Dr Rhiannon Falconmore (Vector Informatik) - The evolution of a PhD student: Academia to Industry (top)

From project inception to the viva and beyond, the path of each PhD student is completely unique, rewarding, daunting and especially non-linear. This talk covers the timeline of my academic evolution from 2017 (beginning of PhD) until now, including all the research I undertook, the ups (paper acceptances, conference experiences), the downs (burnout, COVID), why I chose to move into industry and what I do now (still a lot of research!).

Dr Tom Gur (Warwick) - How to write a research statement (top)

The importance of writing a strong research statement cannot be overstated. Yet despite the career-changing implications of these documents, there is often not enough discussion and guidance on how to approach writing them. In this talk, I will discuss how to write a strong and effective research statement and how to update and use it throughout one's academic career.

Theory and Foundations

This is Driving Me Loopy: Transforming `loop` in Arrowized Functional Reactive Programs (top)

by Finnbar Keating

Arrowized Functional Reactive Programming (AFRP) is a useful paradigm for writing reactive programs: programs which take in inputs throughout their execution and return outputs at the same rate to react to their environment. Its usefulness comes from being embedded in Haskell, allowing programs to utilise Haskell's expressive type system to check for a variety of runtime errors - which is key in domains like robotics where errors can be costly. Unfortunately, performance issues of AFRP means that it does not see much use. One such issue is that its `loop` construct is defined in a way that requires the execution order of a program to be worked out at runtime. We avoid this particular overhead with a novel program transformation which turns `loop` constructs into a restricted form with known execution order. We discuss the performance benefits and prove the correctness of this transformation.

Communication-avoiding Optimizations for Large-scale Unstructured-mesh Applications with OP2 (top)

by Suneth Ekanayake

We investigate data movement-reducing and communication-avoiding (CA) optimizations and their practicable implementation for large-scale unstructured-mesh applications. Utilizing the high-level abstraction of the OP2 DSL for unstructured-mesh based codes, we reason about techniques for reduced communications across a consecutive sequence of loops - a loop-chain for distributed-memory parallelization. A new CA back-end for OP2 is designed with an extension to operate on a cluster of GPUs, such that these techniques can automatically be applied to any OP2 application. The new CA back-end is applied to the loop chains of two non-trivial applications, including the OP2 version of Rolls-Royce’s production CFD application, Hydra. Performance is investigated on both CPU and GPU clusters using 8M and 24M node mesh sizes. The results demonstrate 30–50% runtime reductions for select configurations. We model and examine the determinants and characteristics of a given unstructured-mesh loop-chain providing insights into the general feasibility and profitability of CA optimizations.

OP-PIC – An Unstructured Mesh Particle in Cell Domain Specific Language (top)

by Zaman Lantra

Particle-in-cell (PIC) calculation is a well-established procedure for modelling the behaviour of charged particles in the presence of electric/magnetic fields. Performance portability of PIC applications on emerging massively parallel heterogeneous HPC hardware have become an increasing concern for scientific organizations relying on these applications. One solution is the development of high-level abstractions such as a domain-specific-language (DSL) that allows domain specialists to specify the problem while leaving the implementation to a lower level through automatic code generation techniques, translating the specification to hardware specific parallelisations such as OpenMP, MPI, CUDA, etc. DSLs for PIC codes, especially unstructured-meshed PIC have not been well developed in the HPC community. In this work, we present a new DSL for unstructured PIC calculations, using the source-to-source translation and code generation techniques developed in the OP2 DSL (a DSL for unstructured-mesh algorithms), together with some preliminary results of single-node sequential, OpenMP and CUDA parallel versions.

Performance Portable Multiphase Flow Solutions with Discontinuous Galerkin Methods (top)

by Toby Flynn

In this work we present a performance portable solver toolkit based on the OP2 DSL, targeted at developing multiphase flow simulations. The new toolkit brings together the state-of-the-art in multiphase flow solutions based on Discontinuous Galerkin (DG) methods for achieving high-order accuracy within OP2's established performance portable framework. A key aspect of the modelling is how the fluid interface can be managed through high-order methods while maintaining the necessary stability, accuracy and performance. The new OP2-DG methods are used for developing a number of applications, including a non-trivial 2-fluid liquid whistle problem. The OP2-generated code for these problems are benchmarked, validated and compared to experimental results on a wide range of multi-core and many-core systems demonstrating near-optimal performance and scaling.

