RA Symposium 2017

Welcome to the 2017 Department of Computing Research Associate Symposium

• Date: Friday 23th June 2017
• Venue: Huxley Building LT311

The RA Symposium is a one-day workshop to showcase the wide range of interesting research undertaken by postdoctoral research associates and fellows in the department to other research staff and students of the department. As we aim to celebrate your work in the department, all postdocs are invited to present their work at the symposium, regardless of areas of research.

Lunch and coffee will be provided on the day. There will be multiple prizes (£500, £250 and £100) awarded to top presenters at the symposium, to be voted by registered attendees.

Prize winners

1. Javier Andreu-Perez
2. Dimitrios Letsios
3. Luis Muñoz-González

Congratulations!

Keynote

The keynote of the symposium will be delivered by our new Head of Department, Prof Daniel Ruckert, where he will talk about:

• Overview of department’s research focus and the future plans for the department
• Information on postdoc career opportunities and department support, such as fellowship opportunities and grant opportunities
• Q&A from audience

We are looking forward to your participation!

Registration

Attending the symposium is free of charge. We welcome and invite all members of the Department including students, staff and visitors to attend the symposium.

Please REGISTER in advance to help us estimate attendance.

Schedule

The venue will be LT 311 in Huxley Building unless otherwise specified

Abstracts

Session 1

Context-Sensitive Super-Resolution for Fast Fetal Magnetic Resonance Imaging

Steven McDonagh

3D Magnetic Resonance Imaging (MRI) is often a trade-off between fast but low-resolution image acquisition and highly detailed but slow image acquisition. Fast imaging is required for targets that move to avoid motion artefacts. This is in particular difficult for fetal MRI. Spatially independent upsampling techniques, which are the state-of-the-art to address this problem, are error prone and disregard contextual information. We propose a context-sensitive upsampling method based on a residual CNN that learns organ specific appearance and adopts semantically to input data allowing for the generation of high resolution images with sharp edges and fine scale detail. By making contextual decisions about appearance and shape, present in different parts of an image, we gain improved structural detail at a similar contrast provided by high-resolution data. We experiment on 145 fetal scans and show increased PSNR over upsampling baselines that in turn improves brain volume reconstruction on motion corrupted data.

Efficient reflectance models for diffractive structural colours in nature

Daljit Singh Dhillon

Computer Graphics often employs simplistic formulations based on ray optics to render virtual scenes efficiently. With increasing emphasis on realistic appearances in virtual reality, such simplifications fall short in modelling complex effects that can only be explained with wave-optics. This includes rich, vivid and iridescent reflectances on plant, animal and insect bodies as well as inanimate plastic or metalic surfaces. The physics behind such reflectances may involve wave interference, diffraction, off-surface scattering or even sub-surface scatering. Our research presently focuses on diffractive reflectances. Previous research has shown that a large array of 2D textures precomuted by applying scalar wave-optics to actual, scanned surface nano-gemoetries enable real-time rendering of colourful iridescence effects accurately. We have made this reference approach highly efficient by reducing the required number of lookup tables to as few as two. In effect, this reduces shader programs memory footprint by an order of magnitude and improves its speed by a factor of 6X. We achieve this by using Chebyshev approximations for precomputing view-dependent variations in diffractive reflectances. Furthtermore, we are developing an effective approach to capture and model spatially varying diffractive colours on surfaces. Such surfaces are commonly found on modern, decorative and fashionable fineries as well as bright and shiny insects like butterflies and bees. Our optimizations allow virtual reality applications to achieve real-time performance on low-end GPU platforms as well.

Towards Everyday Human Computer Interaction with Non-invasive Brain Computer Interfaces

Javier Andreu-Perez

Individuals with a neurological movement impairment can partially recover their lost motor function from rehabilitation that involves repetitive movements. Although it is difficult for some stroke survivors to move the stroke-impaired limb, most of them can perform motor imagery, i.e. the mental process of imagination of movements without physical movement. The rationale for performing motor imagery arises from the similar brain activations it shares with performing physical movements. However, whilst a physical movement can be observed, motor imagery is a mental process that cannot be observed. Recent advances in the analysis of brain signals and improvements in computing capabilities have enabled people with motor disabilities to use their brain signals for communication and control without using their impaired neuromuscular system. Performance variation is one of the main challenges that BCIs are confronted with, when being used over extended several sessions. We posit that the level of self-adaptation provided to the BCI user should be adjusted in real time to enhance BCI reliability over time. Many of the factors contributing to suboptimal BCI performance listed above could be addressed through BCI adaptation.

