Associate Professor Marcus Frean
School of Engineering and Computer Science
Victoria University of Wellington
Log In
Marcus is an academic at Victoria University of Wellington with research interests in machine learning, complex systems, and evolution.

What's on this page:


Machine Learning

Anomaly detection in astronomical images. Working with the radio astronomy group at Victoria and David Hogg at NYU, we have developed new algorithms for detecting very faint distributed sources. Our approach is fast yet sensitive and robust. Joint work involving Tony Butler-Yeoman and Chris Hollitt in ECS. See also Anna Friedlander's MSc thesis and our paper in the MaximumEntropy conference.
dependentGPexample-tiny.png Optimization of expensive functions: Consider finding the highest point on some unknown surface - ie the problem of "non-convex optimization". The interesting problems are those in which putative solutions are expensive to obtain - afterall, all others can be solved by some variant of "brute force" search of random solutions! In domains where data is precious, it pays to build a model of the family of possible surfaces, and use that family to reason about the best place to try next. This raises all kinds of interesting questions: what family? how to learn it from limited data? what property of it to use to decide where to try next? Better answers to these lead automatically to better optimization algorithms. Current PhD student Mashall Aryan is developing this approach, working with Marcus and JP (John Lewis).
factorgraphExample.png Deep belief networks and Boltzmann Machines: there is a lot of topical interest in deep belief nets due to their fast learning algorithm. Our current work (with Stephen Marsland (Massey)) concerns inference and learning in multiple-cause deep belief nets. Previous MSc student Russell Tod (MSc student) looked at how spiking neurons might implement probability propagation.
Multi-target tracking: In work with Praveen Choppala and Paul Teal, we have developed new ways to apply particle filtering to the task of detecting and tracking multiple faint objects in datastreams characterised by low signal-to-noise ratio, and an unknown (and changing) number of targets.
Frean-Feild-hexapod.png Reinforcement learning: Tim Field (MSc student) used the policy gradient algorithm to train local policies for motor control; James Bebbington (ditto) looked into reinforcement learning of policies for active sensing by deep belief nets.


histograms.png How network structure affects the rate of evolution: The rate of evolution depends on the "shape" of the underlying network. This is work with Gareth Baxter (former postdoc at ECS, now in Portugal), Stephen Hartley (VUW's SBS) Paul Rainey (Massey Albany) and Arne Traulsen (Max Planck Institute). See "The effect of population structure on the rate of evolution", Marcus Frean, Paul Rainey and Arne Traulsen. In Proceedings B (Proceedings of the Royal Society, Biological) (2013), for example.
miles-of-smiles-CarinGoldberg.jpg Evolution of cooperation. (i) Early work examined the strong effects that details of spatial positioning and timing can have on the cooperative strategies that we might expect natural selection to come up with, under payoffs given by the Prisoner's Dilemma; (ii) Natural selection can warp mutually productive relationships into exploitative ones even to the point of enslavement (work with Edward Abraham) - this might explain why some endosymbionts cooperate with their hosts while others are parasitic; (iii) Humans are good at forming ad hoc cooperative groups (especially solutions to the Stag Hunt coordination game), and this relates to possible evolutionary drivers behind our strong religious tendencies (work with Joseph Bulbulia).
rsp.png Cyclic competitions in nature: With Edward Abraham I have studied rock-scissors-paper situations in ecological systems. Such systems show paradoxical behaviour - for example the slowest invader is the one most likely to survive while its competitors go extinct - a phenomenon we dubbed "survival of the weakest". Richard Mansfield extended these ideas significantly in his PhD thesis. Richard and I showed how the 3-way competition could arise from even simpler systems of just two species.
retinotectal.png Topographic projections in the brain, specifically the retinotectal projection. The mush in your head wires itself up by matching molecular cues between the innervating axons and their targets elsewhere in the brain. Despite a wealth of interesting data, there is still argument about exactly how this process of self-organisation occurs - particularly given its astonishing robustness to damage and other interference. Jerome Dolman completed an MSc exploring some of these ideas.


Grad students

Marcus has taught on the following courses:

  • COMP307 - 3rd year AI paper
  • COMP421 - Machine Learning
  • COMP103 - Introduction to Data Structures and Algorithms
  • COMP303 - occasional guest lectures
  • COMP101 - was an introduction to computer science taught via dynamic web design
  • SCIE401 - an Applied Bioinformatics course
  • COMP473 - reading course in complex adaptive systems (case by case basis)


This page is
Contact Us | Section Map | Disclaimer | RSS feed RSS FeedBack to top ^

Valid XHTML and CSS | Built on Foswiki

Page Updated: 24 Jun 2016 by marcus. © Victoria University of Wellington, New Zealand, unless otherwise stated