Code Translation to High-Level Abstraction FPGA Design of Stencil Computation (Problem Exploration) (top)

by Beniel Thileepan

Structured-mesh based stencil computations are a common motif appearing in many numerical algorithms. Recent work on FPGA has shown promising performance, both runtime and energy-efficient execution in this class of applications. FPGA vendors have introduced high-level synthesis tools (HLS) based on languages such as C/C++, OpenCL, and SYCL to ease the development of FPGA applications, which would otherwise require low-level RTL languages. However, there are still a significant amount of custom optimizations needed to obtain the best performance. Recent works have demonstrated how Domain Specific Languages (DSLs) are used to generate optimized implementations for various accelerator architectures. In this presentation, I will detail research into using DSLs to generate optimized FPGA code with an overview of the techniques to generate code using the OPS DSL and present some initial performance results on Xilinx FPGAs with comparison to hand-coded performance and performance from traditional architectures such as CPUs and GPUs.

Machine Learning + AI

Supervised Machine Learning Model for Diabetic Students’ Glucose Levels Classification System (top)

by Mona Alotaibi

Accurate and timely blood glucose prediction is essential for students who have been diagnosed with diabetes to prevent hypoglycemia and hyperglycemia episodes throughout the school days. Continuous Glucose Monitoring (CGM) is to manage blood sugar levels effectively which is a form of sensor that can be incorporated in the Internet of Things . In addition, machine learning (ML) models can precisely be used for predicting glucose levels using a CGM system. We aim to develop a decision support system based on ML algorithms and the IoT, for which we have applied Random Forest, Logistic Regression, AdaBoost (Ada), Multi-layer perceptron (MLP), and linear SVM algorithms, followed by a majority voting algorithm. Evaluation metrics such as the Confusion Matrix, Accuracy, Precision, F1 score and Recall, were implemented to evaluate the performance of each algorithm. The experimental results demonstrate that the accuracy of the majority voting model is 99.61%.

Different cortical connectivities in human females and males relate to differences in physical workload and body composition, reward and emotional systems, and memory (top)

by Ruohan Zhang

Gender differences in functional connectivity between females and males are well established, but what they relate to is unclear. In this investigation, we investigated this by identifying functional connectivity links that were different between females and males, and then analysing associations of these with behavioural measures. Most of the functional connectivities were lower in females, with the mean Cohen’s d = -0.18. The lower functional connectivities in females were especially of somatosensory/premotor regions including the insula, opercular cortex, paracentral lobule and mid-cingulate cortex, and were correlated with lower maximum workload, and with higher whole body fat mass. But some functional connectivities were higher in females, involving especially the ventromedial prefrontal cortex and posterior cingulate cortex, and these were correlated with higher liking for some rewards such as sweet foods, higher happiness/subjective well-being, with better memory-related functions.

Weakly Supervised Learning for Predicting Viral Oncoprotein in Nasopharyngeal Carcinoma (top)

by Made Satria Wibawa

Nasopharyngeal carcinoma (NPC) is a type of cancer originating in the nasopharynx, often associated with Epstein-Bar virus infection. This virus encodes several viral oncoproteins, such as Latent Membrane Protein (LMP1) which may promote NPC progression. Because of this nature, LMP1 is an excellent therapeutic for EBV-associated NPC. However, the detection of LMP1 on NPC with standard testing may cause additional costs and delays the diagnosis. In this study, we utilised haematoxylin and eosin-stain (H&E) histology images which is more viable. Using deep learning-based method, we attempt to predict the LMP1 status of NPC. We utilised an attention mechanism with residual connection in our weakly supervised learning. In our 3-fold cross-validation experiment, we achieved average accuracy, AUC and F1-score 0.936, 0.995 and 0.862, respectively. This method allows us to examine model interpretability. To the best of our knowledge, this is the first attempt to predict LMP1 status on NPC.

Spatial-temporal graph convolutional neural networks on evolving cell surfaces (top)

by Edward Offord

Many cellular processes involve complex deformations of the cell surface that change over time, which are difficult to automatically detect and analyse in 3D microscopy images. Attempting to use modern machine learning methods for extracting dynamic features from surface triangulations of a cell is a major challenge, because the topology is constantly changing over time. To incorporate time as a feature to inform segmentation on cell surfaces this talk presents heterogeneous graph neural networks, using a time based hypergraph of a cell timeseries. Here, we focus on identification of macropinocytic cups on the surface of Dictyostelium cells, structures involved in the uptake of extracellular fluid. The network classifies each node into belonging to a cup or not, enabling subsequent studies of the detailed distribution of molecules regulating fluid uptake in cells.

