26 May 2016
A joint student conference between the 3 CDTs in the School of Informatics:
This one-day will showcase the research across the three CDTs and PhD and MSc(R) students from across the School of Informatics and students from the School of Maths in related research areas.
The day will present a snapshot of the wide-ranging research from across the School of Informatics, and will facilitate networking and collaboration across CDTs, companies and research areas. The event will also give PhD and MSc(R) students the chance to learn more about each others' research and practice their research communication skills.
For additional details, including how to register to attend, please contact datascience[a]inf.ed.ac.uk.
Justs Zariņš, CDT in Pervasive Parallelism
Runtime management for asynchronous algorithms
I will talk about iterative asynchronous algorithms - in essence trading correctness for performance. Specifically I will share my findings on trying to manage at runtime an asynchronous algorithm running on a shared memory node. The goal of the management is to limit work progress imbalances without sacrificing convergence speed and tolerance to noise.
Rafael Karampatsis, CDT in Data Science
A comparison of different large scale language models for source code
Abstract to follow.
Matt Graham, Institute for Adaptive & Neural Computation
## Sampling under constraints in deep density models
There have been several models proposed for learning probability densities across high-dimensional continuous data by optimising the parameters of an invertible map which transforms the data to a new space where it is (approximately) distributed with some simple density e.g. Gaussian. To independently sample from the learnt joint distribution we can then draw independent samples from the simple density and transform these through the inverse of the learned map. I will present my work on a MCMC method for sampling in such models under constraints in the original data space e.g. due to conditioning on a subset of dimensions having known values.
Victor Dumitrescu, CDT in Pervasive Parallelism
Formally verifying large software systems
Formal methods can give the strongest guarantees about the correct behaviour of software, but formally specifying and proving the correctness of real-world systems is notoriously difficult and expensive. As more and more ambitious verification projects are being completed, such as the CompCert C compiler and the seL4 microkernel, the challenges they faced reveal how interactive theorem provers can be improved to better support these large, collaborative efforts.
Charlie Nash, CDT in Data Science
Modelling 3D Object Classes with Parts
A model of the geometric and semantic relationships present among object classes can be a considerable aid in computer vision applications. Additionally, content creation is a central task in computer graphics, and a shape model can be used to synthesize realistic objects that can be placed within a scene. We present a system that can learn to generate novel instances of an object class using a collection of examples from that object class as training data.
Arpit Joshi, Institute for Computing Systems Architecture
Non-volatile Memory: Optimizing for Performance
The emerging non-volatile memory (NVM) technologies (like 3D XPoint) provide the opportunity for tremendous improvement in performance of applications that require access to persistent data. However, this opportunity comes with the challenges of ensuring consistency of data in the presence of failures. In this talk we will discuss some techniques on how to improve performance in systems with NVM while ensuring consistency of data in the presence of failures.
Simon Fowler, CDT in Pervasive Parallelism
Relating Channel- and Actor-based Models of Concurrent Programming
Akka, Elixir, Erlang, Go: these are a few of the languages and frameworks which take concurrency and communication to be a focal point in language design. Actor-based languages such as Erlang have found much use in implementing scalable and fault-tolerant distributed systems, and channel-based languages such as Go have found much use in systems programming. But how do they relate to each other, and what can we learn from this?
In this short talk, I'll give a high-level overview of my work in relating these two models of communication-centric programming, both in the case where channels and actors can send and receive values of a single type, and where channels and actors can send and receive values in a sequence corresponding to a communication protocol.
Ross McKenzie, CDT in Robotics & Autonomous Systems
Soft Robotics Grippers
Soft Robotics is the study of the use of flexible and extensible materials in robots. Soft robotic grippers can deform to mimic the shape of an object automatically, allowing for robust gripping of unknown and delicate objects. This gripping can also be achieved with less complex control than standard grippers. This presentation provides a brief introduction to the field of soft robotics and showcases its use in grippers.
Bobby Ikonomov, CDT in Data Science
The application of Quantum Computing to Machine Learning
Quantum computing is becoming an increasingly popular topic as of late, with progress being made both in theoretical and experimental fronts. Not surprisingly, researches have recently started to explore the possible application of quantum computing to another prominent field - machine learning. This talk will give a brief summary on the very new field of Quantum Machine Learning and attempt to answer several important questions, such as: How can quantum computing potentially be used to improve machine learning? What relevant algorithms have been already proposed? What are the challenges in implementing these algorithms? When can we expect quantum computers to become practical for machine learning?
Vanya Yaneva, CDT in Pervasive Parallelism
Accelerating Software Testing using GPUs
In software engineering, careful validation of the developed system is a crucial task, which involves the generation and execution of a large number of test cases. Thus, rigorous testing of any non-trivial system could be extremely time consuming, putting an enormous strain on software development cycle. One way to accelerate software testing is to execute test cases in parallel on the GPU threads. In the talk, I am going to present this idea, the challenges of implementing it and how I am addressing them.
James Owers, CDT in Data Science
Deep learning has had its greatest successes in the domain of images. Can we apply similar techniques to the domain of music? What can we do with deep representations of music? How hard would it be to make a generative model?
Philip Ginsbach, CDT in Pervasive Parallelism
Enabling Compilers to Exploit Heterogeneous Computing
Heterogeneous computing is becoming increasingly important for exhausting the potential of current hardware platforms, but established programming models using OpenCL and CUDA require much technical expertise. Specialized libraries such as cuBLAS or cuFFT expose more concrete functionality and are well optimised by expert programmers but are each limited in scope. We present an approach that will enable compilers to automatically spot functionality that suits particular interfaces and then substitute it with library calls.
Andy Brock, CDT in Robotics & Autonomous Systems
Introspection: A Walk in Latent Space of Variational Autoencoders
A brief discussion on the relationship between generative and discriminative modeling, and a real-time visualization demonstration.
Panos Stratis, Labaratory for Foundations of Computer Science
Test Case Permutation to Improve Energy Consumption
Executing large test suites is expensive, both in terms of time and energy with cache misses being one of the main factors contributing to execution time and energy consumption of software. Cache misses are reduced by increasing the locality of memory references. In the present work we explore cache locality optimizations across several program runs, which is the case when we run several test cases.
Streaming Communication Patterns for Domain Decomposition Methods
One of the most significant challenges faced by parallel computing is that of efficient communication between processors. This is becoming more and more important as the industry advances towards exascale computing, as many current algorithms will not scale to the required numbers of processors.
One such algorithm is halo exchange - a standard communication pattern for problems that are decomposed by domain. This talk will discuss some of the problems with using halo exchange on exascale machines, and also explore the advantages of a stream-based alternative.
Valentin Radu, Institute for Computing Systems Architecture
Deep sensing with wearable devices
The large volume and rich data that can be collected with our pervasive mobile devices gives rise to a new perspective of interacting with our environment. In this talk I'm going to present my work concerned with the first step of this process, detecting user context from sensor data, the challenges involved in squeezing deep learning onto limited resource devices and the vision of how these can lead to more natural human computer interactions.
Adam Harries, CDT in Pervasive Parallelism
Abstract to follow.