Session 2

Poisoning Neural Networks with Back-Gradient Optimization

Luis Muñoz-González

A number of online services nowadays rely upon machine learning and neural networks to extract valuable information from large data volumes. However, learning from data collected in the wild exposes machine learning algorithms to the threat of poisoning, i.e., a coordinate attack in which a fraction of the training data can be controlled by the attacker and purposely manipulated to subvert the learning process. In this work, we propose a novel poisoning algorithm that, compared to current poisoning strategies, can target a wider class of learning algorithms, including neural networks and deep learning models, while also being less computationally demanding. Our approach relies on back-gradient optimization applied to a bi-level optimization problem, where by reversing the learning process we can efficiently compute the gradients needed to update the parameters of the optimization objective. We empirically validate the effectiveness of our attack against neural networks on application examples including spam filtering, handwritten digit recognition, and malware detection.

Heuristics with Performance Guarantees for the Minimum Number of Matches in Heat Recovery Networks

Dimitrios Letsios

Joint work by Georgia Kouyialis, Dimitrios Letsios, Ruth Misener

Optimal heat exchanger network synthesis improves heat recovery in chemical processes [1, 2, 3, 4]. Heat exchanger network synthesis is a non-convex mixed-integer nonlinear optimization problem that may be solved simultaneously, i.e. to generate an optimal network without decomposition [5, 6, 7, 8]. An alternative, sequential method, decomposes the synthesis problem into: (i) minimizing utility cost, (ii) minimizing the number of matches, (iii) deriving many possible superstructures, (iv) minimizing the investment cost, and (v) configuring the final network [9]. The sequential method cannot guarantee global optimality of the original problem, but it reduces the computational burden.
We investigate the minimum number of matches subproblem of the sequential method. This subproblem is an NP–hard mixed-integer linear program (MILP) exhibiting: combinatorial explosion in the possible hot and cold stream configurations and symmetric mathematical structure creating unnecessarily large search trees [10, 11, 12]. Because the minimum number of matches subproblem still poses computational difficulty for moderately-sized networks [13], we develop heuristic methods and provably efficient approximation algorithms with guaranteed solution quality and run-time bounds. In the sequential method, many possible stream configurations are required to evaluate the minimum overall cost. So, a complementary contribution of this presentation is developing efficient algorithms producing multiple solutions.
Previous work introduces a collection of approximation algorithms which are based on rounding an optimal fractional solution for the transportation MILP formulation of the problem [14]. A unified worst-case analysis of their performance guarantees shows a 2-approximation ratio for a single temperature interval and a non-constant approximation ratio for multiple temperature intervals scaling with the number of hot and cold streams.
This presentation shows a tight analysis of the deterministic LP rounding algorithm, which does admit a constant factor approximation algorithm. This finding points out an inherent drawback of LP-based approximation algorithms for the minimum number of matches problem, namely their dependence on the “big-M constraints”. So, the presentation elaborates on the design of heuristic methods exploiting the combinatorial structure of the problem. The resulting algorithms are obtained by efficiently solving further subproblems of the minimum number of matches and novel relaxations. Improved guaranteed solution quality and run-time bounds are obtained. This work also strengthens the existing MILP formulations of the problem by tightening of the big-M constraints and the addition of novel cuts. The proposed algorithms and the strengthened MILP formulations are validated experimentally.

1. C.A. Floudas and I.E. Grossmann. Synthesis of flexible heat exchanger networks with uncertain flow rates and temperatures. Comput Chem Eng, 11(4):319–336, 1987.
2. J. A. Elia, R. C. Baliban, and C. A. Floudas. Toward novel hybrid biomass, coal, and natural gas processes for satisfying current transportation fuel demands, 2: Simultaneous heat and power integration. Ind Eng Chem Res, 49(16):7371–7388, 2010.
3. R. C. Baliban, J. A. Elia, R. Misener, and C. A. Floudas. Global optimization of a MINLP process synthesis model for thermochemical based conversion of hybrid coal, biomass, and natural gas to liquid fuels. Comput Chem Eng, 42:64–86, 2012.
4. M. Escobar and J. O. Trierweiler. Optimal heat exchanger network synthesis: A case study comparison. Appl Therm Eng, 51(1-2):801–826, 2013.
5. T. F. Yee and I. E. Grossmann. Simultaneous optimization models for heat integration – II. Heat exchanger network synthesis. Comput Chem Eng, 14(10):1165–1184, 1990.
6. A. R. Ciric and C. A. Floudas. Heat exchanger network synthesis without decomposition. Comput Chem Eng, 15(6):385–396, 1991.
7. K. P. Papalexandri and E. N. Pistikopoulos. Synthesis and retrofit design of operable heat exchanger networks. 1. Flexibility and structural controllability aspects. Ind Eng Chem Res, 33(7):1718–1737, 1994.

