Dermatologically Inspired Deep Face Analytics

Introduction
The ability to detect face ageing which is localised to specific regions of the face has shown to be useful to both the medical and research communities. In medical practice it could be used to inform the diagnosis of conditions, and subsequently the speed and accuracy of screening. However, large-scale facial screening by human experts is laborious and requires specialised expertise. Hence, there is great value in the development of automated methods for feature extraction and decision making from facial Imagery in a health and well-being context

Explaining the Science
The main challenge in training any machine learning problem is data. In order to train a deep neural network to predict something accurately, there must be a large number of known data points. When approaching the field of ageing, there is no dataset in existence labelling the rate at which different face regions have aged.

Our approach to this lack of data is to build on the work of others who have developed models for global face ageing and semantic face segmentation. Semantic face segmentation is a field which splits face regions at a pixel level. We intend to learn from both existing datasets of segmented faces and datasets containing global face ages. The combination of these two types of data provide a framework to develop novel new methods for grading the ageing of faces at a feature level.

Aside from grading the uneven rate of ageing across the human face, we also access a novel dataset of elderly individuals who have died since the images were taken. We use these faces in combination with the number of years they lived following the photograph, to attempt to extract features relevant to their mortality.

Project aims
The aims and objectives of this project are interdisciplinary at their core. Ranging from novel contributions in data annotation, to neural network architecture; to dermatological analyses. We intend to explore the active research area of ageing and vitality, covering gaps in existing research and propose a new deep learning model to analyse the heterogeneous ageing of human faces. While there is existing work detailing both human and computational methods for grading face ageing, little research has been done to apply deep learning to this problem. We aim to train convolutional neural networks to extract features related to age, health and ultimately mortality.

Applications
Our work is highly applicable to several practical and research areas, with many more possible applications materialising as the wealth of academic data grows.

The ability to accurately extract age and health related features from face images allows Dermatological research to be carried on sample sizes which are several orders of magnitude larger than without. Presently clinical trials for dermatological treatment are run in labs, with expensive photography equipment and highly skilled reviewers needed to evaluate outcomes. Given an effectively trained neural network, these analyses can be carried out automatically using selfie images from mobile phones; allowing participants to be located anywhere globally. The benefits of remote face skin analysis also extend to every-day practice, where pre-appointment screening can be carried out to filter patients who may not need to see a doctor immediately.

Given a mortality feature extractor with a good correlation to the endpoint of death, it is possible to implement screening systems in high risk locations such as GP surgeries and Care Homes. These systems could be designed to flag individuals with a higher relative mortality risk for their age, allowing them to receive care in a timely fashion.

Estimating individual pig liveweight trajectories from group-level weight measurements

Introduction
Recent years have seen an increase in the use of advanced data collection systems to improve management in pig farms, however, the cost of implementing such systems remains a barrier to many. Some of the most useful data for farm management are on an individual pig level, which requires investment of labour and expense of systems such as RFID tagging. Machine learning approaches can help to alleviate this issue for instance by estimating individual growth trajectories from unknown weight measurements without requiring expensive RFID systems.

Explaining the science
Machine learning methods can be used to estimate individual-level pig growth trajectories as a cheaper alternative to using RFID tag systems. However, the use of traditional machine learning methods in this task has several limitations, namely that they cannot incorporate all data on a group-level time series simultaneously. This limitation could have a potentially profound impact on their predictive accuracy and thus also their applicability.

Project aims
This project aims to develop more accurate methods for the prediction of individual-level growth trajectories. To achieve this, attention methods capable of integrating all contextual data within group-level time series will be developed, and comparisons between these methods and existing machine learning methods will be made.

This project will also involve the exploration of attention weights produced by the attention methods to allow for the interpretation of resulting predictions. In addition, this project seeks to develop additional metrics for evaluating the accuracy of predicted growth trajectories, and to improve the ability of methods to generalize to pigs of different breeds than those that they were trained on.

Applications
The methods developed during this project have the potential to be used as a significantly cheaper alternative to RFID tagging systems for the tracking of individual-level pig weights when combined with either weighting platforms or camera-based pig weight estimation systems. This work could also be used as a foundation for the development of methods that can be applied to other livestock.

Semantic segmentation using a small, labelled dataset and larger, unlabelled dataset, producing predictions that incorporate uncertainty

Introduction
A lack of high-quality volumetric medical imaging ground truth can limit the performance of neural networks, particularly around generalization to unseen data. As well as this semantic segmentation predictions lack uncertainty measures in the general case.

Explaining the Science
This project uses volumetric MRI data, resulting in memory-intensive, multi-day training runs, and trains networks in four separate configurations. Tuning requires a large number of hyper-parameters to balance supervised and unsupervised tasks. As such, access to Jade and its high-end GPUs is essential to the successful completion of this project.

Project aims
The approach was to develop a network whose properties should enable it to learn in a semi-supervised fashion from a small, labelled dataset and a larger dataset, such as is typically available in clinical settings, and that can produce uncertainty in labelling by generating plausible variations through generative mechanisms.

Applications
Datasets typically available in clinical settings.

Increasing the training efficiency of sophisticated function of mobile robots

Introduction
With the development of mobile robots, increasingly sophisticated functions are required in order to satisfy diverse application requirements. This work is likely to be the first method that involves a fuzzy system in the graph neural networks’ attention mechanism in order to effectively employ a spatial relational inductive bias.

Explaining the Science
In this project, JADE is used to conduct reinforcement learning experiments based on customized OpenAI gym environments and FAM-GNN PPO algorithms. Several parallel environments are used as inputs to decrease the physical sampling time, and observations are then fed into the systems and CUDA in order to realise GPU training. The JADE HPC system increases the training efficiency and accelerates the experiment stage.

Project aims
As a sub-research domain of path planning, obstacle avoidance is one of the compulsory functionalities for ensuring the safety of both robots and environment. Inspired by existing work and in order to fill the research gap in obstacle avoidance applications, the project proposes a fuzzy attention mechanism (FAM) based graph neural network (GNN) architecture in order to realise representation learning based proximal policy optimisation (PPO) reinforcement learning structure.

Applications
Obstacle avoidance applications to ensure the safety of both robots and the environment.

Sucess Stories

Dr. Gangapurwala’s work within Oxford University’s Robotics Institute was demonstrated on a live Chevron facility:
Oxford quadruped robot research-tested on a live Chevron facility
Additional information on Dr. S. Gangapurwala’s work:
Legged robot navigates by learning from its mistakes

We are always looking for new ideas and feedback.

Any questions or comments, please report it via GitHub issue tracker.

GitHub