Introduction

An important component of an intelligent system is its ability to adapt to differing user needs and environments. NICTA's Statistical Machine Learning (SML) program researches methods of creating intelligent devices with the ability to learn. Ultimately, the aim is to build intelligent systems that adapt to user needs without needing a programmer to encode rules about how to act.

The systems researchers are constructing collect data from their environment, extract knowledge from data, and respond in an intelligent manner. The SML program aims to develop ICT products, processes and mechanisms that are increasingly usable; that hide sophisticated and complex processes behind simpler interfaces; that make use of information in vast databases; and that adapt to different environments and users.

For example, researchers are working to develop systems whereby devices are able to learn how to recognise a user's voice or handwriting. These devices should learn how their environment works by analysing and understanding large sets of data.  They should also learn how to interact with their environment in order to reach the objectives they have been assigned.

The primary areas of interest to the program are kernel methods and statistics; rapid stochastic gradient methods; reinforcement learning and planning; and bioinformatics.

Kernel Methods and Statistics

Kernels Methods are systems that describe similarities between objects, for example, the similarity of two faces or the similarity between the pen strokes used to write a letter or character. The resulting information is used to find efficient methods of identifying a person (in the case of face recognition), or a digit (in the case of optical character recognition). Kernel methods provide some unique ways of detecting unusual and novel items in a set of data.

Support Vector Machines (SVMs) is a method of classifying data using convex optimisation methods. In particular, text classification, optical character recognition, bioinformatics, and natural language processing.

Rapid Stochastic Gradient Methods

Gradient descent methods provide the engines that drive much machine learning. To cope with the flood of data we find ourselves in today, it often becomes necessary to approximate gradients from subsamples of data. Unfortunately the noise this introduces into the gradient is not tolerated well by the classical gradient methods - with the exception of steepest descent, which however is very slow to converge.

We have used local step size adaptation to accelerate the convergence of stochastic gradient descent, culminating in our recently developed stochastic meta-descent (SMD) algorithm. SMD not only tolerates noisy gradients, but approaches the rapid convergence of a seond-order method at low computational cost, allowing us to scale the algorithm to extremely large optimization problems.

Reinforcement Learning and Planning

Programming a machine to achieve complex tasks requires a considerable investment of resources. Reinforcement learning (RL) aims to create machines that program themselves to achieve specific goals. The designer defines goals with reward signals, which reinforce the desired behaviours of the machine.

The study of planning involves learning by way of models that describe a particular problem. Learning refers to behavioural learning by trial and error. The algorithms employed in learning and planning must consider the fundamental randomness of nature and, therefore, that actions do not always have predictable outcomes.

Applications include robotics, operations planning, resource management, logistics, and operation research.

Document Analysis and Understanding

The information revolution has led to an the explosive growth of unstructured documents available for analysis. Much information is also available in the context of law enforcement and intelligence scenarios. Such abundance can lead to a glut, meaning there is a need to understand the raw information and transform it into usable knowledge. Analysis and understanding will be accomplished via methods based on natural language processing and state of the art non-parametric statistics, such as kernels on structured data, Conditional Random Fields (CRFs), and low-dimensional data representation.

Bioinformatics

We concentrate on applications of machine learning techniques for building predictive models for determintaion of phenotype from genomic profiles. Moreover we work on methods for diagnosing cancer based on DNA microarray measurements.