Network dissection in Diffusion Model (top)

by Yu Cao

Recently, the diffusion model has become a rising class of generative models by virtue of its power-generating ability. Compared to other deep generative models, however, diffusion models are certainly a bigger black box for humans. In this work, we use network dissection, an analytical framework that systematically identifies the semantics of individual hidden units in image classification and image generation networks. First, we find evidence that the units in the VDM model are significantly more efficiently utilised than the GAN by analysing the role of the units of those. Secondly, by analysing the changes when small units are activated or deactivated, we find that although the mathematical principles of the two models is the same, in reality the two models have completely different operating mechanisms. Thus, we answer the question of why the DDPM performs extensively better than the VDM, both from a theoretical and experimental perspective.

Fault Tolerant Task Offloading in Edge-Based IoT networks, A Multi-armed Bandit Approach. (top)

by Alya Alhazmi

IoT devices have limited computing capabilities, and cannot execute computationally-intensive tasks such as deep learning. To address this issue, devices offload some tasks to powerful servers at the edge of the network, to bound communication delays. As users are mobile, they can only communicate with a varying set of edge nodes (ENs) at any given time. Hence, as IoT devices move and offload their tasks, the capacities of the available ENs vary, giving rise to a dynamic network, exacerbating the offloading problem, which can be stated as follows: Given multiple ENs with unknown capabilities and a mobility pattern, select the most appropriate EN set that guarantees minimum delay and energy consumption. We formalise this problem as a delay-energy optimization problem by jointly considering EN selection for task offloading and fault tolerance in the edge network. We develop a multi-armed bandit framework that learns under the various uncertainties in system dynamics.

Research in Federated Graph Neural Networks (top)

by Yan Qian

Federated Learning (FL) has been a hot research topic in enabling the collaborative training of machine learning models among different organizations under the privacy restrictions. As researchers try to support more machine learning models with different privacy-preserving approaches, there is a requirement in developing systems and infrastructures to ease the development of various federated learning algorithms. Federated learning systems are equivalently important and face challenges from various aspects such as effectiveness, efficiency, and privacy. Graph Neural Network (GNN) research grows rapidly that help in learning distributed representations from graph-structured data. On the other hand, centralizing a massive amount of real-world graph data for GNN training is prohibitive due to privacy concerns, regulation restrictions, and commercial competitions, and FL could help GNN to train in the distributed system that benefits dealing with the large-scale dataset, as well as the privacy concerns.

Monocular Depth Estimation (MDE) (top)

by Bashayer Abdallah

Computer vision applications, such as scene understanding and reconstruction, rely on monocular depth estimation from images. Traditional monocular depth estimation (MDE) methods, such as defocused methods, rely on depth cues for depth prediction and which unsuitable for images of a blurred texter. Deep learning (DL) techniques for MDE were recently proposed, showing great promise in dealing with the classic ill-posed problem. However, those methods continue to suffer from edge blur challenges, especially in complex object structures. The blurry edges contribute to depth estimation errors and result in flying pixels. Our initial work is to review edge detection state-of-art studies and how to implement these algorithms on our MDE baseline to comprehend the effects of such methods on capturing the fine edges in predicted depth maps.

Exploration in NAS (Neural Architecture Search) techniques (top)

by Minghao Xu

Neural Architecture Search has been widely used to automatically discover better neural network models in an ample search space of candidate network architectures. However, due to the massive number of architectures, it usually takes hundreds and thousands of GPU hours to train and test to get the most optimal structure. Recent works focus on new strategies named zero-cost proxies to remarkably reduce training time. By introducing such metrics, we can predict the expected performance of the network in several seconds, which is efficient enough for computers to search in the vast search space at a significantly lower cost. My research interests lie in designing such strategies to optimize the procedure of NAS. Not only by proposing new zero-cost proxies but by developing more intelligent model search algorithms, which conserve previously trained parameters to retain the accumulated accuracy like transfer learning.

What would Galileo think of Neural Architecture Search? (top)

by Rosco Hunter

This talk will introduce the field of Neural Architecture Search (NAS), a technique interested in automating the discovery of the best-performing neural network architectures with respect to a chosen combination of accuracy, latency, etc. The talk will cover examples of metrics that are predictive of network architecture performance; the role of NAS benchmarks in evaluating the effectiveness of different metrics and the potential of NAS to advance our understanding of why neural networks even work.

Multi-Camera Multi-Object Trajectory Forecasting (MCTF) (top)

by Amey Noolkar

The objective of MCTF is to predict the trajectory of a moving object across a network of cameras. The existing approach (Styles et al) tracks one object across multiple cameras. However, its use for surveillance applications is limited. There exist robust object-tracking and prediction methods for single cameras, however they are not easily scaled for a network of cameras. We address these shortcomings by proposing an MCTF model with a location-tensor-based framework which would represent the location of objects across multiple cameras using the position of their bounding boxes over time. The model shall follow a Which-When-Where approach and learn to predict the movement of the bounding boxes across the network of cameras in an image-agnostic way. Concretely, we want to predict in which camera(s) the bounding-boxes corresponding to the tracked objects are most likely to appear, and when and where within the camera views would they appear.

