David Barber : Main - Home Page browse

Current Highlights

Generating Sentences Using a Dynamic Canvas accepted in AAAI 2018.

Overdispersed Variational Autoencoders IJCNN 2017. Supplementary Material and code.
Complementary Sum Sampling for Likelihood Approximation in Large Scale Classification in AISTATS 2017. A very simple but effective way to train large classifiers (such as in language models).
JMLR paper Approximate Newton Methods for Policy Search in Markov Decision Processes
Nesterov's Accelerated Gradient Descent and related algorithms arxiv paper.
How to deal with a large number of classes in softmax models arxiv paper.
Some very simple Julia code for AutoDiff. Self explanatory I hope from the demos. See also note.
A note and slides on AutoDiff, parameter tying in Deep Learning (Neural Nets) and Backprop Through Time
Some preliminary ideas on solving ODEs using Gaussian Processes arxiv paper.

Using Gaussian Processes for Parameter Estimation in Ordinary Differential Equations ICML 2014 paper
Variational Optimisation ESANN paper arxiv paper.
A simple demo for string matching (a kind of Aho-Corasick algorithm for noisy sequences). Matlab Code and explanation of the method.
Inference and estimation in probabilistic time series models by D. Barber, A. T. Cemgil and S. Chiappa in Bayesian Time Series Models.
JMLR paper on Gaussian Kullback-Leibler Approximate Inference which describes much of Ed Challis' recently completed PhD work. There is also extensive software available that demonstrates Bayesian inference on very large-scale models, including sparse regression and logistic regression.
Chris Bracegirdle's successfully defended thesis on Inference in Bayesian Time-Series Models. This includes a thorough description of our work on efficient inference in switching linear models and more recently on Bayesian Conditional Cointegration.
A tech report on Variational Optimization
A note on Occam's razor
A note on quickly calculating your nearest neighbour, with demo code. Updated to include KD tree discussion and code.
NIPS 2012 ORAL paper A Unifying Perspective of Parametric Policy Search Methods for Markov Decision Processes supplementary material. Perhaps the world's greatest tetris player?! See the video and Tom's site for more of our work on MDPs.
NIPS 2012 paper on Affine Independent Variational Inference supplementary material.
ICML 2012 paper on Bayesian Cointegration supplementary material (runner up best paper)
An Approximate Newton Method for Markov Decision Processes
Promotional Discount on Bayesian Timeseries models. Note that you can change your locale in the checkout cart to pay and deliver in your local country.
Promotional discount on Bayesian Reasoning and Machine Learning. Note that you can change your locale in the checkout cart to pay and deliver in your local country.
Cambridge Science Festival talk on How to Engineer Intelligence.
New books and our popular MSc programme:

Recent research highlights

On the computational complexity of stochastic controller optimization in POMDPs
Just a bit of fun with Mona Lisa -- reconstructing an image (left) using 20000 randomly placed lines (centre) and 2000 circles (right). Each line/circle is just one colour. See here for an explanation. [click on images for higher resolution]:

ECML paper on Lagrange Dual Decomposition for Finite Horizon Markov Decision Processes
UAI paper on Efficient Inference in Markov Control Problems

AISTATS paper on Concave Gaussian Variational Approximations for Inference in Large-Scale Bayesian Linear Models supplementary material software
AISTATS paper on Switch-Reset Models : Exact and Approximate Inference supplementary material
Signal Processing Magazine article on Graphical Models for Time Series
I'm promoting an old TR that I never made very visible (apologies for a minor typo in the information filter -- I'll upload a corrected version shortly):

The auxiliary variable trick for deriving Kalman smoothers The derivation explains how to express the backward (beta) message in a form that is numerically stable.

Paper on Variational methods for Reinforcement Learning
Identifying graph clusters using variational inference and links to covariance parametrization Please email me if you would like a reprint.
LEMIR workshop paper on Solving deterministic policy (PO)MPDs using Expectation-Maximisation and Antifreeze

MSC in Computational Statistics and Machine Learning

The MSC in Computational Statistics and Machine Learning is an exciting course that imparts key skills for analyzing our data rich world, including techniques for Information Retrieval and Machine Learning. We anticipate graduates from the programme to take key positions in research and leading organizations involved with large-scale information processing and analysis. The programme is taught by world-renowned researchers in Machine Learning from UCL Computer Science and the Gatsby Computational Neuroscience Unit.