Rough Path and Machine Learning Research Group



The supplementary code repository and examples for our research papers showcase how to use our methods to solve real-world problems, e.g., recognising human movements from skeleton-based video data and detecting early sepsis from electronic health data.

Path development network


The path development network is a generic neural network architecture built on the Cartan development, originated from the Rough path theory. Essentially, the path development network lifts the time series to a matrix Lie group by learning the Lie group representation in a data-driven way. It can be applied to many sequential data applications, including but not limited to speech recognition, handwriting classification and N-body simulation. The empirical performances have shown effectiveness by exploiting the Lie group properties. The hybrid model staking the LSTM with the path development can improve the convergence of model training and achieves state-of-the-art on several sequential learning tasks. This repository (Github link) is the official implementation of the paper entitled "Path Development Network with Finite dimensional Lie Group Representation.

Real-time sepsis prediction pipeline

The early detection of sepsis is a key research priority to help facilitate timely intervention. Criteria used to identify the onset time of sepsis from health records vary, hindering comparison and progress in this field. The subtle variation of sepsis criteria may have a significant impact on the predictive performance of machine learning algorithms. In this repository (Zenodo link), we implemented several commonly used methods for labelling sepsis onset time on the publicly available ICU data -- MIMIC-III database, and three representative models for early sepsis detection. The three models are (1) Light gradient boosting machine, (2) Long short term memory and (3) Cox proportional-hazards models. Our codes provide a pipeline of early sepsis prediction including data downloading, sepsis onset labelling, feature extraction, model training and model evaluation. This repository is the official implementation of our paper entitled "Subtle Variation of Sepsis-III Definitions Influences Predictive Performance of Machine Learning".

Logsig-RNN for skeleton-based action recognition

Given sequences of human skeleton joints over time representing different human actions, the action recognition task aims to train a classifier to label the sequences with their corresponding action classes. The Logsig-RNN method is a combination of a log signature layer and a recurrent neural network, where the former uses log-signatures as representations for streamed data, which can manage high sample rate streams, non-uniform sampling and time series of variable length. The hybrid method outperforms the traditional RNN on the skeleton-based action recognition tasks. This repository (Github link) contains the implementation of the novel method invented in our paper entitled "Logsig-RNN: a novel network for robust and efficient skeleton-based action recognition".