Explaining predictions of Neural Network with hidden paradigms (top)

by Bhushan Atote

Along with the enormous achievement, Artificial intelligence (AI) plays an integral part in our daily life. A major barrier to the deployment of AI-based systems, despite these tremendous breakthroughs, is that they frequently lack transparency. Thus, these systems lose people’s trust and has reignited the discussion on explainable AI (XAI). XAI has a lot of potential for enhancing the transparency and trust of AI-based systems. Strong predictions can be made using Deep Neural Networks (DNN), but they cannot be directly explained. In terms of perceptibly human-friendly representations, such as word phrases in text or super-pixels in images, concept-based explanations justify model judgements. Convolutional neural networks (CNNs), in particular, are DNNs used in computer vision. In this work, we introduced a deep neural architecture – Hidden Paradigm Network (HiPaNet): the network analyses the image to find hidden paradigms and then produces the classification result by highlighting its importance.

Machine Learning for Histopathology Images (top)

by George Wright

Pathologists manually examining slides under a microscope is considered best practice in the diagnosis of many medical problems. This is a costly and time-consuming process due to a shortage of pathologists and an increase in demand. Computational pathology aims to automate clinical workflows using machine learning algorithms to analyse histopathology whole slide images (WSIs). These algorithms are advantageous as they can quickly provide a standardised diagnosis incorporating several different data sources and can offer valuable insights such as the factors affecting prognosis. However, due to the large size of multi-gigapixel WSIs, the use of traditional image analysis algorithms is not possible due to the large memory requirements. Specialist computational pathology algorithms exist in numerous medical domains that in many cases can outperform a trained pathologist. This problem exploration talk will explore recent advances in the methods used for processing histopathology images and their possible applications.

Source Code Plagiarism with Supervised Machine Learning: Open Issues and Challenges (top)

by Fahad Ebrahim

Plagiarism in programming assignments can be referred to as: "source code plagiarism". As manual checks for plagiarism are time-consuming, the automation of this process has been initiated targeting a fair evaluation and integrity. Supervised Machine Learning (SML) has been applied to source code plagiarism as plagiarism can be treated as a binary classification problem with two outputs: 'yes' (plagiarised) or 'no' (not plagiarised). While surveying the SML approaches as part of the Master’s dissertation, there were several open issues. There is a lack of adequate academic datasets due to legal/social issues. The available reuse datasets have a severe imbalance between training and testing sets as the percentage of real plagiarism is minor compared to non-plagiarism ones. The performance of the models is based heavily on feature extraction and selection. Most of the developed models are language-specific and treat source code as a plain natural language that limits their capability.

Towards Efficient Large-Scale Federated Learning (top)

by Sara Alosaime

Data privacy concerns and strict government legislation (e.g. GDPR) have raised interest in privacy-preserving distributed training. Federated Learning (FL), proposed by Google in 2016, allows clients to synchronise locally trained models rather than their own data. Mobile Edge Computing is ideal for large-scale FL with hierarchical layers to train a generic and robust model on many clients. However, as deployments scale up, there may be unique FL challenges related to clients’ heterogeneity, lack of resources, availability and communication issues. Consequently, the number of stragglers and dropped participants will increase, preventing the FL from reaching the target convergence within the expected time and wasting clients' resources. In light of these real-world challenges, our aims to design multi-layer FL architecture for Human Activity Recognition system that balances training progress, model accuracy, and resource preservation.

Optimal Transport for Causal Abstractions (top)

by Giorgos Felekis

Complex systems can encompass a range of fields, from physical systems like particle behaviour to social systems like decision-making. The challenge with studying complex systems is their level of detail regarding the dependencies and interactions within their parts or with its environment. Structural Causal Models (SCMs) are commonly used to model these systems by representing the cause-and-effect relationships between variables and predicting the consequences of changes/interventions to them. To simplify their analysis, we create abstractions which are simplified high-level versions of them but retain the crucial information. Our goal is to connect these abstractions back to the original, more detailed models, allowing us to jointly reason about evidence across models with multiple levels of granularity. To achieve this, we introduce a novel approach that leverages Optimal Transport theory to align the probability distributions resulting from interventions on the Structural Causal Models of both the detailed and abstracted systems.