europe, Israel and
  Australia
• Innovative management style and broad impact

14
VISDEM

• Inference in Complex Stochastic Environmental
  Models
• EPSRC funded multi-site project
• Involving Manfred Opper from TU Berlin, Aston,
  Surrey and Met office reps
• Bayesian inference for stochastic differential
  equations modelling uncertainty and finer scale
  effects
• Generating interest in the Met office
• Many other potential applications of the
  technology

15

• Learning the Structure of Music (LeStruM) EPSRC
  funded project with University of Plymouth,
  Leibniz Institute of Neurobiology and Johannes
  Kepler University Linz
• Making fMRI and EEG scans of subjects listening
  to music
• Relating structure of responses to structure in
  the music
• Insight into how music creates effects and
  possibly how the mind works?

16
SMART

• Statistical Multilingual Analysis for Retrieval
  and Translation (SMART) EU funded STREP
• 8 site project lead by Xerox research labs in
  Grenoble
• Apply recent advances in machine learning to
  statistical machine translation and cross-lingual
  retrieval
• Particularly correlation analysis and
  discriminative kernel methods for structured
  output learning together with on-line adaptation
  of translation models

17
Advanced Research Fellowship

• 5 year EPSRC Advanced Research Fellowship for
  Massimiliano Pontil
• Includes a standard EPSRC project
• Learning multiple tasks
• Theoretical analysis and practical implementations

18
New recruit David Barber

• Expert in Bayesian inference over graphical
  models
• Has been working in modelling discrete dynamical
  systems
• Will join the department during 2007

19
Statistics Group
20
Gatsby Computational Neuroscience Unit
21
Summary

• Centre aims to set the agenda for the fundamental
  analysis in computational statistics and machine
  learning
• Can provide eclectic support covering the main
  well-founded approaches to data analysis
• Hence providing innovative and informed advice
  for users with a wide variety of problems and
  data from science or commerce
• Registration on website to be activated soon
• First clinic to be launched in the new year
• External launch in April 2007
• But drinks NOW!