SquirrelJoin: NetworkAware Distributed Join Processing with Lazy Partitioning

William Culhane

To execute distributed joins in parallel on compute clusters, systems partition and exchange data records between workers. With large datasets, workers spend a considerable amount of time transferring data over the network. When compute clusters are shared among multiple applications, workers must compete for network bandwidth with other applications. These variances in the available network bandwidth lead to network skew, which causes straggling workers to prolong the join completion time.
SquirrelJoin is a distributed join processing technique that uses lazy partitioning to adapt to transient network skew in clusters. Workers maintain in-memory lazy partitions to withhold a subset of records, i.e. not sending them immediately to other workers for processing. Lazy partitions are then assigned dynamically to other workers based on network conditions: each worker takes periodic throughput measurements to estimate its completion time, and lazy partitions are allocated as to minimise the join completion time. We implement SquirrelJoin as part of the Apache Flink distributed dataflow framework and show that, under transient network contention in a shared compute cluster, SquirrelJoin speeds up join completion times by up to 2.3X with only a small, fixed overhead.

Algorithms and Framework for Energy Efficient Parallel Stream Computing on Many-Core Architectures

Nicolas Melot

The rise of many-core processor architectures in the market answers to a constantly growing need of processing power to solve more and more challenging problems such as the ones in computing for big data. Fast computation is more and more limited by the very high power required and the management of the considerable heat produced. Many programming models compete to take profit of many-core architectures to improve both execution speed and energy consumption, each with their advantages and drawbacks. The work described in this thesis is based on the dataflow computing approach and investigates the benefits of a carefully pipelined execution of streaming applications, focusing in particular on off- and on-chip memory accesses. As case study, we implement classic and on-chip pipelined versions of mergesort for Intel SCC and Xeon. We see how the benefits of the on-chip pipelining technique are bounded by the underlying architecture, and we explore the problem of fine tuning streaming applications for many-core architectures to optimize for energy given a throughput budget.
We propose a novel methodology to compute schedules optimized for energy efficiency given a fixed throughput target. We introduce Drake, a tool that generates pipelined applications for Many-Core architectures and allows the performance testing in time or energy of their static schedule. We show that streaming applications based on Drake compete with specialized implementations and we use Schedeval to demonstrate performance differences between schedules that are otherwise considered as equivalent by a simple model.

LibSEAL: Detecting Service Integrity Violations Using Trusted Execution

Florian Kelbert and Pierre-Louis Aublin

We will present LibSEAL, a SEcure Audit Library for internet services that creates a non-repudiable audit log of service operations and checks invariants to discover violations of service integrity. LibSEAL acts as drop-in replacement for TLS libraries used by services, and thus observes and logs all service requests and responses. It runs inside a trusted execution environment (TEE), such as Intel SGX, to protect the integrity of the audit log. Logs are stored using an embedded relational database, which permits service invariant violations to be discovered using simple SQL queries. We evaluated LibSEAL with three services (Git, ownCloud and Dropbox), showing that it is effective in discovering integrity violations, while reducing throughput by at most 32%.

Session 4

Meander: Cooperative Resilient Data Stream Processing for the Internet of Things

Dan O’Keeffe

Real-time data-intensive Internet of Things (IoT) applications, e.g. for environmental monitoring, video analysis, or object tracking, can be specified as streaming queries, in which operators transform tuples in data streams. Without a fixed IT infrastructure, IoT devices can process sensor data cooperatively, i.e. using the resources of multiple devices interconnected by a wireless mesh network. In such deployments, however, the performance of streaming queries is heavily influenced by the wireless network connectivity between devices, which can vary dramatically due to environmental interference and device mobility. An open challenge is how a stream processing system should adapt to changes in network bandwidth between IoT devices, and how it should handle intermittent connectivity.
We describe Meander, a distributed and resilient data-parallel stream processing system for IoT devices. Its key idea is to improve the throughput of streaming queries in an unreliable wireless mesh network by exploiting network path diversity: a streaming query includes replicated operators at different nodes, which increase the number of possible network paths that tuples can take. Meander dynamically routes stream data to operator replicas based on network path conditions: nodes probe the path throughput and use backpressure stream routing to decide on transmission rates, utilising multiple operator replicas in a data-parallel fashion. If a node disconnects from the network for a prolonged period of time, a node disconnection recovery mechanism reprocesses the exact set of lost tuples. Our experimental evaluation of Meander shows that network path diversity can improve throughput considerably, while being resilient to intermittent connectivity of nodes.

Implementation of Edge-Analytics for Anomaly Detection in Water Networks by an Arduino based Wireless Sensor Network

Mehrdad Babazadeh

A novel distributed data analytics architecture and algorithms that have been applied to infrastructure anomaly detection. In particular, we have focused on smart water networks and demonstrate that high sensor rate analytics can be performed at on edge nodes without the requirements of having to send all data back to servers therefore saving both communications costs and lengthening battery powered node lifetimes. To achieve this we are required to schedule a complex set of multiple tasks including data mining and communication on a lightweight single core Intel Curie processor, Arduino 101. Instead of analysing the raw data as it is read from the sensors, we compress the data using our customized Lempel Ziv compression algorithm tailored to resource-constrained embedded systems. From this, we analyse the compression rate figures only and send only the data associated with the anomalous condition using a LoRa platform which again minimizes transmission payload.