Research Theme: (Applied) Machine Learning

Metric learning for dynamical models
A machine learning pathway to early dementia diagnosis
BIGHARVI - Big data harvesting
Multilinear classification
Vehicle classification from inductive loop signature
Recent work in action recognition has exposed the limitations of methods which directly classify local features extracted from spatio-temporal video volumes. In opposition, encoding the actions' dynamics via generative dynamical models has a number of attractive features: however, using all-purpose distances for their classification does not necessarily deliver good results. We propose a general framework for learning distance functions for generative dynamical models, given a training set of labelled videos. The optimal distance function is selected among a family of pullback ones, induced by a parameterised automorphism of the space of models. We focus here on hidden Markov models and their manifold, and design an appropriate automorphism there. Experimental results are presented which show how pullback learning greatly improves action recognition performances with respect to base distances.
Early recognition of the various forms of dementia is a difficult task, as mild symptoms and other health issues with similar traits can complicate their clinical evaluation.
We propose an integrated framework for the early discriminative assessment and monitoring of dementia, based on the development of novel machine learning methodologies for classifying time series and unconstrained videos, and their application to both daily activities and laboratory tasks involving a combination of gait and cognitive tasks, which have demonstrated an ability to discriminate in early stages. Data are captured from multiple cheap and unobtrusive sensors (a smartphone's accelerometer, gyroscope and camera, and a Kinect-style range camera), as part of a sustainable approach to accurate and reliable diagnosis and severity assessment. An easy-to-use smartphone app is designed for data gathering, providing a clear route for engaging with the NHS.
The number of companies active in big data analysis is already large and keeps growing. Most of such businesses work in isolation, without much cooperation between partners within the market. The best results of in terms of analysis and prediction, however, come from combining the results of different analytical tools.
This project is motivated by the need of our industrial partners in unifying usage of data analysis tools and methods which should allow plugging in new functionally (for example, algorithms and components developed by third parties).
The aim of the platform is to provide applications with an environment, which will enable the applications to expand new geographical markets together with the platform if new data sources will be exploited by the platform. For supporting this transition a shared application translation manager is provided. The proposed solutions will be delivered as pluggable components of the open harvesting platform.
In most real-world problem however, observations are influenced by a number of nuisance factors. To tackle their influence, it is natural to resort to multi-linear or "tensorial" decompositions: approaches such as Higher-Order SVD have indeed been formulated to address the problem of decomposing a tensor into its constituent factors.
We show how HOSVD can be exploited to formulate a natural generalization of Tenenbaum's bilinear classifiers, which we call 'multilinear classifiers', able to classify observations depending on one content label and several style labels. A set of style-specific linear maps are learned by HOSVD of the training set, represented as a tensor. When a new observation in a different combination of styles is presented, EM is applied to alternative learn a new style matrix and classify the content of the observation. This approach is validated on the UCF gait ID dataset, demonstrating how explicitly modelling the different nuisance factors delivers superior performances.
Inductive loops are sensors that are widely deployed on road networks for the purpose of traffic data collection. Our aim is to classify vehicles in a 10 category scheme such as the SWISS10 from inductive loop signals. This level of classification is proving problematic with simple peak and valley detection algorithms.
Vehicle-classification is a small player in the field of machine-learning, and literature published so-far only addresses a handful number of classes. Equipment in the field runs on low-power embedded platforms, typically ARMv4T architecture, clocked at 8MHz, which delivers in the order of 10 MIPS.
We looked at two machine-learning algorithms: Support Vector Machines and Adaptive Boosting with decision stumps. We used the two most common algorithms for multiclass classification, One-versus-One and One-versus-Rest, and we looked at addressing class-imbalance with Under-sampling, Oversampling and SMOTE.